One aspect that semiotic models of genres focus on is the multilayered meanings of generic terms, which point to the many communicative levels that genres can be defined on and the complexity of genres as signs. As signs, the generic terms can be understood as models for the even more complex models that the genres themselves are conceived as (Raible 1980, 334). Two semiotic models for generic terms are presented in more detail here. These are used as a basis for an empirically established discursive model of subgenre terms for the corpus of nineteenth-century Spanish-American novels created and analyzed in the context of this dissertation.The corpus-specific model for generic terms is presented in chapter 3.2.3 below. The first of the two models has been formulated by Raible (1980, 342–345) and involves six dimensions from which generic terms usually draw their meaning and classificatory features: the communicative situation between sender and recipient (“Kommunikationssituation”) the object area of the texts involving persons and things (“Objektbereich”) the higher order structure of texts (“übergeordnete Ordnungsstruktur”) the relationship between text and reality (“Verhältnis zwischen Text und Wirklichkeit”) the communicative medium that the text uses (“Medium”) the way of linguistic representation (“sprachliche Darstellungsweise”). An example of a generic term that addresses the communicative situation is “children’s book”. In this case, the genre’s name specifies the target group of the texts labeled with it. A term concerning the object area is, for example, “picaresque novel”, which refers to the protagonist’s social status. According to Raible, instances of the third group of terms, which refers to the higher-order structure of the texts, are relatively rare. He gives jokes as examples, as they involve the expectation of something unexpected in their structure. A generic term that addresses the relationship between the text and reality is, for instance, the fable. Terms that relate to the communicative medium can refer to language, other media (music, mimic, rhythm), and carrier media, for example, the “epistolary novel”. Finally, an example for a genre label that alludes to the linguistic representation of the text is “short story”, which refers to the length of the form. Raible sees his model as principally open and refinable through applied literary genre analyses (Raible 1980, 342–345). A similar semiotic model of generic terms has been developed by Schaeffer (1983, 64–130). Like Raible, Schaeffer chooses to approach the complex signs that genres are through the also complex but more tangible names of the genres: Commençons par une question banale: quel est le statut des classes générique? Ou, pour éviter de nous encombrer dès le début d’entités peut-être fantomatiques, demandons-nous plutôt: quel est le statut des noms de genres? [...] l’identité d’un genre est fondamentalement celle d’un terme général identique appliqué à un certain nombre de textes. [...] les noms générique traditionnels sont la seule réalité tangible à partir de laquelle nouns en venons à postuler l’existence des classes génériques [...]. (Schaeffer 1983, 65–67)

The names of the genres are significant because they witness that a generic class of texts has a communicative existence, condensed in a name. Furthermore, names applied to texts signal that the texts participate in the genres, and the participating texts, in turn, contribute to the genre’s identity. The above quote is intentionally reduced to concentrate on the aspect of relevance of the generic names. In the wider context, Schaeffer explains that the relationship between the generic names and the texts is not at all simple. The names can have different statuses, as analytical ex-post terms, as words in use in a very specific historical situation, or as something between both of these poles. They can be applied collectively to a set of texts at once or to individual texts so that multiple applications of the terms, or their sum, form the genre. The meaning of generic names is not fixed, not synchronically, and especially not over time, and they can be related to other terms. One text can be associated with several generic names, which may involve different levels of significance (Schaeffer 1983, 65–66, 69). Schaeffer emphasizes that he does not want to replace a theory of genre with a lexicological study of generic names but wants to start from them to account for the fluent character of the genres and to understand the kind of phenomena that are covered by the generic names (Schaeffer 1983, 75–76). For Schaeffer, a generic term is any term, “á condition qu’il soit utilisé pour classer des œuvre ou des activités verbales linguistiquement et socialement marquées et encadrées (framed), et qui se détachent par là de l’activité langagière courante” (Schaeffer 1983, 77). He thus does not start from a strict separation of generic terms for literary and non-literary works, although his study focuses on the former. A condition for a term to be generic is that it is used for classifying works or other linguistically and socially marked and situated verbal activities. This definition again focuses on the communicative function of genres. It is limited to oral and written linguistic acts because it presupposes verbal activities, which excludes, for example, communicative acts in the visual or performing arts or music. However, within the literary-linguistic frame set by Schaeffer, all kinds of generic terms are considered.

A central observation that Schaeffer makes is that generic names do not all refer to one specific dimension of a literary work. Possible dimensions are, for example, the syntactical and semantic chain of a text that expresses a work. Instead, generic terms pick up all kinds of levels of a work as a global discursive act. That way, the generic identity of a literary work is not unique and fixed but depends on the perspective or perspectives taken towards it (Schaeffer 1983, 79–80). Similar to Raible, Schaeffer defines a literary work as a complex semiotic object. He describes dimensions of this object to which generic terms usually refer. Schaeffer’s model comprises five principal dimensions: utterance of the discursive act (“énonciation”) its destination (“destination”) its function (“fonction”) its semantic realization (“sémantique”) its syntactic realization (“syntaxique”). The first three levels belong to the communicative frame of the discursive act, and the other two concern its textual realization. Each level is further differentiated by reference parameters: for example, a real, fictitious, or simulated instance of utterance, or grammatical, phonetic, metric, or stylistic constraints on the syntactic level. Schaeffer also stresses that the model should not be considered complete but representative or exemplary (Schaeffer 1983, 116). As examples of generic terms pointing to different levels, he mentions, amongst others: a letter or a prayer as instances of a directed utterance, a love poem or ode as examples of an expressive function, a science-fiction story or a western as specific semantic realizations, and a lipogram (forms requiring that specific letters are left out) as a kind of syntactic realization (Schaeffer 1983, 96, 102, 108, 114).

The approaches that view literary genres as complex semiotic objects are characterized by differentiation and openness. An advantage is that they enhance the comparability of different genres by clarifying on which discursive levels generic terms operate without restricting the functioning of genres to a specific communicative level. On the other hand, some genre theoretical aspects are not clarified by these models because they focus on the communicative nature of generic structures. For instance, the semiotic models do not include the generic terms’ provenience and their theoretical or historical nature into the core model, nor do they make statements about the kind of categories that genres can be (if they are to be understood as classes of texts, as prototypical structures, or other types of categoric relationships between literary works). Therefore, these two genre theoretical aspects are discussed further in the subchapters 2.1.3 (“System and History”) and 2.1.4 (“Categorization”).

The focus of the semiotic models on generic names leaves out another aspect of genres: who says that there are no communicative patterns without a name? There may be genres that have not been explicitly discussed or labeled but nonetheless exist as frames for communicative acts. A sign of this is that there are genres that have primarily been labeled by literary scholars in retrospect but that were not explicitly named in their historical peak phase. This does not necessarily mean that the scholars made arbitrary classifications without considering contemporary communicative practices. For example, both the novela gauchesca and the novela indigenista could not be found as explicit generic labels in the bibliography and corpus of nineteenth-century Spanish-American novels that were prepared for this dissertation.See the chapters 3.2.3  and 3.3.4 on the assignment of subgenre labels to the works in the bibliography and the corpus. They are, nonetheless, established subgenres of the novel in the corresponding literary historiography (see, for instance, Ghiano 1957 and Meléndez 1961). Furthermore, generic signals, i.e., text-external or -internal aspects of literary works that indicate in which genres they participate, are not limited to explicit mentions of generic names. They can also be implicit or established through intertextual references (Fowler 1982, 88–105). However, such indirect signals are not directly congruent with a sign-based linguistically oriented approach and thus need to be taken into account in addition.

Up to this point, the ontological status of genres has been discussed, with a particular focus on semiotic theories of genres. In the following chapter, the general, genre-theoretical considerations will be directly related to the approaches of digital genre stylistics. Which possibilities of knowledge about genres arise from digital analyses, if they are carried out starting from certain genre-theoretical foundations? Which methodological aspects of computational genre analyses play a role in this context? Which genre theoretical approaches have been used in digital genre stylistics so far? In the following, these questions will be discussed, focusing on the special role of digital text corpora, genre labels, textual features, and text style in digital genre stylistics.

Regarding the question of the ontological status of genres, Hempfer’s synthesis can be productively related to approaches pursued in digital genre stylistics. In digital stylistics, the anchoring point between genres as communicative norms and their descriptions in terms of textual features is just the style of the texts. Whenever literary works are associated with specific genres, a basic assumption for a digital stylistic genre analysis is the following: that the text style can be analyzed to assess to what degree there are normative facts expressing the generic participations of the works in the genres and of what these facts consist. Obviously, the analysis of text style is limited to the syntactic realization of the discursive act, but this does not mean that digital stylistic concepts of genre are reduced to this level of communication. The level on which digital genre stylistics principally operates is the linguistic, strictly speaking, even the orthographic surface of certain manifestations of literary texts as they are transmitted in a form that is determined by the digital medium. Still, many kinds of discursive aspects can be analyzed on this level. The crucial point is how the participation of the texts in genres is modeled and defined. As several literary genre theorists have pointed out, texts can be classified arbitrarily by any criterion, and this would include any computationally tractable aspect of text style.In his set of terms, Hempfer, for instance, also includes the term “Sammelbegriff” (“collective term”), which he uses to designate logically disjunct groups of texts established on any characteristic: “Genauso wie man ‘Kopf’, ‘Apfel’, ‘Platz’, ‘Tisch’ u.ä. mit dem Prädikator ‘rund’ belegen und somit eine Klasse von Gegenständen bilden kann, der die Eigenschaft ‘rund’ zukommt, kann man Texte aufgrund ihrer Länge, des Vorhandenseins oder Fehlens eines Erzählers, der Fiktionalität oder Nichtfiktionalität, der Tatsache, ob sie in Vers oder Prosa geschrieben sind, usw., einer bestimmten Textklasse zuordnen. Wie das Beispiel der Klassenbildung mit dem Prädikator ‘rund’ darlegen sollte, braucht eine solche Klassifizierung keineswegs aufgrund von für die dergestalt klassifizierten Objekte wesentlicher Eigenschaften zu erfolgen, und dieselben Objekte können, je nach der Eigenschaft, die man wählt, verschiedenen Klassen zugeordnet werden” (Hempfer 1973, 28). Schaeffer too, when discussing the problems of differentiating between theoretical and historical genres, notes that an infinite number of traits can be chosen to compare texts: “En second lieu, et inversement, le nombre de caractéristiques selon lesquelles on peut regrouper deux textes quelconques est indéfini sinon infini. Cela est dû au fait que, lorsqu’on compare deux textes, on ne part pas de leur identité numérique (toujours simple), mais de ce que Luis J. Prieto appelle leur identité spécifique (défini comme un ensemble de caractéristiques non contradictoires). Or, «comme chaque objet possède un nombre infini de caractéristiques, il peut posséder un nombre infini d’identités spécifiques; et comme n’importe quelle caractéristique que présente un objet donné peut toujours aussi faire partie des caractéristiques d’un autre objet, chaque objet peut partager n’importe laquelle de ses identités spécifiques avec un nombre infini d’autres objets 3»” (Schaeffer 1983, 67–68). The potential arbitrariness of the textual features that describe a text category is thus an important point to consider when text classification as a computational activity is used to determine literary genres. Such an endeavor is not the point in itself. The question is to which communicative norms of genre the texts relate and in what way. A good example of taking the relativity and significance of generic assignments into account is a study conducted by Underwood, in which he analyzes different definitions of Gothic novels at different points in time.He differentiates between “detective fiction”, “science fiction”, and “Gothic” (Underwood 2016, 4). As Underwood states: Distant reading may seem to lend itself, inevitably, to literary scholar’s fixation on genre as an attribute of textual artifacts. But the real value of quantitative methods could be that they allow scholars to coordinate textual and social approaches to genre. This essay will draw one tentative connection of that kind. It approaches genre initially as a question about the history of reception — gathering lists of titles that were grouped by particular readers or institutions at particular historical moments. But it also looks beyond those titles to the texts themselves. Contemporary practices of statistical modeling allow us to put different groups of texts into dialogue with each other. (Underwood 2016, 2)

The debate that Underwood engages in is the question of the life cycle of genres. Some critics sustain that genres have relatively short life cycles, roughly corresponding to one generation and about 25 years. Others say that genres can sustain an identity over periods much longer than that, even if there are shifts in the concept of the genre. “Textual analysis won’t prove either claim wrong, but it may help us understand how they’re compatible” (Underwood 2016, 2). With this assertion, Underwood outlines an important task of digital genre stylistics: not necessarily to refute or confirm claims that are made on other levels of genre investigation (as here in the history of reception), but to look for textual and more specifically stylistic evidence as traces of these other levels. That way, genres are not established in style, but through style. Underwood aims to investigate how textually coherent the Gothic is over time: Evidence of this kind [that only a one-generational linguistic coherence could be found] wouldn’t rule out the possibility of longer-term continuity: we don’t know, after all, that books need to resemble each other textually in order to belong to the same genre. But if we did find that textual coherence was strongest over short timespans, we might conclude at least that generation-sized genres have a particular kind of coherence absent from longer-lived ones. (Underwood 2016, 3)

So digital genre stylistics can help to find out which genres imply stylistic coherence of the texts attributed to them at all and on which levels of text style they do. Underwood finds out neither strong evidence for the succession of genre generations nor for a gradual consolidation of the genres over time. The sensation novel is short-termed but textually not very coherent, while detective novels and science fiction novels are textually connected for longer periods (Underwood 2016, 4).Underwood’s method is predictive modeling with L2-regularized logistic regression based on the top 10,000 word features in the text collection (Underwood 2016, 7). The findings of his article from 2016 have been integrated into his book “Distant Horizons” (Underwood 2019, 34–67).

If digital stylistic genre analysis functions as a connector between social and communicative norms of genre and stylistic textual evidence, several aspects in the connective chain need to be defined and selected with care. Usually, a corpus of texts is analyzed, and the generic conventions in question are expressed as genre labels of the texts, which themselves are representatives of literary works. The assignment of genre labels to the texts is the first crucial point. Which kind of genre labels are selected, and how are they assigned to the texts? The semiotic models of generic terms provide a way to differentiate between different discursive levels on which genres can be defined, which can help not to compare “apples with oranges”. That would happen if one would, for example, contrast primarily formally defined genres with thematically defined ones. The sources of the genre labels should always be indicated to document which kind of generic convention they represent. Do the genre attributions go back to assignments made to individual texts by different authors, editors, or publishers? Or are they collectively defined, for example, established in a discussion of a set of works by a contemporary critic or poet, by modern institutions, or by a literary historian? Are the assignments made based on explicit generic terms or implicit signals of the texts? Or are they derived from specific theoretical definitions of genres that are applied to the texts? Depending on the answers, quite different kinds of generic conventions can be analyzed. In the worst case, it is not clear which type of genre an analysis aims at, and the goal of addressing a communicative pattern that lies outside of the analysis itself would be missed. Awareness of the kind of analysis target can still be raised in digital genre stylistics.The case of Underwood is an example of a clear definition and transparent documentation of the target genre convention: “To investigate these questions, I’ve gathered lists of titles assigned to a genre in eighteen different sites of reception. Some of these lists reflect recent scholarly opinion, some were defined by writers or editors earlier in the twentieth century, others reflect the practices of many different library catalogers (see Appendix A) [...] By comparing groups of texts associated with different sites of reception and segments of the timeline, we can ask exactly how stable different categories have been” (Underwood 2016, 4). In their classification of subgenres of the German novel, Hettinger et al. also explain that they analyze genre attributions made by literary scholars: “Literary scholars and common readers use labels like educational novel, crime novel or adventure novel to organize the large domain of fiction. In both discourses the use of these categories is well-established even though they are evolving and tend to be inconsistent. [...] Our corpus consists of 628 German novels mainly from the nineteenth century [...]. The novels have been manually labeled according to their subgenre after research in literary lexica and handbooks” (Hettinger et al. 2016). Kim et al. also explain the provenience of the genre labels in their investigation of the prototypical emotion developments in literary genres: “We collect 2113 books from Project Gutenberg that belong to five genres found in the Brown corpus [...] namely adventure (585 books), romance (383 books), mystery (380 books), science fiction (562 books), and humorous fiction (203 books). [...] The selection is based on the Library of Congress Subject Headings in the metadata” (Kim, Padó, and Klinger 2017). A case in which the provenience of the generic assignments to the texts remains implicit is Schöch (2017c). Schöch analyses subgenres of French Classical and Enlightenment Drama by applying topic modeling to a corpus of plays that was initially curated by Paul Fièvre (called the “Théâtre classique” collection). The data which is used in the analysis is presented in detail, and the subgenres are also mentioned, but it is not made explicit where the labels that are finally used come from: “detailed metadata has been added to the texts relating, for instance, to their historical genre label (e.g. comédie héroique, tragédie, or opéra-ballet) as well as the type of thematic and regional inspiration [...]. A large part of this information can fruitfully be used when applying Topic Modeling to this text collection. [...] Finally, all the plays included belong to one of the following subgenres: comedy, tragedy or tragicomedy” (para. 9–10). Looking into the TEI collection on GitHub (e.g., https://github.com/cligs/theatreclassique/blob/master/tei/tc0001.xml, accessed November 28, 2020), it can be noticed that there is a general subgenre assignment in the TEI header (Comédie), but no further information about its provenience is given. As Schöch mentions the historical subgenre labels in his text, it can be assumed that these are the source. The classification of works into dramatic subgenres is probably less debated than into subgenres of the novel. Likely, dramatic subgenre labels are also more often explicitly given on title pages than in the case of novels. Still, it would be better to make the generic convention that is analyzed more explicit because it makes a difference whether labels assigned by librarians or literary historians or genre labels from the historical paratexts of the works are used. Despite all good advice and intentions to analyze genres or subgenres only on a defined discursive level or on the basis of homogeneous sources of subgenre labels, especially large-scale digital analysis using hundreds or even thousands of texts have to face challenges in defining which generic conventions they refer to. Even in qualitatively oriented genre analysis, selecting works for a corpus that aims to cover one or several specific genres is not trivial. A starting point using either certain labels or definitions of the genre(s) has to be found.An overview of proposals that have been made for corpus building in literary genre studies is given in chapter 3.3 (“Text Corpus”) below. Beyond that, some strategies of text selection and genre assignment are not viable for very large corpora. It is, for example, not possible to read every text and check it against a genre definition that relies on qualitative textual features, i.e., characteristics of the texts that are not (yet) easily formalized and computationally analyzable. Furthermore, it is very likely that large corpora also cover lesser-known texts which critics have not considered yet. That way, existing critical approaches may only cover part of a text corpus. On the other hand, depending on the kind of genre, explicit labels on historical editions may also not be the norm. These are additional difficulties that quantitative genre analysis faces in defining of its object of investigation through the selection and preparation of the text collection. Such challenges make it even more important to clarify which genre convention is addressed and how this is done.

Besides assigning genre labels to the texts, another fundamental point for a digital genre analysis is the selection of textual features. In the end, the normative facts that can be found in the texts, that can be related to genre conventions, and that can be used to establish definitions of genres, depend on which kind of textual material is analyzed. There are different opinions regarding the importance of which kind of features are selected. Underwood, for example, highlights the predictive power of statistical models, which is based on specific features but not directly dependent on them: Leo Breiman has emphasized that predictive models depart from familiar statistical methods (and I would add, from traditional critical procedures) by bracketing the quest to identify underlying factors that really cause and explain the phenomenon being studied. Where genre is concerned, this means that our goal is no longer to define a genre, but to find a model that can reproduce the judgements made by particular historical observers. (Underwood 2016, 5–6)

Taking the example of science fiction, Underwood explains that a very reliable textual clue for this genre are adjectives of size such as “huge” or “tiny” and that a set of some more hundred words would be enough for a statistical model to recognize instances of the genre. Still, he argues, these genre markers do not need to correspond to any definition of the genre that has been formulated so far, and they might not lead to any definition that could be articulated verbally in a useful way (Underwood 2016, 6). Underwood’s stance towards selecting textual features is characterized by a reproductive strategy on the one hand and an explorative one on the other. It is a reproductive strategy in the sense that text analysis is used to replicate social constructions of genre to find out about their general relationship to the textual basis. It is explorative in that the kind of features used is not defined in a top-down approach and controlled in advance, but tested as for their reproductive relevance: “To put it more pointedly: computational methods make contemporary genre theory useful. We can dispense with fixed definitions, and base the study of genre only on the shifting practices of particular historical actors – but still produce models of genre substantive enough to compare and contrast. Since no causal power is ascribed to variables in a predictive model, the choice of features is not all-important” (6). Following this approach, the normative facts found as traces of social constructs of genres would not lead to descriptions of them in scholarly terms, at least not to definitions focusing on the kinds of facts found.

A different view on the question and relevance of feature selection is formulated by Jannidis, who outlines a set of general methodical working steps for computational text analyses: “1. Thesenbildung, 2. Bestimmung der Indikatoren, 3. Korpuszusammenstellung, 4. Korpusvorbereitung, 5. Suche, 6. Quantitative Erhebung, 7. Überprüfung von Indikatoren und Korpuszusammenstellung sowie Diskussion der These im Licht der Ergebnisse” (Jannidis 2010, 110). In this setup, the choice of indicators is directly linked to the formulation of an initial thesis and is more driven by prior theoretical assumptions than in Underwood’s approach.However, Jannidis himself comments on the status of the work steps: “Diese hier skizzierte Vorgehensweise ist natürlich stark idealisiert. Nicht selten steht am Anfang nicht die These, sondern ein auffälliger Befund in Texten, der dann als Indikator für eine These gedeutet wird. Doch selbst in diesem Fall einer induktiven Vorgehensweise ergibt sich zuletzt eine ähnliche Forschungsstrategie, wie hier skizziert” (Jannidis 2010, 110). It represents a deductive procedure. In the case of genre analysis, a genre-related thesis would need to be formulated, for instance: “In social novels, there are more different characters than in sentimental novels”. To be able to verify or falsify the hypothesis through computational text analysis, it would be necessary to define textual indicators representing the concepts mentioned in the hypothesis. The above example would require an approach to identify characters in the text, for example, by detecting mentions of character names and other linguistic references to characters and resolving to which character they point. It would be necessary to automatically detect the set of different characters in a novel, which is a difficult task. Somewhat easier to formalize and closer to a stylistic analysis would be a hypothesis such as “In social novels, there are more mentions of different character names than in sentimental novels”. Like the explorative procedures, also top-down approaches have several advantages and disadvantages. What is good about them is that they start directly from the scholarly field that is also the target context. If the goal is to find out something about literary genres and the hypothesis is formulated in literary scholarly terms, the hypothesis is compatible with the epistemological frame of the investigation. In addition, the selection of textual indicators and features can be motivated theoretically so that meaningful and interpretable results can be expected. The main disadvantage is that the possibility of formalizing the hypotheses depends on the available technical methods. Although research is done in this direction, many literary theoretical concepts still need to be formalizable.For an approach to automatically recognize characters in German language novels, see Jannidis et al. (2015). Barth and Viehhauser (2017) made the first attempts to formalize concepts of literary space. Existing text mining methods, for example, topic modeling or sentiment analysis, may also be used to operationalize the hypotheses. Then it must be explained in what way they can be considered formalizations of literary theoretical concepts, such as themes, topoi, or emotions. In any case, in such an approach, the selection of specific textual features is not at all arbitrary or negligible. The features represent the texts and are assumed to cover stylistic aspects that are traces of generic conventions in the texts. The chosen indicators must be suited to check the plausibility of the literary theoretical hypothesis. At the same time, the choice of the indicators themselves is based on assumptions: “Das Verhältnis zwischen Indikatoren und These ist allerdings in vielen Fällen keineswegs selbstverständlich, sondern hat selbst hypothetischen Charakter” (Jannidis 2010, 116). Even in hypothesis-driven digital genre analysis, the suitability of the features needs to be tested empirically to some degree.

In the same way as the kind of generic convention that is analyzed and the selection of corresponding genre labels, the choice of textual features for a digital stylistic genre analysis should also be motivated. How specific the chosen features are and how important it is to clarify their relationship to literary theoretical concepts depends on the kind of strategy that is chosen for the genre analysis: it can be primarily deductive, inductive, or experimental, and it can be theoretically or historically oriented. In principle, digital genre analysis can be used for all kinds of investigations. The goal can be, for instance, to find out if and in what way works with specific historical or critical genre labels are textually coherent, as Underwood did. Another goal can be to test if a specific scholarly definition of a genre holds when it is formalized and applied to a corpus of texts that have been assigned to the genre in question. The results of a stylistic genre analysis could also be used to formulate new, statistically-based definitions of genres. Even if textual variables in statistical models do not necessarily reflect causal relationships, their distribution can be interpreted to reach empirically based genre definitions if genres can be distinguished based on these variables. A sentence in such a definition could be, for example, “the genre X is defined such that the probability of a love topic is significantly higher in texts that participate in the genre X than in texts that do not participate in this genre”. In Hempfer’s terms, the normative fact that was found is the different probability of a certain kind of topic in two groups of texts that are associated with different genres by convention. The genre is constructed in the interaction of the scholar who decides which textual features to use and which texts to analyze on the one side and the texts themselves on the other. Moreover, the scholar has to interpret the topics and decide that one of them can be described with the term “love”. Furthermore, it has to be decided what “significantly higher” means. A definitory phrase such as the one above can itself be used as a hypothesis that can be tested in other empirical settings, for example, with a different corpus of texts. To what extent the found textual characteristics of exemplars of different genres actually correspond to conscious social norms can only be clarified by analyzing contemporary or historical discussions about what the genres in question are. That would not be different in non-computational text analysis. In addition, besides starting from known generic conventions or scholarly definitions of genre, a digital stylistic genre analysis can also start from the texts themselves. For instance, a corpus of texts can be built for a certain period and a specific cultural, geographical, and linguistic context. It can then be analyzed which groups of texts emerge as textually coherent when specific textual features are used. Such an approach would allow for the possibility of detecting faits normatifs as signs of communicative patterns that might not have had a high degree of explicitness. They might not have been frequently named or broadly discussed in the historical context, and possibly they have not been described yet in scholarly terms. In such cases, it would as well be possible to complement the quantitative analysis with a qualitative study of intertextual links, of implicit or explicit signals in the texts and paratexts, and of metatextual statements that could substantiate or question the communicative relevance of new findings of text groups.

When stylistic text features are interpreted as signs of generic conventions, a difficult point is how clearly the relationship between both characteristics of texts can be established: having certain features or a specific distribution of them on the one hand and participating in a particular generic convention on the other. In this context, only the causal relationship between the genre label and the textual features is meant, not the question of the kind of social, generic norms and their unconscious or conscious application. In many cases, literary-historical studies of genres focus only on one, positively described genre.In the context of subgenres of Spanish-American novels, see, for instance, Gnutzmann (1998) and Schlickers (2003) on the naturalistic novel, Löfquist (1995) and Read (1939) on the historical novel, and Rivas (1990), Rosell (1997), and Sparrow de García Barrío (1977) on the anti-slavery novel. Often subtypes of one genre are distinguished as part of the investigation, especially if a major genre is concerned,Suárez-Murias (1963), for instance, is a study of the Spanish-American romantic novel by country, in which subtypes of the romantic novel are presented, such as sentimental novels, historical novels, or novels of customs. A comprehensive overview of thematic subtypes of Spanish-American novels is given in Sánchez (1953). In Molina’s (2011) study of the early nineteenth-century Argentine novel, four classes of novels are established (“novela histórica”, “novela política”, “novela socializadora”, “novela sentimental”). but explicitly contrastive studies are rare.The tradition to directly contrast genres is stronger in linguistics than in literary studies. For text linguistic contrastive genre analyses, see, for example, Adamzik (2001), Danneberg and Niederhauser (1998), Gnutzmann (1990), Kaiser (2002, 2008), and Theisen (2016). In literary studies, contrastive analyses are, in particular, used in comparative studies on different cultural and linguistic literary systems. See Lamping (2010) for an overview of comparative genre studies and Jacobs (1986) for a case study on the picaresque and the education novel. In corpus-based and empirical digital genre stylistic analysis, it is more usual to directly contrast different genres or subgenres to find distinctive features for each group (see, for instance, Schöch 2018 and Schöch et al. 2018).The concept of distinctiveness or keyness, which aims to find words characteristic of one group of text compared to another, is a general one, not limited to analyses of genre. See Burrows (2007), who developed the Zeta-measure for questions of authorship, and Scott (1997) for a general approach to keyword extraction. Approaches based on statistical classification also bring forth features that are decisive for distinguishing different classes of texts.For an analysis of the text features that are decisive in detecting the nationality of authors of Spanish language novels, see, for instance, Zehe et al. (2018). Sentiment features that are important in the classification of subgenres of nineteenth-century Spanish-American novels were explored in Henny-Krahmer (2018). If only the characteristics of one genre are worked out, one cannot say with certainty that there are not other genres that share part of these characteristics. This possibility is avoided in contrastive studies. On the other hand, the results of comparative approaches obviously depend on what is compared. A contrastive text analysis aiming to define the sentimental novel, which is based on a corpus of sentimental, historical, and adventure novels, will lead to a definition of the sentimental novel that is relative to the other subgenres. If sentimental novels were instead compared to science fiction and crime novels, different aspects might be decisive in their distinction and recognition. The choice of the corpus that is used to determine a genre in the context of other genres is, therefore, a central aspect of digital stylistic genre analysis. In the case of contrastive analysis, the relationship between the textual features and the genre in question is established relative to other genres on which the outcomes depend.

Another challenge in this regard is that text style is not only a function of genre. All kinds of intra- and extra-textual phenomena shape the style of a text. A special awareness of this circumstance has developed in the field of authorship attribution, where it was repeatedly noted that, for example, the period a text was written and published in but also its genre interfere with authorship signal.Oakes mentions the following issues that can mask the individual authorial style and make it challenging to attribute texts correctly to a particular author: heterogeneity of authorship over time, genre, gender, variation within a single author, and topic (Oakes 2009, 1072–1073). In the same way, authorship, time period, and other factors can also obscure the link of textual features to genre.An attempt to neutralize authorial signals in genre analysis has, for example, been made by Calvo Tello et al. (2017). It may thus happen that conclusions are drawn about genre that are instead due to unrecognized effects of other variables. Considering the different discursive levels on which generic terms and genres can be defined, it is also likely that these different levels may correlate or interfere with each other. For example, in the bibliography of nineteenth-century Spanish-American novels that was prepared for this dissertation, there are 172 novels with the primary thematic subgenre “novela sentimental”. For 85 of these, the literary current to which they belong is unknown. 72 have been associated with the romantic current, 15 with the realist, eight with the naturalistic, and two with the modernist current.The sum of the numbers related to currents is higher than 172 because the same work can be associated with several literary currents. The bibliographic data containing this information is available at https://raw.githubusercontent.com/cligs/bibacme/master/app/data/works.xml, and the script used to retrieve the numbers of sentimental novels can be viewed at https://github.com/cligs/scripts-nh/blob/master/concepts/subgenre-label-combinations.xsl. Both links accessed November 29, 2020. Putting aside the cases of unknown literary currents, the numbers suggest a strong correlation between a primarily sentimental theme and the romantic current. Definitions of the sentimental novel derived from this set of novels will therefore be strongly influenced by the period in which the romantic current was the dominant aesthetic program.

If there are undesired factors that influence the target style that is analyzed (e.g., the influence of period on genre style if genre is the primary concern), then several strategies are possible to cope with such factors. A crucial aspect is the construction of the text corpus that is used for the genre analysis. For example, it is possible to include only one text or an equal number of texts per author. This prevents the results from being too much influenced by authors that were very productive writers of texts that can be attributed to specific genres. However, in most cases, creating a balanced corpus means moving away from a text collection that represents the historical proportions of works according to specific criteria. It means that one tries to create a synthetic setting that can be used to get results that are free from unwanted influencing factors in order to reach a definition of the subject that is theoretically “clean”. This may be very difficult if there are not enough sources for such a corpus. Going back to the Spanish-American sentimental novels, taking an equal number of romantic and realistic sentimental novels would reduce the number of novels to analyze to 30 instead of 172. Besides controlling external factors through corpus building, another possibility is to try to choose the text features in such a way that they are likely to be relevant for specific generic distinctions but not for other kinds of differences. However, it has so far not been possible to isolate features that are only connected to a single extra- or intratextual influencing factor.In a presentation about the differentiation of authorship, form, and genre of literary texts, Schöch and Pielström tried to identify statistical components that are clearly attributable to either of these factors. Analyzing French dramatic texts, they found two components that primarily covered differences in authorship but none that were predominantly related to genre (Schöch and Pielström 2014a, 2014b). A third way would be to model the factors that are assumed to influence the target variable, for example, by collecting corresponding metadata and using this information when the analysis results are inspected. In principle, a corpus that is created by random sampling may represent historical imbalances. However, it would still be possible to estimate how much influence other factors have on the stylistic features that are interpreted in terms of the genres that are investigated. The decision for a specific strategy to control different factors that may influence the text style may depend on the aim of the study. The wider the claim of validity is for the results that are reached, i.e., the more independent they should be from contextual determinations and intra-textual correlates, the more important it is to build a corpus and features that are theoretically adequate. The narrower the scope of the genre study, the more it will be acceptable to have a corpus and feature set that is influenced by the specific historical setting, and that leads to a less theoretical but more organic description of the genres in question. Looked at another way, the stronger the theoretical claim on the results of a corpus-based study, the more important it is to control for possible factors influencing the target variable of the study, i.e., literary genres. On the other hand, a primary interest in historical adequacy can be pursued by considering and investigating influencing factors, but not necessarily controlling them, for example, by balancing a corpus of texts. Awareness of influencing factors is still indispensable in both cases to make sure that the assertions made are about genre at all.

It is clear that the ontological status of genres as conventions or norms of communication or social interaction – if they are understood in that way – makes the access that digital genre stylistics has to them one that is mediated by text style. Text style, in turn, is influenced by a number of factors other than genre. These influence factors are never captured as a whole but only on selected levels of textual features that are chosen for analysis. One could say that genre “hangs by a thread” for digital stylistics, but in general, in the debate about the status of genres, the field can be characterized as inclined towards the realistic side. How strong the connection between genres, genre labels, and genre signals, on the one hand, and common features of text groups, on the other hand, is, can be analyzed in detail with digital text analysis. It is to be expected that the results may be quite different depending on the kinds of genres and the literary-historical contexts that are investigated. Quantitative approaches lend themselves very well to analyzing big corpora of formula fiction, that is, popular genre fiction for which uniform patterns of style are expected. In such a research setting, genres are probably more tangible than in an analysis of highly canonized, individualistic works of art where generic references may be weaker.Jannidis, Konle, and Leinen (2019), for example, analyzed a corpus of 9,000 dime novels in German language, which were published between 2009 and 2017. They aimed to find out how the subgenres of the dime novels can be differentiated and in what way the corpus as a whole is different from high-prestige novels. They used the 8,000 most frequent nouns as features and classified the novels with Logistic Regression. In addition, a clustering was done on the basis of the 2,000 MFW. To analyze the relevant features for the different subgenres, topic modeling and a contrastive analysis with the Zeta measure were performed. They found that the subgenres can be distinguished well both on a stylistic and a thematic level. Regarding the complexity of dime novels compared to high-prestige literature, they found that sentences are shorter in dime novels. However, they did not find any clear differences in the vocabulary richness or length of the words. Furthermore, as was pointed out above, digital genre stylistics is a field that is closely linked to computational linguistics, and the existence of linguistic text types is less questioned than the one of literary genres, mainly because it is more evident that the text types constitute communicative norms.

There are further aspects that are as well relevant to the relationship between literary genre theory and digital genre stylistics and that have not been covered yet in this discussion of the role of corpora, genre labels, features, and text style. One aspect is the relationship between genre systems and their history, a field of tension that is also linked to terminological questions. Another aspect is the debate about the type of categories that genres can be conceived as. These issues will be touched upon in the next chapters.

In digital genre stylistics, it has sometimes been proposed that one should not speak about the analysis of “genres” in this case, but of “text types”. This is because of the focus of digital genre stylistics on the literary texts themselves and more specifically on their linguistic surface and style, and the relatively limited inclusion of information that is related to the literary-historical and social context. At the same time, genre is a concept that appears to be strongly influenced also by extra-textual historical factors. On the one hand, the terminological distinction between genres and text types has its origin in linguistics, where literary genres are differentiated from linguistic text types. On the other hand it has also been proposed in literary genre theory itself as a means of distinguishing between genres that are based purely on textual and linguistic criteria and historical genres. This is similar to Genette’s proposal to differentiate between mode and genre. In the context of text linguistics, text types (in German, “Texttypen”) are described as follows: Texte werden zu Texttypen zusammengefasst auf der Grundlage linguistischer Kriterien. Texttypen verlaufen quer zu den Textsorten in verschiedenen Kommunikationsbereichen. Als linguistische Kriterien gelten dabei textinterne Merkmale (Merkmale der Textinfrastruktur) wie Stil (Stiltyp, z.B. Ironie, Nominalstil), Medialität (medialer Typ, z.B. digitaler Text, konzeptionell mündlicher Text), Textfunktion auf der Basis sprachlicher Indikatoren (Funktionstyp, z.B. Kontakttext), Themenentfaltung/Vertextung (Vertextungstyp, z.B. explikativer Text). (Gansel 2011, 13)

Style is directly mentioned as a defining characteristic of text types in the linguistic sense. However, the linguistic term that is used in a way that can be compared to the general conception of literary genres is also “text type” (in German, “Textsorte”): Wir definieren Textklasse als das Vorkommen einer Menge von Texten in einem abgegrenzten, durch situativ-funktionale und soziale Merkmale – also textexterne Merkmale – definierten kommunikativen Bereich, in dem sich Textsorten ausdifferenzieren. [Z. B. Textklasse] Religion – [Textsorten] Predigt, Ordensregel, Enzyklika oder [Textklasse] Politik – [Textsorten] Koalitionsvertrag, Parteiprogramm, Regierungserklärung. (Gansel 2011, 12–13)

In the latter sense, which is predominant in text linguistics since the “pragmatic turn”, the linguistic text types are also primarily determined by text-external factors that characterize the communicative situation, and they are understood as norms and conventions, which influence how texts are produced and received (Gansel 2011, 8–10; Krieg-Holz and Bülow 2016, 220). These two uses of the term “text type” should thus not be confused. The terminological differentiation between pragmatically determined text types and syntactically and semantically distinguishable ones is also recognized in computational linguistics, where it is labeled as the opposition between genres and text types. It is probably through this influence that the conceptual separation between genres and text types has also been taken up in digital genre stylistics. In his computational study of linguistic genre variation, Biber, for example, declares: I have used the term ‘genre’ (or ‘register’) for text varieties that are readily recognized and ‘named’ within a culture (e.g., letters, press editorials, sermons, conversation), while I have used the term ‘text type’ for varieties that are defined linguistically (rather than perceptually). Both genres and text types can be characterized by reference to co-occuring linguistic features, but text types are further defined quantitatively such that the texts in a type all share frequent use of the same set of co-occurring linguistic features. Because co-occurrence reflects shared function, the resulting types are coherent in their linguistic form and communicative functions. (Biber 1992, 332)

Biber says that also genres are related to the occurrence of specific linguistic features, but in a way that is less consistent than in the case of functionally determined text types. Biber determines the text types in a bottom-up approach by using factor analysis and interpreting the resulting dimensions in terms of linguistically expressed communicative functions. Biber finds five dimensions of variation (informational versus involved, narrative versus non-narrative, elaborated versus situated reference, overt expression versus persuasion, and abstract versus non-abstract style) on which he bases his definitions of text types (Biber 1992, 334–335, 339–340). In their computational linguistic approach to the detection of text genre, Kessler, Numberg, and Schütze, on the other hand, only use the term “genre”, which they define broadly as encompassing both literary and non-literary texts: “We will use the term ‘genre’ here to refer to any widely recognized class of texts defined by some common communicative purpose or other functional traits, provided the function is connected to some formal cues or commonalities and that the class is extensible” (Kessler, Numberg, and Schütze 1997, 33). Both Biber and Kessler, Numberg, and Schütze recognize that genres are difficult to grasp on the formal linguistic level, but they expect (as Biber) or require (Kessler, Numberg, and Schütze) that some common formal elements can be found for texts that have been associated with the same genre.

In linguistics and computational linguistics, the discussion of the distinction between genres and text types focuses on the differences between conventional and pragmatic characteristics on the one side and structural linguistic features of text groups on the other side. In literary genre theory, in contrast, the point of debate in this terminological question is more oriented towards the systematic versus historical nature of the objects. Fricke, for instance, argues that it is not necessary to abandon neither the function of generic terms as names for historical groups nor their use as classificatory terms but that they should be distinguished. He differentiates a “literary text type” (“literarische Textsorte”) from a “genre” and defines the first one as a purely systematic term to categorize literary texts and the latter as a term for historically bound and delimited literary institutions. To determine if a text belongs to a text type, its grammatical, semantic, and textual-pragmatic functions must be analyzed (Fricke 1981, 132–133). According to Fricke, more criteria need to be fulfilled in order to attribute a text to a genre. First, the text needs to conform to a clearly distinguishable literary text type. Second, the literary text type has to be established in the national literature of the text in question when the text is created, so that it corresponds to the expectations that contemporary readers have regarding the characteristics of the literary text type. Third, the text needs to explicitly carry an established name of the literary text type or exhibit other established signals for it. Finally, Fricke also uses the term “Gattung” (German for “genre”) as a general term that can be used to both designate literary text types, genres, or any other establishment of groups of literary texts (Fricke 1981, 132). Fricke’s proposal to speak of literary text types is fruitful for cases in which literary texts are classified based on syntactic, semantic, and pragmatic linguistic properties or any other textual characteristics without establishing any direct connection to a specific historical genre discourse. This use of the term “text type” is congruent with the definition of “Texttyp” in general linguistics and “text type” as used by Biber. However, Fricke’s definition of “genre” is very strict. According to his concept, genres are a subset of the text types that they correspond to, because a common text type is a prerequisite. In addition, the text type has to correspond to contemporary expectations, which means that only generic conventions that are realized as distinguishable text types actually represent genres. Fricke does not say how he would call generic expectations or norms that are not mappable to specific combinations of textual features in a consistent way. Moreover, distinctions of literary genres made after the text’s creation, whose participation in a genre is analyzed, are not called genres if the genre concept had not been established before. For example, the gaucho novel would not be a genre of nineteenth-century Spanish-American novels if it could not be verified that the term or the concept already existed as a norm that was conscious to the historical authors, publishers, readers, and critics. Furthermore, texts that are similar to others because of linguistic criteria but do not carry explicit or established implicit generic signals would also be considered as not participating in a genre. It makes sense not to assume a common genre simply because of textual similarities – one could, for example, find similar verse structures in the poetry of different centuries and continents by accident, for which no historical relationship could be verified. Nevertheless, if a text was created in the same context as a group of other texts that carry an explicit genre name and they all share textual features, one might argue that the non-labeled text also participates in the genre. Determining where the same context begins or ends is another question that needs to be solved. One can also not exclude the possibility of cases of generic references spanning different temporal, geographical, or linguistic contexts, even if this is not the usual case. A wider definition of genre would be more appropriate for the purpose of digital genre stylistics. Such a definition should be able to relate conventional or critically established generic groupings of different kinds with findings of groups on the textual level without restricting the kinds or relationships.

Examining Fricke’s definition of genre further and following his explanations, even his definition of text types appears too narrow for digital genre stylistics. In requiring the congruence of texts that share the features of a certain text type and that belong to a certain genre convention, Fricke already assumes that the text types are linked to generic expectations. Any arbitrary text classification could not be expected to be congruent with the genre concepts if not by accident. Thus Fricke requires a precise and systematic definition of a text type that can serve as a starting point for an empirical study that aims to verify if the text type was institutionalized as a genre in some historical period (Fricke 1981, 133). The text types should not be defined arbitrarily but according to their scholarly appropriateness: “Um ihrer heuristischen Eignung willen wird man die Textsortenbegriffe folglich in der Regel so festlegen, daß sie aufgrund hypothetischer literaturgeschichtlicher Vorannahmen voraussichtlich auf historisch belegte Texte zutreffen und daß nach Möglichkeit sogar irgendwann einmal ein dieser Textsorte entsprechendes Genre bestanden hat” (Fricke 1981, 134). Defining text types beforehand that probably correspond to historical genres requires a deductive procedure starting from assumed prior definitions of these genres. As was already discussed in the previous chapter, digital genre analyses can be conducted in a number of different ways: they can start with the formulation of theoretically based hypotheses about genres, which are then formalized, but they can also start from genre labels that stand for conventional ideas of genres and directly test if and how they relate to textual characteristics, without following a continuous chain of formalizing steps. The hypotheses about the relevance of specific textual features for genres can be quite loose. A strength of the computational approaches is specifically that series of different hypotheses can be tested and also that unforeseen results can be achieved through experimentation. In many cases, the formalization of literary concepts is not so mature that it would allow for continuous deductive procedures. Consequently, text types found through digital stylistic analyses do not necessarily have to be in line with literary-theoretical text types or historical genres. Rather, what can be looked for are the points of intersection between generic conventions and groups established based on textual features. They can be approximated to each other, but it should not be expected that they depend on each other or that there is a relationship of inclusion between them. I propose to widen the understanding of genre, but in turn to differentiate between two kinds of genres to establish a connection to text types. For the purpose of digital genre stylistics, as it is conducted in this dissertation, the following working definitions are proposed: A literary text type is an intensional, systematic term used to designate groups of literary texts that are established on the basis of common or similar immanent textual features and feature distributions of any kind and that can be distinguished from other literary text types based on these features and feature distributions. A conventional literary genre is a term referring to the extension of genres as historically bound literary institutions (in the sense of Voßkamp 1977) and as codifications of discursive expectations towards literary texts that participate in the genres (in the sense of Jauß 2014). The term “conventional” is used in a broad sense here and refers to historical as well as modern conventions. It includes socially and communicatively established genre conventions but also conventions of how texts are to be systematically assigned to genres, e.g., by librarians or literary scholars, because considered on a large scale and across time, systematic approaches are also historically bound. A textual literary genre is the convergence or intersection of a conventional literary genre and a literary text type. A literary text type and a conventional genre can be congruent to a certain degree, depending on the extent to which the groups of texts participating in them coincide. If the perspective starts from the conventional genre, it may be textually coherent to a certain degree. This means that a certain percentage of the texts that are attributed to the conventional genre are also part of a corresponding text type. With a correspondence of more than 50 %, one can speak of a certain textual coherence of the conventional genre. A text can participate in several conventional genres, and it can also belong to several text types. As a consequence, a text can be associated with no, one, or several textual genres.

In what follows, these determinations are explained and justified in more detail. With the differentiation between conventional and textual literary genres, the genres can be both detached from specific textual features or linked to them. To find out about the textual literary genres in which literary text types and conventional literary genres overlap is then a central task of digital genre stylistics. All kinds of relationships between conventional genres and text types can be assumed. For example, an extreme case would be a generic norm that has been discussed in poetological writings and can be described as a conventional literary genre, but that has never been realized on a textual level through a set of literary works. At the other extreme would be a text type that is recognizable in terms of recurrent patterns of feature distributions and that can be distinguished from other text types based on these features, but that has never been named or discussed as a conventional genre. Between these two extreme cases, many other constellations are possible, for example, groups of texts that carry signals of a particular conventional genre but of which only a part is coherent on the textual level. Another possible case is a conventional genre, which is held together by a common name but which relates to several different text types so that it can be described as several different textual genres. Such a result could stimulate further investigation into the conventional genre to see if there are signs of several subtypes.An example of such a constellation is the German novella, which Schröter described as an instance of “historically discontinuous and heterogeneous genres” (Schröter 2019, 227). Rigorous poetics of the novella existed, but these were not consistent with the texts called novellas. Furthermore, the texts carrying the label “novella” could not be described as a homogeneous group (Schröter 2019, 229). Schröter proposes to define “intersections” between the texts referred to as novellas and classificatory sets of texts so that, for example, texts with the name novella that were published as fictional journal prose and share the characteristics of the latter type of text would form one subtype of the genre novella. Schröter states: “It is important not to summarily define such intersections as novellas, as is often done. Such a definition would result in a classificatory concept of the respective genre, which in turn would no longer be suitable for comprehending the historical use of the generic label and hence the semantics of a genre in literary-historical communication.” (Schröter 2019, 228). Using the terminology proposed here, the intersection of texts referred to as novellas and fictional journal prose would be called a “textual literary genre”, whereas the semantics of the novella in literary-historical communication would be referred to as the “conventional literary genre”. The novella, as a conventional literary genre, would then consist of several different textual genres. Also possible are cases where two different conventional genres relate to the same text type. For example, it could be analyzed if different terms of shorter narrative prose in the Spanish-American nineteenth-century tradition that do not seem to be limited very clearly as conventional genres (“narración”, “relato”, “cuento”, “novelita”) have a common textual basis. Three of the many possible constellations between text types, conventional genres, and textual genres are illustrated schematically in figure 1.

The first case on the left side is one where a conventional genre and a text type overlap to a certain degree so that one textual genre can be identified. The parts of the conventional genre that lie outside the textual genre refer to literary texts that have been marked as being part of the convention but that do not conform to the text type. On the other hand, there are texts whose characteristics correspond to the text type but that have not been assigned to the conventional genre. In the three examples shown in the figure, there is no relationship of inclusion between the conventional genre and the text type, but this would, of course, also be possible. In such a case, all the literary texts that conform to the text type could also be part of the conventional genre, only that some other texts that are associated with the conventional genre are not congruent with the text type (inclusion of the text type in the conventional genre), or all the texts that participate in the conventional genre are in line with the text type, and, in addition, there are texts that fulfill the criteria of the text type but that have not been marked as belonging to the genre by convention (inclusion of the conventional genre in the text type). The second case in the middle of the figure illustrates a relationship of overlap between one conventional genre and two different text types, so that two textual genres are identified. In such a setting, it is conceivable that one genre name is used for different text types that can be distinguished either synchronically or diachronically. The third case to the right shows how two conventional genres, which are, for example, characterized by two different genre names, can be covered by the same text type. Here, too, there result two different textual genres. An example of this constellation would be if the same textual characteristics are found in different historical contexts and different genre names are used in these contexts to refer to them. In the schematic cases shown here, only up to two conventional genres or text types are involved, but there could, of course, also be scenarios where more parts are involved.

What is the status of generic signals that relate literary works to genres in this setup? Explicit genre names, implicit textual genre signals such as conventionalized titles or names, or beginnings of the textsFor a short description of the usual types of generic signals, see Fricke (1981, 135). A more comprehensive overview is given in Fowler (1982, 88–105). are understood as belonging to the level of conventional genres. The presence of such signals does not necessarily mean that a text conforms to a specific textual genre, and even their absence does not mean that a text cannot be described in terms of the same text type that is connected to a textual genre. An unconventional use of generic signals, for example, a subversive or an ironic one, can be detected when there is a discrepancy between the individual text carrying the signal and other members of a textual genre.

A question that needs to be addressed is how assignments of literary texts to genres that are not established by paratextual or textual signals but text-externally relate to the three terms defined above. Part of this question is how generic terms and definitions of genres, on the one hand, and the text types, conventional, and textual genres, on the other hand, behave in relation to the literary system, literary theory, and literary history. As Schaeffer notes: “les termes génériques on un status bâtard. Ils ne sont pas de purs terms analytiques qu’on appliquerait de l’extérieur à l’histoire des textes, mais font, à des degrés divers, partie de cette histoire même” (Schaeffer 1983, 65). What Schaeffer means is that also terms that are defined externally are part of the history of genres. Very important is his remark that generic terms do so to different degrees. Furthermore, the degree also depends on the perspective. If a nineteenth-century Mexican editor labels several books that are published for the first time as “novelas”, he applies this term collectively to several works. He thus has an understanding of the genre that already involves a systematic element – he compares the books to his concept of the genre novela and decides if they fit into the category or not. However, if we analyze these nineteenth-century novels today, we perceive the editor’s decisions as part of the history of the genre and his opinion as formed by the convention of the time. The labels would therefore be signs of the conventional genre. In addition, the understanding that an editor has of a genre is probably not much formed by theoretical considerations. If, instead, a contemporary writer expresses his views on what the novela is in poetic terms and lists several works that conform to his definition of the genre, this, too, would represent a systematic and possibly also theoretically motivated approach to the genre.A Mexican writer who reflected on the novel’s role in Mexican national literature was, for example, Ignacio Manuel Altamirano, who expressed his ideas in the essay “Revistas Literarias de México” (Altamirano 1868). To his or her contemporaries, the initiative could appear as not being part of the genre convention but rather as an attempt to define a textual genre. For the literary historian, again, it would clearly contribute to the genre convention of the time, even if the historical, poetic definition of the genre is by degree more systematic and theoretical than the practice of individual writers or editors who provide the texts that they write or publish with generic labels. Yet another case is that of a modern librarian who classifies historical texts by genre. If one librarian does this, a unique concept of the genre (the one the librarian has) is applied collectively to a set of texts. If several librarians classify the texts together, multiple genre concepts are involved.The terms “collectif” versus  “individuel” and “unique” versus “multiple” are introduced by Schaeffer when discussing the status of different kinds of generic terms. Furthermore, Schaeffer distinguishes between “noms génériques endogènes”, which are used by authors or their public, and “nom génériques exogènes”, which are established by literary historians. The generic terms can have a textual (attached to the literary text, for example, as a paratextual element) or a meta-textual status (if they are used as terms to discuss a work but are external to it), and the functions of the labels vary depending on the category they belong to (Schaeffer 1983, 65, 77–78). The labeling that the librarians do is not part of the historical genre convention of the context in which the literary texts were initially created and published. Instead, it is part of the modern genre convention, even if that convention has a systematic background and is informed by literary studies. This is the more obvious the more different people participate in applying a specific concept of genre. Finally, literary-theoretical or literary-historical positions can also be perceived as conventional if they are analyzed collectively. When the genre assignments of twenty different literary scholars are examined together, it is not very probable that they all share the exact same theoretical understanding of the genres in question. The sum of their judgments constitutes itself a theoretically based genre convention. Conventional genres can thus involve a systematic and theoretical element.Rather than on the question of historical convention and institutionalization, Schaeffer focuses on the place and concept of the genre names in the communicative situation. He uses the property of all texts as speech acts as an argument for characterizing genres as more analytical or historical: “Je pense qu’il faut aller plus loin: les genres théoriques, c’est-à-dire en fait les genres tels qu’ils sont définis par tel ou tel critique, font eux-mêmes partie de ce qu’on pourrait appeler la logique pragmatique de la générécité, logique qui est indistinctement un phénomène de production et de réception textuelle. En ce sens on peut dire que l’Introduction à la littérature fantastique de Todorov est elle-même un des facteurs de la dynamique générique, à savoir une proposition spécifique pour un regroupement textuel spécifique et donc pour un modèle générique spécifique [...]” (Schaeffer 1983, 68). The more weight that systematic and theoretical element has, the more probable it is that the genre convention largely correlates with a text type. On the other side, historical poetical genre definitions as well as modern literary theoretical ones could also be understood as theoretical entities and be used as hypotheses about textual genres. They could then be tested by applying them to the texts that have been classified as instances of the genres based on the respective genre definitions.Schaeffer too argues that there should not be a strict separation of theoretical and historical genres and genre labels, although he recognizes that they follow quite different rules: “le système des genres théoriques, construit à partir d’oppositions différentielles, simples ou multiples, obéit à des contraintes de cohérence qui ne sont pas celles des genres historiques (quelles que soit la réalité de ces genres désignés par les noms de genres traditionells)” (Schaeffer 1983, 67). Schaeffer demonstrates why it is not useful to assume a direct deductive relationship between both. However, if theoretical definitions of genres are used as hypotheses about textual genres, no claim about the integrity of the conventional historical genres is made, so that such a deductive procedure seems viable. Just as genre conventions are not understood as purely historical, also text types are not conceived as purely systematic or theoretical entities here. Whenever text types are derived from a corpus of literary texts, they involve a historical element because the texts were produced in a specific historical setting. Following Hempfer’s assumption that generic norms and expectations enter literary texts and are readable from them as normative facts, the conventional genres influence the text types. Otherwise, there would not be any textual genres at all. More distant to the historical genre conventions are text types that are defined on general theoretical principles without recourse to specific text corpora, for example, text types that are formulated according to a general theory of language or literature. Similar to theoretically defined genres, these could be used as testable hypotheses with regard to literary text types, i.e. to check to what extent they represent actual textual patterns. When such theoretical text types also refer to generic conventions, they are theories about textual genres. That is, theories of literary textual genres can be more inclined towards genre conventions – if they concentrate on historically bound discursive expectations towards literary texts and relate them to textual characteristics – or they can be inclined towards text types if they start from systematic definitions of text groups and relate them to conventional generic terms.

Which consequences does this terminological system have for digital genre stylistics? First: there is room for experimentation. For Moretti, through digitally-based, formal, and quantitative analysis, literary criticism is transformed into an experiment: a few years, and we’ll be able to search just about all novels that have ever been published, and look for patterns among billions of sentences [...] By sifting through thousands of variations and permutations and approximations, a quantitative stylistics of the digital archive may find some answers. It will be difficult, no doubt, because one cannot study a large archive in the same way one studies as text: texts are designed to ‘speak’ to us [...]  but archives [...] say absolutely nothing until one asks the right question. [...] it turns criticism on its head, and transforms it into an experiment of sorts: ‘questions put to nature’ is how experiments are often described, and what I’m imagining here are questions—put to culture. Difficult; but too interesting not to give it a try. (Moretti 2008, 114)

By not presupposing theoretically how text types should be constituted to match historical genre conventions in the best way and by not demanding a relationship of inclusion between text types and conventional genres, experiments with different kinds of textual features are possible and can lead to new insights about textual genres. One possible analysis scenario is outlined here: generic labels for the texts in a corpus are collected from different modern library catalogs a most frequent words-based classification is used to find out how well the texts can be classified by the genre labels it is determined which features are most important for the textual distinction between the different genres. In this scenario, the goal is to work out the degree of textual coherence of the conventional genres and to define the text types underlying the textual genres based on a general set of linguistic features. Texts that are repeatedly misclassified can be assumed to be part of the genre convention to which they were assigned but not of the textual genre to which the genre label in question is primarily attributed. If such texts are assigned to another class several times by the classifier, it can be checked if they are described better by another text type.

However, the distinction between literary text types, conventional genres, and textual genres does not only allow for testing how relevant certain general textual features are to capture the connections between the levels of text and convention. On the other hand, one can as well start from theories of textual genres. Such an analysis scenario is sketched here as an example. For instance, it would be possible to depart from a definition of the historical novel as a novel that is characterized by the temporal distance between the writing, publication, and reception of the novel and the past in which the narrated events take place, the co-occurrence of invented and historical personages, places, and events, and the localization of the narrated events in a precise historical past.For definitions of the historical novel in which the mentioned characteristics play a role, see, among others, Fernández Prieto (1996), Lefere (2013), Lukács (1955), Maxwell (2009) and Spang (1998). A specific corpus of historical and non-historical novels could be built by using specific kinds of genre labels, for example, indications of the subgenre on book covers. These then function as markers of the conventional genre against which the definition of the textual genre is checked.In this case, it would be important not to derive the labels directly from the theoretical definition of the textual genre in order to avoid circular reasoning. It would then be necessary to formalize the elements that are part of the literary theoretical definition. In this case, it would not be enough to just use the most frequent words because the deductive chain from the genre definition to the textual features needs to be maintained. Instead, named entity recognition could be used to detect the mention of historical figures in the novels, combined with their identification through authority files. The detection of temporal expressions with temporal taggers could be used to check if the historical novels are characterized by more frequent uses of explicit temporal expressions than other subgenres of the novel that are represented in the corpus. As a result, it could be verified how many of the novels that were marked as historical novels by convention are captured with the text type that is derived from the theoretical definition of the textual genre. In this case, a challenge would be to quantify the boundaries between historical and non-historical novels based on the literary theoretical criteria and to find a way to delimit the text type. This is because the criteria are positively formulated and refer to the presence of features (for example, “there is a temporal distance between the publication date of the novel and the narrated past”). However, it is not said how much of the feature should be present (how large should the temporal distance be? 20 years, or 50, or hundreds of years?) and how the same features are distributed in texts that are part of other genres (are they totally absent there – no temporal distance at all – or just less frequent or less strong – only up to 10 years of temporal distance?). This example shows how difficult it can be to directly map literary theoretical assumptions about genres to a digital quantitative genre analysis. In the above example, possible strategies could be to cluster the texts based on the chosen features and, if several clusters can be clearly distinguished in the data, to interpret them as text types. It could then be checked on which feature ranges and distributions these text types are based and where the boundaries between them are in terms of feature values. In a next step, it could be analyzed how the texts associated with the conventional genres are spread over the different clusters to see if the theoretical assumptions about the relevance of the three text characteristics for the distinction of historical from non-historical novels on a textual level hold. Still, at least two problems would persist in this case. The first problem is the one of underspecification of the assumptions. How can the theory of the historical novel be confirmed or rejected if the decisions about quantifying the definition’s different parts are part of the analysis results but not part of its starting point? One could always say that the theory may be valid or invalid for other quantifications. If not to confirm or reject a specific definition of a textual genre, the results of the quantitative analysis could still be used to refine the theory by providing the specification of its assumptions in numerical terms. The second problem that persists is that it would be necessary to define a limit for the ratio of conventional historical or non-historical novels that must be present in the found text types in order to say that the text types correlate with the conventional genres. It is implausible that 100 % of the novels with the conventional label “historical” would end up in the same cluster and none of them in the other(s). Is 90 % enough to say that the theory is adequate for the corpus at hand and describes the textual genre satisfactorily, or is just 60 % enough or only 51 %? The same problem would remain if classification was used instead of clustering. Again, instead of confirming or rejecting the hypotheses entirely, the results could be used to describe the extent of the intersection between the conventional genre and the text type.

There is another consequence of the terminological distinction between literary text types, conventional genres, and textual genres for a corpus-based digital genre analysis aiming to mediate between genre conventions and text characteristics. It is important to clarify and explain several aspects of the data and the analysis: the text selection, the kind of genre convention and / or genre theory that is addressed, the assignment of genres to the works in the text collection, and the modeling of factors other than genre that are assumed to have an influence on the distribution of textual features. These requirements are, in essence, not different for a non-digital, non-quantitative, and non-stylistic analysis of genres, but the ways in which they can be addressed are different. In quantitative text analysis, a genre is determined by contrasting and distinguishing texts associated with it from other texts – otherwise, the analysis would concentrate on the internal structures of the genre but not on its defining characteristics. Consequently, a corpus for the analysis of a specific genre always contains also texts that are not assumed to participate in that genre. This can facilitate the text selection because what has to be defined is the initial broader context of the genre. For example, an analysis aiming to find out about seventeenth-century tragedies can rely on a corpus of seventeenth-century dramatic texts, including comedies, tragicomedies, and other dramatic subgenres. Therefore, it is not all-decisive which texts exactly are assumed to be associated with the genre of interest at the outset of the analysis. A prior definition of the genre that is to be determined by the text analysis is not mandatory. Instead, the extension of the genre can be approached by collecting texts that are similar to the object of interest. In the case of the tragedies, these would be texts that were written and published in the same historical period, and that participate in the same major genre. What is important, though, is to remember that any delimitation of a textual genre that results from such a corpus-based contrastive analysis is relative to the texts that are part of the text collection but are not considered instances of the genre. If the tragedies are only compared to comedies, they are determined as non-comedies. This means that the problem of genre definition is relocated to a more general level, by selecting the wider context of texts in which instances of the textual genre of interest are assumed to be found. Inside the contextual text collection, the search for intersections between conventional genres and text types, and thereby the definition of textual genres, is possible, and its results are not predetermined.

Clarifying the target genre convention and making the procedure of assignment of genre labels to the texts in the collection transparent is crucial if the digital stylistic analysis aims to contribute to existing literary-historical research and if it seeks to find out about genres and not only text types. For very large corpora, assigning genre labels by hand may be unfeasible. One option is to use available bibliographic sources in digital format that are processable with scripts and to collect genre labels automatically. Even if labels are assigned individually and manually to the works, it is often necessary to rely on several sources. Single sources usually do not cover all the texts in a corpus, so different kinds of genre conventions may be involved. For example, the genre labels can originate from scholarly handbooks or library catalogs, thus involving two different kinds of systematic approaches to the genres. Even if several sources are used, it is probable that there remain texts of the broader context whose generic status cannot be clarified beforehand, for instance, if they have not been in the focus of scholarly, librarian, or public attention at all. To what extent historical labels can be used depends on the availability of historical editions and book covers. Because the use of large corpora makes it more difficult to pursue a uniform strategy for genre assignments, it is even more important to set out in detail how the task was solved and to discuss the consequences for the textual genre that is determined in the analysis.

As the text types in digital genre analyses are determined through text style, it is important to model factors which influence it, and this is also connected to the question of corpus building. Ultimately, the goal is not just to find arbitrary text types but to delimit textual genres in which textual similarities and communicative generic norms converge. How to find out if the facts found in the texts actually represent normative facts that are due to genre conventions and not traces of other intra- or extra-textual factors? Every genre analysis, a stylistic and also a non-stylistic one, has to make sure that the text corpus does not only consist of texts written by one or a few authors so that the claims built on the corpus analysis are not only about individual authors but about the genre. The issue is, however, more complicated in a stylistic analysis because all characteristics of the texts that are taken into account, modeled, or found lead to the text style or are derived from it. For instance, authorship can be primarily marked by a certain authorial style, and that style can be captured through the analysis of high-frequency words. Yet, the same words can also be properties that give evidence of some higher-order structural concept. For example, the type of narrator that is used in a narrative text can influence the number of specific personal pronouns. There can also be properties that are related to the textual genre, for instance, many general descriptive passages in novels of customs, which lead to a comparatively higher number of undetermined articles. To make sure that it is the genre convention that is captured with the overlap of genre labels and text types, it is, therefore, necessary to control and check possible interfering factors in the text style by modeling them as metadata and considering them in the process of text selection.

Several aspects remain that need to be clarified in connection with the systematic or historical status of text types and genres. In chapter 2.1.2.1 above, there was a focus on semiotic theories of genres and on generic terms. How can these models be understood in connection with text types, conventional genres, and textual genres? When they are used to analyze historical generic terms, they approach these with a language-based, communicative theory to find out which levels of discourse the terms address. This means that the norms that a conventional genre involves are described as characteristics of texts as speech acts. Following Raible’s model, for example, the subgenre novela documentaria would point to the level of the relationship between text and reality (a documenting activity purports that reality is depicted) and the level of the kind of linguistic representation (a documentary style is likely to be neutral and descriptive). These characteristics can be taken as hypotheses about the text types that relate to the conventional genres and thereby also as hypotheses about the features of textual genres, considering that for a stylistic analysis, all these hypotheses would have to be broken down to the level of linguistic representation. Furthermore, the semiotic models of genre constitute a possibility to differentiate between the various discursive levels on which sets of genres are defined, which can help to define more meaningful comparisons of genres. In the case of the novel, for instance, there is a wealth of subgenres, some of which emphasize similar aspects of speech acts and others entirely different ones. The more different the discursive levels are that the generic terms point to, the more probable it is that also the corresponding textual genres are defined on different levels, that the groups of texts associated with the genre conventions overlap, and also that the generic conventions might be different in kind. For example, it seems useful to compare sentimental novels to historical novels if both are interpreted as being defined primarily on a thematic level. On the other hand, it is less fruitful to directly contrast sentimental and romantic novels because the latter often (but not always) have a sentimental theme. In addition, the label “romantic” points to a certain aesthetic movement that was dominant in a specific period of time and not necessarily to thematic elements. This is not to say that sentimental or historical novels are only defined thematically, only that comparing these subgenres makes the most sense on a discursive level that is a prominent reference point for both.

A closer look at Schaeffer’s explanations is useful to further understand how the semiotic approach to genre relates to the three terms of text type, conventional genre, and textual genre proposed here. In the last part of his book “Qu’est-ce qu’un genre littéraire?”, he discusses generic regimes and logics, by which he means the kinds of relationships that individual texts have with “their genre”, and how these regimes are connected to the generic names (Schaeffer 1983, 156). Basically, he distinguishes between the relationships of exemplification and modulation. By exemplification, he means that texts instantiate genres globally (the whole text as a communicational act represents the genre, and not only parts of it) and without modifying the genre itself (the text is only an example of the genre, the genre stays the same). The discursive level to which generic names refer in such cases is the communicational act. In that case, the genre is not defined in textual (syntactic and semantic) terms but in pragmatic ones, concerned with the discursive attitude and the global intention. The type of convention that is involved in such cases is a constituting one. That is, one cannot depart from the convention to a certain degree but only fulfill it completely or fail to do so entirely. According to Schaeffer, cases of genres that can be exemplified are, for instance, drama or narration. If a literary work is conceived as a drama, that is the case independently of its exact form. Even if it contains narrative passages, it is still a drama, because the genre is not understood as being dependent on the actual textual realization but on the outer communicative level (Schaeffer 1983, 156–164). Schaeffer’s regime of exemplification is a bit difficult to grasp. It seems close to the concepts of mode and Schreibweisen that Todorov, Hempfer, and Genette develop because it refers to basic discursive attitudes that are linguistically motivated and tend to be historically invariant. However, Schaeffer uses nouns and not adjectives when he refers to this type of genre, and he relates the terms to literary works as a whole, which makes it more challenging to differentiate the genres by example from the historical genres. Schaeffer relates this generic regime to a “constituting” convention.

The second main type of generic logic that Schaeffer describes, the modulation, is, on the other hand, concerned with the syntactic and semantic entities that texts are. As soon as these discursive levels are addressed, Schaeffer claims, the relationship to the genre cannot be simply exemplifying anymore. Except for the special case of exact copies, each text is individual in this respect and always modifies the genre with which it is associated. In the case of modulation, the determination of the genre depends on the individual works and, thereby, also on the specific historical context in which the works are situated. The same generic names “drama” and “narration” can also be examples of the modulating logic if they refer to specific textual elements such as character constellations, themes, or elements of style. Schaeffer divides this regime further into a modulation by the application of rules, a hypertextual modulation based on historical genealogies, and a modulation by textual resemblance. The first kind of modulation is related to a regulative convention, the second to a traditional convention, and the third to no convention at all (Schaeffer 1983, 164–180). Schaeffer summarizes his system of generic logics in a table, which is reproduced here in table 1The colors were added by me. (Schaeffer 1983, 181). That way, it can be related graphically to the three terms of text type, conventional genre, and textual genre as they are used in this study. Generic logics according to Schaeffer. niveau référent relation définition description convention écart acte communicationnel propriété exemplification globale en compréhension contrastive constituante échec texte règle modulation par application prescriptive énumérative régulatrice violation classe généalogique modulation hypertextuelle heuristique spécifiante traditionelle transformation classe analogique modulation par ressemblance statistique typisante — variation

Three colors are used here to highlight the parts of the table that are related to text types (orange), conventional genres (dark green), and textual genres (light green) – no colors are used in Schaeffer’s original table. Text types correspond to the logic of modulation by resemblance, which leads to classes of similar texts that are defined statistically. Schaeffer has a double line between the last row of the table and the upper three ones, showing that he sees a clear difference between this generic logic and the other types involving different kinds of conventions. However, for the three generic logics of global exemplification, modulation by application, and hypertextual modulation, the conventional level is not clearly separated from the textual level. They mainly describe how that genre is constituted, for example, as a set of rules that are listed in a poetic formulation of a genre and that have a prescriptive character or as a heuristic description based on works that are traditionally associated with a genre. The two columns “relation” and “écart” (deviation), in contrast, describe to how a specific text instantiates a generic convention and how it may deviate from it. They can be interpreted in terms of the textual genre – as the point of contact between the text type that the text in question belongs to and the conventional genre that it participates in, even if Schaeffer focuses on divergence (modulation, deviation, violation, transformation) rather than intersection. Nevertheless, in Schaeffer’s approach, the levels of conventional genre and textual genre are intertwined. On the other hand, resemblance on the textual level is seen as not being related to the conventions at all, disconnecting the text types from the generic conventions. If one were to transfer this view to digital stylistic genre analysis, which is based on text features that are statistically evaluated, this would mean that it operates in another sphere, which has nothing to say about the literary-historical context of genres. Against this, the initial separation of the textual from the conventional level that is proposed in this study is not meant to isolate both concerns but to provide a more open basis for comparison so that the connections and disconnections of the two levels can be analyzed.

Another question that needs to be addressed is which kind of phenomena that are related to literary-historical conventions can be considered conventional genres. Is an aesthetic movement a genre? Literary genre theorists have different views on this topic. Todorov, for example, explicitly excludes literary movements from his concept of genre, but he does so because he presupposes textual coherence for genres, which according to him, cannot be expected with certainty for literary movements: Since a genre is the historically attested codification of discursive properties, it is easy to imagine the absence of either of the two components of the definition: historical reality and discursive reality. In the absence of historical reality we would be dealing with the categories of general poetics that are called – depending upon textual level – modes, registers, styles, or even forms, manners, and so on. The ‘noble style’ or ‘first person narration’ are indeed discursive realities; but they cannot be pinned down to a single moment in time: they are always possible. By the same token, in the absence of discursive reality, we would be dealing with notions that belong to literary history in the broad sense, such as trend, school, movement, or, in another sense of the word, ‘style’. (Todorov 2014, 200–201)He thus refers to textual coherence in a broad sense involving all the discursive levels of a speech act. The difference between factual or fictional utterances, for instance, would in fact be difficult to pin down to textual features in a narrower sense.

With discursive properties, Todorov means textual elements like phonetic or phonological features, thematic elements, and plot structure, but also aspects of the communicative level of texts, such as the factual or fictional status of the utterance (Todorov 2014, 199). Using the example of symbolism, he explains that it is known that this literary movement existed historically but that it is not proven that the works of authors identified with this movement (which would then be, for example, symbolistic poems) are characterized by common discursive properties. Instead, the movement may only have been based on friendships or manifests (Todorov 2014, 202). How can one be sure that there is no discursive reality or textual coherence behind works that are associated with a literary current? I am in favor of the position that this should not be ruled out in advance per definition but that it should be tested empirically, just as the textual coherence of conventional genres in the narrower sense should be examined. Digital genre stylistics provide good opportunities to do so. If there are explicit manifests and poetic writings that define literary schools and movements as frames of expectation for the creation and reception of literary works, this can be taken as proof of their relevance as historical, literary institutions. By locating literary movements on the level of conventional genres, nothing is said in advance about their relationship to text types and their significance as textual genres. However, for Todorov, like for Fricke, purely conventional genres are no genres, or at least they should not be called “genres”: “Genres are the meeting place between general poetics [concerned with the theoretical definition of text types] and event-based literary history [concerned with the historical study of the expression of literary conventions]” (Todorov 2014, 201). Both Fricke and Todorov only see textual genres as genres, which shows that their genre concepts are primarily theoretically anchored, despite the references that they make to the relevance of a historical foundation for the definition of textual genres.

In the semiotic models of Raible and Schaeffer, literary movements are not directly included either. Even so, Schaeffer discusses their status in the explanations accompanying his core model of discursive levels to which generic terms point. He explains that there are generic terms that cannot be ascribed to the five levels of the verbal act that he defines, for example, terms that refer to the context, place, and time of the speech act. Schaeffer explains that he does not directly include them in his model but that numerous generic terms exist that refer to these aspects, for example, the “baroque sonnet” or the “Greek epic” (Schaeffer 1983, 117–118). According to Schaeffer, such genre names reinforce how important the historical context is for the determination of genres. Together with the multiple dimensions of the verbal act that are addressed by generic terms, they show that genres cannot be reduced to texts as entire, whole objects (Schaeffer 1983, 119). In the context of this thesis, it was noted that also many generic terms that are associated with nineteenth-century Spanish-American novels refer to time and space. As an extension of Schaeffers model, the levels of temporal and spatial context are therefore included to organize the subgenre labels in the bibliography and corpus presented in chapter 3.See, in particular, chapter 3.2.3.6, where the empirically based discursive model of generic terms is described. Here, references to literary currents are described as labels that relate to the temporal context because literary periods are often named after such movements, and the currents are usually phenomena that are temporally limited.

Fowler utters another critical view on literary movements in relation to genres. In his book on kinds of literature, he mentions them in a section called “Other types”, which is part of the chapter devoted to the definition of modes and subgenres: “The system of generic categories is complicated by the existence of several other quasi-generic groupings. These include [...] the collective productions of ‘schools’ or movements (Metaphysical; Romantic; Georgian). They must be mentioned only to be dismissed” (Fowler 1982, 126–127). Fowler describes them as a “collective œuvre” and says that the features that literary works associated with schools or movements have may be “extensive and coherent enough to suggest genre” but that “they exhibit them independently of the historical kinds” (Fowler 1982, 128). With that, Fowler means that the common features that different works belonging to a particular movement share crisscross the boundaries of the genres that he calls the historical kinds. As an example, he mentions the Metaphysical poets who wrote in different poetic genres. Fowler states that these poems could be more similar to each other than works of the same poetic genre that belonged to various schools, for instance, different love elegies (Fowler 1982, 128). This is, however, only a problem if genre is not analyzed on different discursive levels and varying levels of generality and if it is assumed that there are no overlapping associations of literary works with different kinds of genres. If the love elegy is conceived as being primarily thematically defined (as involving the lament for a tragic love) and a metaphysic poem primarily formally – on the level of representation (as including innovative use of metaphors and allegories), then these two main characteristics do not exclude each other. The metaphysic poem and the love elegy can be understood as two different conventional genres overlapping with different text types. If Fowler says that metaphysical poems are more similar to each other than love elegies of different schools, this could even mean that the conventional genre “metaphysical poem” is textually more coherent than the conventional genre “love elegy”. Fowler’s argument is, therefore, no obstacle to also considering kinds that are conditioned by literary movements as genres.

A literary scholar who does not see a principle difference between literary genres and movements is Schlickers. In her book about the naturalistic Spanish-American novel, she argues that this distinction is not necessary. First, she confirms that the term “novela naturalista” indeed served as a generic name in the nineteenth century: La denominación del objeto de este estudio, sin embargo, no causa mayores problemas, porque la noción de ‘novela naturalista’ (y sus variantes) funciona(ba) como nombre genérico, indicando, pues, cierta clase de textos literarios y revelando así una conciencia histórica del género, que influía tanto en la producción como en la recepción de las novelas. (Schlickers 2003, 16)This can be confirmed by the explicit genre labels found in the digital bibliography of nineteenth-century Argentine, Cuban, and Mexican novels created for the present study because the label “novela naturalista” is found in the subtitles of three historical editions (“¿Inocentes o culpables? Novela naturalista” (1884, AR) by Juan Antonio Argerich, “Los bandidos de Río Frío. Novela naturalista, humorística, de costumbres, de crímenes y de horrores” (1892, MX) by Manuel Payno, and “Conventillo de intelectuales. Novela de índole rebelde y de género naturalista que no deben leer las almas timoratas” (1904, AR) by Francisco Guillo). It is also referred to in the prefaces of two of the naturalistic novels whose full text was examined (“El tipo más original” (1879, AR) by Eduardo Ladislao Holmberg and “Perfiles y medallones” (1886, AR) by Silverio Domínguez).

Then she discusses the status of the naturalistic novel as a historical and systematic category. According to Schlickers, there is no clear generic model for the naturalistic novel when compared, for example, to the picaresque novel, for which several constant and variable textual characteristics are known: “La novela naturalista, por el contrario, parece constituir más bien una corriente o un movimiento literario” (Schlickers 2003, 16). However, Schlickers says that all of the terms “current”, “movement”, “genre”, and “poetic school” are debatable and have similar extensions, although different intensions. She is in favor of modeling the naturalistic novel as a subgenre because the works that are associated with this convention repeatedly reference the same prototypical works of the French naturalistic tradition. This shows that there was an awareness of the genre and that the works in question constitute a series of naturalistic novels. Schlickers states that conceptualizing the naturalistic novel as a subgenre allows one to analyze its distinctive features, which guarantee the coherence of the texts associated with Naturalism. At the same time, analyzing the naturalistic novel as a subgenre can serve as a constructive and heuristic means by which imprecise and contradictory textual features and purely poetological particularities can be considered (Schlickers 2003, 17). Here she assumes that there is a textual genre “novela naturalista”, but that it is not entirely congruent with the conventional genre. The case of the Spanish-American naturalistic novels is only one specific historical example of a convention that is usually described as part of a literary school or movement, but that could also be conceived as a genre. In the present study, the example of the Spanish-American naturalistic novel is taken as an argument in favor of the possibility of understanding literary currents as a type of subgenre. Digital genre stylistics can contribute to expanding the limited knowledge about common textual characteristics of literary currents, and in particular of the naturalistic novels to which Schlickers points.

Discussing whether literary movements can be understood as genres or not leads over to the issue of theoretically separating genres from other discursive entities and of the place of genres in a system or hierarchy of forms – also beyond the question of the difference between text types and genres. There has been much research in literary genre theory on this topic, and a range of different terminological systems have been proposed. They cannot all be reviewed here, so only selected systems are presented shortly. It will be clarified how the different kinds of theoretical, generic terms that have been proposed relate to the three terms of text type, conventional genre, and textual genre defined above. Furthermore, as the empirical part of this study is concerned with subgenres of the novel, a point of interest is how genres and subgenres relate to each other. A terminological system that is prominent in the German-speaking context is the one proposed by Hempfer.Hempfer (1973, 14–29) dedicates a whole chapter to the discussion of terminological problems. Although his study was first published in the seventies, it is still influential today in the German-speaking area. Klausnitzer and Naschert discuss his approach as one of the positions in genre theory which sparked some debate in the twentieth century (Klausnitzer and Naschert 2007, 387–404). Neumann and Nünning (2007) also refer to him repeatedly in their overview of problems, tasks, and perspectives of genre theory and history. On the one hand, he proposes to use the word “genre” (in German, “Gattung”) as a meta-theoretical term which may include all other terms used to designate kinds of texts (Hempfer 1973, 16–18), as for example “meta-genre”, “subgenre”, “form”, “kind”, “mode”, “species”, “variety”, “text type”, or “text class”. This general, meta- or pre-theoretical use of the term “Gattung” is what also Fricke (1981, 133) suggests. When no further theoretical distinctions are made, the term “genre” is used in this general sense here, as well. On a second level below the most general meta-term, Hempfer defines several theoretical terms for specific kinds of text groupings. The main components of his terminological system are “Schreibweise” (“diction”), “Typ” (“type”), “Gattung” (“genre”), and “Untergattung” (“subgenre”). “Dictions” are defined as ahistorical constants (e.g., “the narrative”, “the dramatic”, “the satiric”); “types” as trans-temporal forms of dictions, that is, as the theoretically possible set of types derived from them; “genres” as historical and concrete realizations of the general dictions (e.g., “novel”, “epopee”, “verse satire”); and “subgenres” as subtypes of single genres (e.g., “picaresque novel”, “pathetic verse satire”). According to Hempfer, types and genres can both be derived from dictions via transformations, which means that his system builds on a dynamical and structural concept and not a hierarchical one. Genres and subgenres do not need to be based only on one diction but can be derived from several ones (e.g., a “comic epic”). In addition, Hempfer uses the term “Sammelbegriff” (“collective term”) for terms that designate classes in a logical sense. Individual texts can be assigned to such classes on the basis of any characteristic (e.g., “poetry” as texts in verse form) (Hempfer 1973, 27–28). Comparing Hempfer’s system to the three terms of text type, conventional genre, and textual genre used here, the following observations can be made. The term that is closest to the text type is Hempfer’s Sammelbegriff because it designates a purely logical grouping without necessary relationships to a certain theory of genre or a genre convention. However, the Sammelbegriff is different in that it presupposes logical classes, which is not done for the text type here. The categorical status of text types and textual genres has to be clarified further, which is done in chapter 2.1.4 below. Not only common but also similar features of texts can be constitutive for text types. Hempfer’s Gattung corresponds roughly to the textual genre in that it designates groups of texts that share textual features for which there is historical evidence and which were relevant as communicative norms. A slight difference is that the textual features of Hempfer’s Gattungen are derived from the features of underlying ahistorical constants, whereas the aspect of the origin and motivation of the textual characteristics is not covered by the terminological distinction proposed here. Hempfer’s Untergattungen can be considered subtypes of textual genres. As far as I can see, there is no equivalence to the purely conventional genre in Hempfer’s system of terms. Hempfer’s Schreibweisen and Typen as purely theoretical terms are not covered here, which shows that the three proposed terms of text type, conventional genre, and textual genre focus on the mediation of text-immanent generic features and text-external communicational as well as purely conventional aspects of genres on an empirical level. It is a further task to build new theories based on textual genres that are found in this process of mediation or to clarify their relationship to existing literary theories of genre.

Another system of terms is used by Fowler, who distinguishes between “kinds”, which are the historical genres (e.g., sonnet, parable), “modes”, which are selections or abstractions from kinds (e.g., comic, aphoristic), “subgenres”, which are subtypes of kinds (e.g., sea eclogue, historical novel), and “constructional types”, which are purely formal patterns (e.g., ring composition, sequence). To distinguish between the different sorts of generic categories, Fowler uses the idea of a “generic repertoire”, which means all possible levels on which features that are characteristic of a genre can be chosen. A basic distinction is made between “formal” and “substantive” features. Formal features include structural characteristics and, for example, verse metres. Substantive features are, for instance, related to themes, purpose, and intended audience. The kinds usually combine both types of features, whereas the constructional types are only based on formal features. Modes are more or less unstructured. They have no or only a few formal features but invoke kinds through samples of their substantive features. The subgenres are defined as having the same formal features as their corresponding kind, plus additional characteristics related to the content of the texts (Fowler 1982, 55–56). Evaluating the relationship of Fowler’s terms to text types, conventional genres, and textual genres, the following observations can be made: kinds can be understood as corresponding more or less to the textual genres and subgenres to subtypes of textual genres. What Fowler emphasizes repeatedly is the historical variability of the kinds’ features.For instance: “Kinds may in this way give the impression of being fixed, definite things, located in history, whose description is a fairly routine matter. As we shall see, there is something in the idea of definiteness. But describing even a familiar kind is no simple matter. We may think we know what a sonnet is, until we look into the Elizabethan sonnet and are faced with quatorzain stanzas, fourteen-line epigrams, sixteen-line sonnets, and ‘sonnet sequences’ mixing sonnets with complaints or Anacreontic odes. Besides such historical changes within individual kinds, there are wider changes in the literary model allowed for, with their repercussions on the significance and even categorization of generic features” (Fowler 1982, 57). This raises the question of whether one kind should be conceived as one textual genre allowing for internal variation or eventually several textual genres based on more compact text types. This question will be addressed in the next chapter 2.1.4 on categorization. However, Fowler also states that despite all variation and historical change, kinds are not indeterminate. Preliminarily, he defines them as follows: “a kind is a type of literary work of a definite size, marked by a complex of substantive and formal features that always include a distinctive (though not usually unique) external structure” (Fowler 1982, 74). As Hempfer’s Schreibweisen, also Fowler’s modes are not directly covered by the concepts of text type, conventional genre, and textual genre. The latter refer to literary texts as structural units, whereas the former are abstracted and disconnected from external structures: “Modes have always an incomplete repertoire, a selection only of the corresponding kind’s features, and one from which overall external structure is absent” (Fowler 1982, 107). A difference between Hempfer’s Schreibweisen and Fowler’s modes is that Fowler does not see the modes as ahistorical constants. He describes them as distillations of kinds, i.e. of the features of kinds that seem permanently valuable. In that respect, they are more durable than the historical kinds because they are not linked to external forms that become outdated faster. However, these distillations may also change or become obsolete, as, for example, the heroic mode, which is only conserved in historical or political novels. (Fowler 1982, 111). When modes are combined with kinds and when they characterize the kinds in more detail, the combination of mode and kind is regarded as a subgenre here (for example, a “comic novel”).

In this study, no principle difference is made between genres and subgenres. Like genres, also subgenres can be conventional genres, and they can overlap with text types to form textual genres. One difference between both is that the genres usually have a more precise formal delimitation than different subtypes of the same genre. The latter tend to be based on differences in subject matter or style so that different textual characteristics and features become relevant in each case. In addition, subgenres can be very inconsistent because they are formally less fixed than genres. As Fowler remarks, “To determine the features of a subgenre is to trace a diachronic process of imitation, variation, innovation—in fact, to verge on source study. At the level of subgenre, innovation is life” (Fowler 1982, 114). It can therefore be more challenging to find the correspondences between conventional subgenres and text types than between conventional genres and text types, and hence be more challenging to determine the textual subgenres than textual genres. However, the degree of innovation that subgenres undergo depends on the type of subgenre that is investigated and on factors such as canonicity and perceived literariness of the works. Especially in popular literature, it is to be expected that there are works that follow quite schematic patterns of subgenres.

Even if the levels of genre and subgenre are not strictly separated here, it makes sense to consider them in the construction of the corpus to be analyzed to compare genres on a similar level of specificity. If subgenres are the point of interest, it makes sense to build a corpus of instances of the corresponding genre and to include works in it that are associated with different types of subgenres. Then the selection criteria for the corpus as a whole can be based on the formal characteristics that define the genre. That way, determining textual subgenres does not need to be restricted by defining text types beforehand. Different subgenres of the contextualizing major genre can then be contrasted with each other. The relationship of subordination between genres and subgenres is a means of combining a deductive with an inductive procedure in constructing a corpus. Other strategies are conceivable, for example, to combine historical labels as signals for conventional genres with preliminary (sub)genre definitions as proxies to textual genres to build a corpus from which to start the analysis.

That genre in the narrower sense (Gattung, kind, conventional, and textual genre) is bound to certain features of formal structure has already been pointed out. Which textual entity is the one that participates in genre? Raible sees an affinity between genres and specific degrees of complexity. He observes that, in general, text is a relative and dynamic notion. A novel such as “Eugenie Grandet” by Balzac is a text, but the series of novels that it belongs to, the “Comédie humaine”, or only a subpart of the novel such as a single chapter, are also texts. Usually, genre is bound to the level of the single, whole novel (Raible 1980, 327). Instances of this kind of textual entity are the ones that are collected in a corpus, that are associated with one or several genres, and that are analyzed as to their textual coherence and compliance with generic conventions. However, a novel like “Eugenie Grandet” is a literary work, and as such, it can be realized as text in different forms and contexts. Is genre linked to the literary work as a whole or a specific realization of it in text form? How stable is the association between the literary work or text and the genre? In principle, it is assumed that the generic identity of a text is identical to the generic identity of the work that the text represents. This means that an English translation of “Eugenie Grandet” would generally be considered as participating in the same genre as the original French text. Different editions of a work in the same language but published in different years or even centuries would also commonly be associated with the same genre. If there is a new work, also the genre may be different. “Eugenie Grandet” as a drama or movie would be a new work and have a different genre. If one looks more closely, this question is, however, not so easy to resolve. First, one can debate whether a new version of a text is another realization of the same work or another work.In the area of bibliographic modeling, this question is, for example, addressed in the conceptual model FRBR of the International Federation of Library Associations and Institutions (IFLA): “variant texts incorporating revisions or updates to an earlier text are viewed simply as expressions of the same work [...]. Similarly, abridgements or enlargements of an existing text [...] are considered to be different expressions of the same work” (International Federation of Library Associations and Institutions (IFLA) 2009, 17). Translations are also considered as different forms of the same work. On the other side, “when the modification of the work involves a significant degree of independent intellectual or artistic effort, the result is viewed [...] as a new work” (International Federation of Library Associations and Institutions (IFLA) 2009, 17). This includes, for example, paraphrases, summaries, adaptations for children, parodies, and changes from one art form to another. As can be seen, in individual cases, it may be difficult to decide whether changes to a text are a minor revision or a significant modification. Certainly, the question of the definition and unity of a literary work is also central in literary studies. For example, the role of authorship or the prerequisite of completion for a work to be a work can be questioned. For an overview, see Thomé (2007). Second, as genres are conventions that can be described as literary institutions and as horizons of expectation for authors, publishers, readers, critics, etc., which are anchored in specific historical settings, the generic identity of a work can be influenced by the context in which it is realized. This is illustrated clearly by Schaeffer, who discusses the example of the story “Pierre Menard, autor del Quijote”, which was published in 1939 by Jorge Luis Borges. The story centers on the idea that the fictional author Pierre Menard publishes parts of the work “Don Quijote” as his own creation in the twentieth century, although formally, they correspond exactly to the text authored by Cervantes and published in the early seventeenth century. As Schaeffer illustrates, the syntactically identical text would have different generic identities because in the twentieth century, it would be considered a historical novel with an archaic style, whereas in the seventeenth century, it was primarily received as a parody on romances of chivalry (Schaeffer 1983, 131–134). In this case, also the author is different, and one could therefore speak of two different works when comparing the original “Don Quijote” to the imagined twentieth-century recreation of it. Nevertheless, the example makes clear that the generic identity of a literary work may depend on its realization as a document in a specific context.

In the digital bibliography of nineteenth-century Spanish-American novels created for this study, there are examples of works that have been associated with different subgenres of the novel in different editions, marked by different subtitles. For instance, the novel “Tomochic” by the Mexican writer Heriberto Frías was first published in 1894 with the subtitles “Episodios de la Campaña de Chihuahua. 1892. Relación escrita por un testigo presencial”. In 1899, it was republished without any subtitle, and in 1906, the subtitle was changed to “Novela histórica mexicana”. This shows how the novel was initially presented as a testimony and a contemporary documentary novel, and only twelve years later, it was considered a historical novel. Modern critics have interpreted it as a historical, political, social, realist, and naturalistic novel.See the mentions of the novel in Dill (1999), Fernández-Arias Campoamor (1952), Gálvez (1990), Read (1939), Sánchez (1953), and Varela Jácome ([1982] 2000). Such generic variability of individual literary works can be clarified when considering it in relation to concepts of text type, conventional genre, and textual genre. Concerning its text-immanent characteristics, and more specifically, its stylistic features, the work does not change over time because these do not depend on the communicative and historical context in which it is embedded through its various realizations. For this to be true, the simplifying assumption is made that the different published editions do not involve considerable textual adaptations of the work. This means that the different text types that the work can belong to do not change. They depend on the textual level and the kind of textual features that are selected for the analysis (for example, most frequent words or topics). What may have changed in a different historical context is the conventional genre, that is, the concept of the genre that was effective at the time, as well as the specific communicative context in which the work is to be seen, which involves the expectations of publishers and readers. It may then be the case that the work did not fit the conventional criteria for a historical novel when it was first published but that it did with the 1906 edition. The concept of the historical novel may have changed by that time. Furthermore, the novel only falls into the definitory pattern of the historical novel once there is a greater distance between its creation and publication year, which means that the perspective on it has changed. If the generic convention that the literary work is associated with is a different one, also the textual genre becomes another one because it is derived from a different intersection of genre label and text type.

Again, this makes clear that digital stylistics must take conventional genres into account if it aims to produce text analysis results that are historically adequate. In addition, when a corpus is created for digital stylistic analysis, the question of generic identity of the individual texts has to be tackled. For the corpus of nineteenth-century Spanish-American novels created here, it was decided to attribute all subgenre assignments directly to the work level. As laid out in more detail in the chapter on creating the bibliography and corpus, it was encoded if the assignments are contemporary or literary-historical. The boundary between both was drawn based on two criteria: first, the temporal limits of the corpus, and second, the origin of the genre label as endogenous and textual or exogenous and meta-textual. As the corpus covers works that were first published between 1830 and 1910, subgenre assignments that were made after 1910 are considered literary-historical, and the ones before that year as contemporary. In addition, the labels that are derived from paratextual signals are differentiated from the ones that external actors conferred. So in this concrete case, for the purpose of subgenre assignment, a simplification was made by defining the period that is covered by the corpus as broadly homogeneous regarding the historical genre conventions. This was done because the change over time of individual subgenres is not the primary concern of the analysis conducted here. If it was, it would have been more important to consider the changing conventional generic identities.

At the beginning of this chapter, several core issues regarding the relationship between the theory and the history of genres were raised. The last one, referring to the origin and evolution of genres, has yet to be discussed. Theories for genre change are an extensive topic of their own, and literary genre theorists have made several different propositions in this regard. For example, Todorov presents a theory about the origin of genres: “Where do genres come from? Quite simply from other genres. A new genre is always the transformation of an earlier one, or of several: by inversion, by displacement, by combination” (Todorov 2014, 197). He thus sees the generic system as one that is in constant transformation, and he approaches the question of the formation of genres not through historical analysis but through systematic considerations. Different types of genres are compared to other kinds of speech acts to which they are related. For example, the prayer as a genre is related to praying as a speech act, the novel to telling, and the sonnet to “sonneting”, which does not exist as an institutionalized speech act. Todorov concludes that different kinds of developments are involved, from general simple speech acts to more complex literary genres, but that in general, genres derive from “normal” language: “that makes it possible to see that there is not an abyss between literature and what is not literature, that the literary genres originate, quite simply, in human discourse” (Todorov 2014, 208). In the more historically oriented genre theories, generic change does not need an independent explanation because it is a central part of the genre concepts. For example, of the three generic logics Schaeffer proposes, three are based on the principle of modulation, which entails that each text that participates in a genre modifies the genre characteristics (Schaeffer 1983, 166). The dynamic of genre change is then a question of the extent to which one or several texts alter the genre concept. In Jauß’s theory, the history of genres is explained with the variability of the readers’ (and authors’) horizons of expectations as a “temporal process of the continual founding and altering” (Jauß 2014, 132). Voßkamp’s idea of genres as literary-social institutions also involves processes of permanent reductions which lead to stabilizations and destabilizations of the institutions (Voßkamp 1977, 30). A literary scholar who devoted several book chapters to the transformation of genres is Fowler (Fowler 1982, 149–212). He provides an overview of the explanations different genre theorists have developed for the origins and changes of genres and generic systems. Similar to Schaeffer, Jauß, and Voßkamp, Fowler assumes that processes of change are at work constantly, but in his view, these processes are inherently literary and not of a general linguistic or historical nature. Taking into account many literary-historical examples, Fowler describes the main types of transformation processes. Among these, there are topical invention, the combination of generic repertoires, the inclusion of generic repertoires into others, a selection of new repertoire elements from other genres, and their mixture (Fowler 1982, 170).Fowler not only sets forth transformations of genre but also differentiates them from modal transformations, which he calls “generic modulation”, with a different sense than the modulation that Schaeffer describes.

How are processes of generic change to be understood in the context of text types, conventional genres, and textual genres? In general, also change can be described on these three levels: there can be textual change leading to transformed and new text types, change in conventional genres if there is, for example, a poetic discussion resulting in new elements that are considered necessary for a specific genre, or if texts are associated with new conventions through the use of genre signals or labels. The transformation of the textual genres is then a result of the shifts that are at work on the other two levels: if either the text type or the conventional genre or both at the same time change, then also the textual genre becomes different. A more difficult question is where the boundary between one text type and another, one conventional genre and a different one, and hence also between one textual genre and a new one lies. If genre transformations are assumed to be constantly at work, how can the unity of a text type, a conventional genre, or a textual genre be defined at all? This question is crucial not only for the historical change of genres but also for their description from a synchronic perspective. The boundaries between different text types, various conventional genres, and textual genres are not necessarily clear-cut. The overlap of text types and conventional genres can also take several forms. These problems are discussed in chapter 2.1.4, on categorization.

With access to very large corpora of digital texts that, in theory, can cover whole periods or several periods of literature, digital genre stylistics is enabled to address genre analysis over time. Different approaches to this problem are possible and have been pursued. Although generic change cannot be comprehensively considered in this study, some general remarks on the importance of corpus design in this regard are made. Obviously, a central aspect to capture change over time is the collection of metadata about the creation or publication dates of the literary works that are part of the corpus that is analyzed. Furthermore, decisions have to be made about the status of different editions of the same works. Questions are, for example, if only first historical editions are considered or if modern editions are used, and also if only one edition per work is analyzed or if different versions of works are compared. In most cases in which transformations of genres over time have been analyzed in digital genre stylistics so far, the reference to temporal metadata is the primary strategy to capture change. For example, a corpus can be subdivided into an earlier and a later period, and the texts contained in both partitions can be compared to find the differences between early and late variants of the genres that are covered by the text collection. Such differentiation can also be made in a more granular way, analyzing, for example, differences by decade or by year. Jockers, for example, analyzes the stylistic change of novels in English by decade (Jockers 2013, 82–89). Using publication years as the temporal unit, Underwood proposes to adopt the Foote novelty, an algorithm that was originally developed to locate points of significant change in music. Underwood uses the algorithm to detect “revolutions” in the history of the novel (Underwood 2015a; Foote 2000).The CLiGS group also experimented with this algorithm to detect phases of accelerated literary development in over 300 French twentieth-century novels in a project presented at the conference “Forum Junge Romanistik”, using topics and temporal expressions as features (Schöch et al. 2017). In the mentioned cases, the feature set used to analyze the novels over time is the same for all points in time or periods. What is measured is how the feature distributions change in relation to the temporal metadata. One could propose different historical stages of text types if the values and constellations of the textual features vary considerably between one and another period. Another way to model textual change is to create different feature sets for different points in time or features that develop over time. In that case, a challenge is to clarify how the different feature sets or features are related to each other so that it is possible to speak about a change. For example, there are dynamic topic models in which the composition of the topics evolves temporally (Blei and Lafferty 2006). As in the previous cases, the primary concern is the change on the level of the text types. This can be linked to developments of conventional genres if, for instance, metadata about different genre labels that the texts had at different points in time is integrated into the analyses, as Underwood did in his account of Gothic novels (Underwood 2016). However, even if the genre labels are constant, changes in the text types produce different intersections with conventional genres and hence altered textual genres.

Several issues that are related to the question of the systematic or historical nature of literary genres were discussed in this subchapter. Different literary genre theories emphasize the dependence of genres from the historical context to different degrees. Several terminological systems have been developed to distinguish between historically constant and more variable generic notions, between linguistically and text-founded types as well as conventional groupings of texts. Digital genre stylistics has been described here as a field of research in which there is a strong focus on the analysis of text types because groups of literary works are formed and analyzed based on stylistic features that are derived from the linguistic surface of the texts. Then again, in its applied form, digital literary stylistics is mainly concerned with empirical analyses of historical text corpora, so that questions of the historicity of the texts inevitably have to be addressed. Such questions are, for example, which literary period to cover with the corpus, how the literary works in the corpus are dated, which types of editions to use, and from which perspective and based on which sources the generic identity of the works is determined. Usually, it is especially the latter aspect of deciding on the type of genre labels that links digital genre stylistic analyses to modern or historical genre conventions. Literary works in very large digital corpora can normally not be labeled according to purely theoretical criteria that have been developed based on close reading of texts and that require close reading for their application. Instead, the genre labeling of works in large text collections usually needs to be based on categorizations that others have made for part of the texts (authors, editors, readers, critics, librarians, scholars, etc.), so that sets of genre labels are the result of collective decisions influenced by the respective generic conventions and theories that apply.

Besides selecting genre labels, selecting works for the corpus also means that a connection to contemporary or historical expectations towards genres is established. If, for example, a corpus of novels is assembled to analyze subtypes of novels, the genre discourse is already involved in the outer delimitation of the corpus, even if the subtypes of the novel are defined by relying exclusively on textual surface features. That is, even a text-centered analysis of literary types addresses questions of genre as soon as this notion is part of the definition of the corpus. Therefore, it is argued that both literary text types as systematic groupings of texts based on textual features and conventional literary genres as historically bound literary institutions are important for digital genre stylistic analyses. However, the potential of digital genre stylistics to enhance the knowledge about the literary history of genres can best be developed if literary text types and conventional literary genres are terminologically separated and defined independently of each other in the first place. That way, insights into textual genres are not limited from the outset by requirements of completeness, integrity, or congruence on either of the other two levels. The initial separation of levels on which genres (in the general sense) are analyzed is the more important as stylistic textual features can be very different from the literary features that have been formulated as defining genres either in historical conventions or literary theories, or both. Defining textual literary genres as convergences or intersections of text types and conventional genres allows for finding new correspondences between both perspectives on the genres.

To consider several levels of analysis for genres and to relate these to each other has also been advocated for in more recent genre theoretical approaches. Gymnich and Neumann (2007), for example, propose a “compact definition” (“Kompaktbegriff”) of genre. They consider four main levels of genre analysis (the textual level, the cultural-historical dimension, the individual-cognitive dimension, and the functional dimension) and outline the points of contact between them.Gymnich and Neumann synthesize the references between the four levels in a diagram: they interpret the individual-cognitive level as mediating between the textual level and the cultural-historical dimension of genres and describe the functional aspects as superordinate to the other three levels. Their model aims not to provide a homogenizing general theory of genre but an integrative view on the different theoretical approaches to it so that scholarly communication about genres is facilitated (Gymnich and Neumann 2007, 34–35). The levels of literary text types and conventional literary genres proposed here cover, in particular, the textual and the cultural-historical dimensions. Individual-cognitive and functional aspects can also be connected to both of them, even if they are not the primary focus of digital genre stylistics. Compared to the integration of the different analysis levels in the “compact definition” of genre, the convergence of text types and conventional genres in textual genres described here is more concrete because it refers to intersections of text groups. The terminological differentiation between literary text types, conventional, and textual genres relates directly to the possibilities and characteristics of digital quantitative approaches. They need their own genre theoretical foundation, not because they should be isolated from other directions of genre research, but to have adequate terms that allow for communicating and clarifying what genre stylistics can contribute to genre theory and history in general. In the following section, the question of how text types, conventional, and textual genres can be held together and delimited as categories is addressed.

The idea that genre categories can be conceptualized as logical classes prevailed in genre theory for a long time. From the early poetics onwards, attempts were made to systematize the field of literary texts by grouping them on the basis of formal, functional, and content-related similarities, with the goal to provide “as exact a classification of concrete texts into clearly disjunct classes as possible” (Hempfer 2014, 405). Besides one-level sets, also hierarchical systems or taxonomies were proposed, requiring distinct and complementary classes. In some cases, they were conceived in analogy to scientific systemizations, particularly biological ones. Bonheim, for example, developed the “cladistic method of classifying genres”, alluding to phylogenetic systematics (Bonheim 1992).Bonheim’s method is based on the idea of separating different classes of literary genres by finding necessary (“megafeatures”) and optional elements (“microfeatures”) for each class. An aspect of his model that is of interest from the point of view of digital genre stylistics is that also “loss features” are considered, i.e., features that are negated or missing in certain genres (Bonheim 1992, 2–3). Strube summarizes the activity of the classifying literary scholar as follows: texts are gathered together in a group based on their similarity as things; then the term that serves to designate these texts is defined by formulating the conditions for its use; the term is differentiated from other generic terms relying on consistent distinguishing criteria and so that neighboring terms exclude each other intensionally or by content; borderline cases are left aside; all terms of the classifying system are organized by juxtaposition or hierarchically; and the terms are defined form a certain literary theoretical perspective (Strube 1993, 42–43). Classifying literary texts by genre is a basic activity for literary scholars and is useful for several reasons. For instance, it helps to build corpora of literary texts that cover a delimited area. In addition, classificatory terms can be used to find out to which genre a work belongs at all. Furthermore, knowledge about necessary features of genres is helpful when texts associated with the genres are interpreted. Classification also provides a way to organize the “embroiled world of literature” (Strube 1993, 41)“der gleichsam verwickelten Welt der Literatur”. in a clear way.

Nevertheless, when classification – in the narrow sense of defining classes that require the presence of a fixed set of features for its members and that fit into a system of non-overlapping categories – is applied to literary genres, there are also limits to its usefulness. Bonheim describes them when pondering the limits of his cladistic method of classifying genres: There are five distinct grounds why cladograms of genres are fraught with such uncertainties. One has been mentioned: our definitions are not sufficiently encyclopedic. A second one is that genres are alive and change from age to age. A third is that genres in France are different from genres in England or Germany; for it is a common observation that genre terms are frequently untranslatable, and we feel forced to speak of a conte or a Bildungsroman in English and of the villanelle and the short story in German. So genres and genre terms are marked on grounds of national origin as well as history. The fourth reason is that genre terms are rarely terms in the scientific sense, but semi-terms, made grubby by repeated use and misuse. The fifth reason is that every new critical school, like every new drive into the bottomless pit of linguistic theories, calls our attention to some qualities of texts which, as far as we are aware, had not before been clearly brought into focus. That means that every model of genre is an open set of features. No cladogram can be final and valid for all time and users. (Bonheim 1992, 15)

So among the main problems that Bonheim sees are the dependence on a cultural and linguistic context and the historical variability of genres, which a classificatory approach cannot directly cover. Further problems derive from models’ theory dependence and selectivity, but these are independent of the categorization procedure. Bonheim concludes that definitions of text kinds are, above all, heuristic tools (Bonheim 1992, 16). That classificatory definitions of genres are not able to cover synchronic variation, including the one caused by the individuality of norm-breaking literary works or historical change, has often been adduced. Alternative concepts of categorization have been proposed.For example by Hempfer (2014, 416–419), who proposes to use the concept of family resemblance instead to conceptualize historical genres, such as the elegy, or Tophinke (1997, 161–163), who suggests a solution based on prototype theory for official municipal charters and the unofficial ones used by merchants in the Late Middle Ages. Strube, on the other hand, defends classificatory procedures by arguing that literary scholars do not classify in the same way as linguists or other empirically working scientists: Die Voraussetzung, die von den Textsortenlinguisten gemacht wird (nämlich daß es auch in der ‘Welt der Literatur’ scharf abgegrenzte Klassen oder Begriffsextensionen gebe), mag berechtigt sein, solange sie sich auf sogenannte Gebrauchstexte [...] bezieht. Sie auf kompliziertere Arten der Literatur zu beziehen, heißt, in die Irre zu gehen. Wenn überhaupt Begriffe zur Einteilung der (bisherigen) Literatur taugen, dann nicht Textsorten-, sondern Art- und Gattungsbegriffe des vom traditionellen Literaturwissenschaftler gebrauchten Typs. (Strube 1993, 58)

His remarks can be interpreted as a rejection of classification in the logical sense as the only valid procedure and as an advocation for a broader definition of “classification”, that includes specific soft classificatory terms, which he calls “univocal”, “paronymic”, “porous”, and “family resemblance terms”, of which the univocal terms are closest to classificatory terms in the strict sense (Strube 1993, 13–28).

As digital genre stylistics is a field in which an empirical methodology prevails, including the use of statistical procedures, also logical classification is part of its repertoire and is repeatedly employed for categorizing literary texts by genre. In many literary stylistic papers, classificatory approaches are used with a focus on which features are most suitable to capture differences between genres, including aspects of style and content, but also structural characteristics of the texts such as text order or representation of speech.Examples of classificatory genre stylistic studies are Calvo Tello (2018), Gianitsos et al. 2019, Henny-Krahmer (2018), Hettinger et al. (2016), Schöch (2017b), Schöch, Henny et al. (2016), and Underwood (2015b). A special focus on the separation of genre from author signals can be found in Calvo Tello et al. (2017) and Schöch (2013). Digital studies on the classification of genres, in the strict sense of logical classes, are relevant for several reasons: they help to find, model, and interpret textual cues that are crucial to recognizing genres, they contribute to assessing established methods of text mining, machine learning, and NLP regarding their value for genre categorization, when tested empirically on corpora of different languages, periods, cultural contexts, and genres, they expand knowledge about the extent to which the methods are sensitive to the kind of data, and they also help to solve practical problems as, for example, to index large collections of texts by genre. That classification is so frequently used in digital genre stylistics is probably due to the methodological influence of computational linguistics and computer science.Classification as a supervised method of machine learning is introduced in more detail in the analysis part of this study. See chapter 4.2.2.1.

As a supervised method, classification has the advantage that the analyzed text classes are directly linked to the genre labels so that the correspondence (or non-correspondence) of generic terms to specific textual features can be checked straightforwardly. Furthermore, the success or failure of the classification can be easily measured by comparing the actual genre labels, that is, the labels that the texts were associated with at the outset of the analysis, to the ones predicted by a classifier. If the genre labels represent the conventional genres and the texts themselves are the data based on which text types can be defined, then supervised classification is an ideal method to test to what extent both levels overlap. The more successful a classification based on selected genre labels and textual features is, the more probable it is that the conventional genres represented by the labels actually correspond to text types defined on the grounds of the selected features. As a consequence, it is also probable that they constitute textual genres. As at least one positive and one negative class are needed for supervised classification, as a minimum, two conventional genres and two text types are involved in such an analysis. The text types are then differentiated not by the kind of features – because the same feature set is used in the whole setting – but by the characteristics of the feature values and how the various feature values are distributed in one text type versus the other. Because the classes are clearly separated from each other, it is possible to inspect the feature distributions in each class to learn about its textual characteristics, and it can also be analyzed which features are decisive to separate one class from the other. This means that text types and textual genres are defined contrastively in a classificatory setting. This highlights the importance of deciding which conventional genres and which literary works associated with them should be analyzed together to find the corresponding text types and textual genres.

However, it can also turn out that classification does not yield very accurate results, depending on the kind of conventional genre that is analyzed, the individual literary works that participate in it and that are selected for the text corpus, and the textual level that is examined. Then the reasons for the failure have to be investigated. Low classification rates can indicate problems with the selection of features or the classification parameters and methods, but they may also be due to the underlying category not having the structure that is assumed. It can, for example, be checked which works are repeatedly misclassified to find out which part of the corpus is not compatible with the textual genres that the classifier learned. Are there only a few works that do not fit? A closer look at them can serve to check if they can be seen as outliers, as texts participating in the conventional genre but not in the textual one because they are not compatible with the text type underlying the other works of the genre. What if a bigger group of texts is affected by misclassifications? This might indicate that there is more than one text type that is connected to the conventional genre or that no textual unity can be found for it at all. In such cases, alternative categorization methods can complement the classic classification methods to find out how the conventional genre in question is structured internally on the textual level. Considering the observations that have been made in literary theory and history about the individual relationships between literary works and “their” genres and about constant processes of modulation and transformation, it is very probable that there are such cases of conventional genres that cannot be neatly mapped to text types and that are not coherent textual genres.

Typological approaches to genre categorization are one of the alternatives that already have a tradition in literary genre theory. A typological description of genres starts from individual literary works seen as typical or especially representative of the genre in question. The genre’s characteristics are then derived from the exemplary texts by isolating, generalizing, or even exaggerating their properties. That way, an ideal type is constructed, constituting an extreme case when compared to other types (Strube 1993, 60–61). Following the cognitive-psychological and linguistic theory of the prototype effect, the ideal types have been called “prototypes”. The prototype effect describes general perception and cognition principles involved when humans categorize things. Some instances of categories are recognized faster than others because they have some typical characteristics. The typical features do not even need to be necessary for the category. For example, a prototypical table has four legs, but tables do not need to have this exact number of legs. When things are categorized, they are not checked systematically for present or absent features but are compared to different ideal types based on their attributes. The proximity or distance to an ideal type is decisive for perceiving something as a member of the category in question (Rosch and Mervis 1975; Taylor 2003, 41–62). The resulting categories are different from logical classes. Membership is not based on a set of necessary and sufficient conditions that can either be fulfilled or not and is not simply a question of belonging or not belonging. Prototypical categories have a core and an internal structure that runs from the core to the edges of the category in a continuous way. The edges are not sharp but fuzzy because membership is defined in terms of closeness to the prototypical core, which also entails that there are degrees of membership. As Hempfer, who sounded out the potential of prototype theory for literary genre theory, summarizes: Given that the difference between categories is constituted by the respective prototypical cores, the problem of ‘boundaries’ loses its poignancy because categories can indeed overlap at their edges when one entity variously resembles different prototypical cores in roughly equal measure. The key factor in the distinction between class inclusion and prototype theory lies in the terminological shift that relocates distinguishing characteristics of categories from their boundaries to their cores and attributes an entity to a category via its relationship of resemblance to the prototypical core. (Hempfer 2014, 412)

Hempfer finds that prototype theory is suitable for capturing the relationship of literary works with modes or transhistorical invariants. As an example, he discusses narrative versus dramatic communication. Paradigmatic (that is, prototypical) forms of narration are the epics of Homer or Cervantes’ “Don Quixote”, realistic novels in the nineteenth century, or popular twentieth-century fiction. On the other hand, there are also deviations from the prototypical norm, for example, when drama is narrativized, or narrative is performatized. Nevertheless, untypical instances of narrative or dramatic communication can still be recognized as such as long as they have a closer relationship to their respective prototypical core than to the core of the other category, which according to Hempfer, is also a question of “reference” (Hempfer 2014, 414–415). In this application of prototype theory to literary genres (in the form of modes), both the levels of the literary text type and the conventional genre are involved. The implicit or explicit reference to a generic prototype must be viewed as a genre signal indicating to which literary institution a text is assigned, while the amount of narrative or performative passages in a text is located on the textual level. This complicates matters because a literary work then has relationships to prototypical cores on two levels: that of the generic convention and that of the text type. Differentiating these levels, as is done in the present study, can mean that a literary work can be closer to a particular conventional prototype because the work carries the respective genre label or other signals pointing to it. At the same time, on the textual level, a work can be closer to a different prototype. As a result, it can be interpreted as an untypical member of two textual genres. As an untypical member of the textual genre (or mode) “narrative”, a narrativized drama can be closer to the narrative literary text type than to the dramatic one but be more distant to the conventional narrative genre than to the dramatic one. Vice-versa, as an untypical member of the textual genre “drama”, it is close to the core of the conventional genre but more distant to the core of the corresponding text type.

For the theory and history of literary genres, the idea of prototypically organized categories is especially attractive when generic terms refer to genealogical categories established on the basis of tradition, and in which the relationship between the literary works and the genre is characterized by hypertextual modulation (in the terms that Schaeffer (1983, 181) uses to describe this type of generic logic).Schaeffer defines hypertextual relationships as follows: “J’accepte comme relation générique hypertextuelle toute filiation plausible qu’on peut établir entre un texte et un ou plusieurs ensembles textuels antérieurs ou contemporains dont, sur la foi de traits textuel ou d’index divers, il semble licite de postuler qu’ils ont fonctionné comme modèles génériques lors de la confection du texte en question, soit qu’il les imite, soit qu’il s’en écarte, soit qu’il les mélange, soit qu’il les inverse, etc.” (Schaeffer 1983, 174). Often, literary works signal their participation in a conventional genre by referencing another work that is viewed as a prototypical representative of the genre, either as a particularly accomplished masterpiece or as the foundational text of the genre.See, for instance, Schnur-Wellpott’s (1983, 149–159) exposition of the two perspectives of a founding text and his followers versus a master and his predecessors. Examples of this are references to Walter Scott’s “Waverley” in historical novels or to Goethe’s “Wilhelm Meisters Lehrjahre” in education novels. Works that reference their prototype are expected to either imitate it closely or vary it to different degrees. In addition, a literary work can contain references to several different prototypical works. The prototypical genre categories that are grounded on hypertextual relationships are necessarily anchored in time and closely related to the literary-historical context. Literary works can only reference other works that existed before them, and the writers must be aware of the prototypes to which they allude. Furthermore, at least two instances of texts are needed to constitute a genre in this sense: the prototype and a text that references it.

If prototypical structures are established based on hypertextual relationships, they are, above all, defined as conventional genres. With the help of digital stylistics, it can be examined to what extent the conventional genres that are constituted by tradition and intertextual references have correspondences in literary text types. It can, for example, be checked whether traditional prototypical works also function as cores of genres on the textual level or not and how close or distant other works participating in the genre tradition are to them stylistically.

Such questions were analyzed by Henny-Krahmer et al. (2018) in an analysis of nineteenth-century Spanish-American novels. The novels are associated with the subgenres of sentimental, historical, costumbrista-, gaucho, and anti-slavery novels. For each subgenre, one or several works were predefined as prototypes, following literary-historical indications of their status as ideal types, either as predecessors or as masterpieces and culminations of the subgenres in question. Subsequently, textual similarities of all the novels to the different prototypical works were analyzed based on MFW and topic features. The analysis faced the difficulty that especially the predecessor prototypes can have their origin in other countries (for example, novels from Spain in the case of the costumbrista subgenre) and also in other languages (such as the English “Waverley” for the historical novel or the French “La Nouvelle Héloïse” for the sentimental novel). Furthermore, they can also be chronologically distant from the follower works. “La Nouvelle Héloïse”, for instance, was first published in the eighteenth century. To be able to compare works that were originally written in other languages, Spanish translations of them were used, but it is clear that such cultural, historical, and linguistic differences significantly complicate comparisons of literary texts on a stylistic level. Therefore, it is not surprising that there was a tendency for the follower works to be more similar to each other than each of them was to the respective prototype. In addition, the stylistic coherence of the works of individual subgenres was not clearly visible in all cases, which means that their integrity as textual genres could not be confirmed with certainty.

The experiment shows that literary-historical generic relationships are not necessarily translatable to a stylistic level. The results of the prototype analysis of the nineteenth-century subgenres of Spanish-American novels also suggest that works which are considered prototypical from a literary-historical point of view might rather be exceptional than typical for a textual genre. Therefore, it could be more useful to look for a prototypical textual core not in the sense of excellent, norm-founding masterpieces but ordinary, average works.

Considering how computational categorization methods in general relate to prototype structures versus logical classes, it can be noted that there is no direct and exclusive relationship between one method and one theoretical concept of genre categories. The two main types of text categorization methods that are used in machine learning and statistical text analysis are classification and clustering. As classification is a supervised method, its aim is to separate the data points in a way that fits best to the class labels provided with the data. As long as only one label is allowed per data point, the results are exclusive classes that do not overlap. So, at a first glance, it seems that statistical text classification is entirely congruent with the idea of treating genres as classes in the logical sense. Looking closer reveals that statistical classification procedures can also be related to prototypical structures. First, many classification algorithms include calculating scores or probabilities for the class assignments, which means that internally, information is collected about the degree to which an instance can be considered as belonging to one class or another, before the final decision for the output is made.See, for example, the availability of scores and probabilities for Support Vector Machines (SVM) or Random Forest Classifiers (RF) (Scikit-learn developers 2007–2023m, 2007–2023e). This internal data can be interpreted as information about the prototypicality of data points for classes. If classification is used to predict the genre of literary works, the probability with which a certain work is assigned to one genre or the other can also be understood as its proximity or distance to a prototypical core of the genre. Second, the features used for statistical classification are usually not only binary but numerical. When numerical and especially continuous numerical features are involved, decisions are not based on the presence or absence of features but on feature distributions that can be similar or more distant from each other. In the output, the classes have clear boundaries, but they have a complex structure internally. The internal structure of the classes is close to how the prototype effect is described theoretically in linguistics:

I will assume in the following discussion that entities are categorized on the basis of their attributes.1 These are not the binary constructs of the classical approach. Consider the ratio of width to depth. The ratio is a continuous variable. Labov’s results show that associated with each of the categories cup, bowl, and vase, there is a certain optimum value, or range of values, for the width-depth ratio. In categorizing an entity, it is not a question of ascertaining whether the entity possesses this attribute or not, but how closely the dimensions of the entity approximate to the optimum value.

1 From now on I shall restrict the term ‘feature’ for the abstract features of the classical approach, reserving ‘attribute’ for alternative, non-classical theories of categorization.

In statistical classification, optimum values and optimal combinations of values are determined quantitatively. If a model is trained with data that shows a concentration of similar values, value ranges, or value combinations for a class, this subspace of features can be interpreted as the prototypical core of the category to which new instances that are to be classified are related. However, the ultimate goal of supervised classification is not to produce models of prototypically structured categories but to define models that separate the data along the lines of the predefined labels that the training data has and to predict membership in distinct classes. Carried over to the definition of genres, it means that what is learned are text types that correspond as closely as possible to the conventional genres – assuming that the labels are expressions of generic conventions. The text types have clear boundaries, but are (usually) not constituted based on of binary features. The model of the text type can be applied to new literary texts to test whether they comply with it or not. True positives can be interpreted as literary works that belong to the text type and the conventional genre and hence to the textual genre. True negatives belong neither to the text type nor to the conventional genre nor the textual genre in question. False positives are part of the text type but not of the conventional genre and, therefore, also not of the textual genre. False negatives are part of the conventional genre but not of the text type and, consequently, also not of the textual genre. Probabilities can be used to check how certain the assignment of a text to the text type is and, thereby, how close it comes to the prototypical core of the text type.Another approach to using statistical classification for analyzing the prototypicality of individual texts with regard to genres is pursued by Konle in his master thesis about word embeddings for literary texts. Using 200-word segments of novels, Konle trained a neural network aiming to find words that are distinctive for different genres. The network is trained for the subgenres of sentimental, crime, science-fiction, and horror novels and is used to classify the segments by genre. When discussing the results, Konle analyses individual segments that were misclassified. He finds out that there are cases in which the misclassifications make sense because the novels are not prototypical for their genres in all parts and contain passages with words that are distinctive for other genres. These can be elements of the plot that are interpreted in terms of another subgenre, for example, passages about the death of characters in sentimental novels that are classified as horror novel segments (Konle 2019, 46–49, 59–64, 69–76). However, nothing is said about prototypes of the conventional genre in such a classificatory setting. Prototypicality on the level of convention (and also tradition) is inaccessible to text classification with stylistic features as long as it is not explicitly modeled in the metadata and available in the form of specific genre labels. The cognitive aspect of the prototype model, which involves recognizing instances of categories as a whole and not only as a sum of its parts, is also out of scope for text stylistic categorization: “Finally, it emerges very clearly from Labov’s experiment that no one single attribute, or set of attributes, is essential for distinguishing the one category from the other. Presence of a handle, or function in the drinking of coffee, merely raises the probability that an entity will be categorized as a cup” (Taylor 2003, 44). What can be done is to make clear on which level, textual or non-textual, literary texts are participating in genres or not. The crux with the proposed interpretation of classification results in terms of text types, conventional genres, and textual genres is that it presupposes a good quality of the model. If, for instance, only an overall accuracy of 60 % is reached – is this a sign that the model was not properly trained? Or is it a sign that the conventional genre expressed through the class labels does not correspond to a congruent text type? In such a case, further experimentation and examination of the results are needed, including an analysis of the historical genre conventions that are involved.

The second main type of computational categorization method, which is also often used in digital genre stylistics, is unsupervised clustering.For different implementations of clustering algorithms in Python, see Scikit-learn developers (2007–2023a). Practice-oriented introductions to clustering are Müller and Guido (2016, 170–209 and VanderPlas (2017, 462–476). A more theoretically oriented introduction to clustering algorithms is contained in Alpaydin (2016, 143–162). With this approach, the data is partitioned into a certain number of clusters so that data points inside of a single cluster are more similar to each other than to other data points lying outside of it (Müller and Guido 2016, 170). That the method is unsupervised means that no category labels are used to group the data. In the results, there can be different degrees of similarity between the points in a cluster, between points in different clusters, and also between whole clusters.This becomes especially clear in hierarchical clustering, an approach that divides the data into several levels of clusters, either in a top-down or bottom-up way. The latter is called agglomerative clustering and starts from merging the two most similar points, which again are merged to the next most similar cluster, and so on, either until a certain number of clusters is reached or until all points are merged into the same overall cluster (Müller and Guido 2016, 184). Hierarchical clustering is often used in stylometric analyses that are concerned with authorship attribution, so that it can be inspected on which close or distant level different authorial candidates are grouped together with the text in question (Eder 2017). For several clustering algorithms, the cluster centers, called “centroids”, are crucial for defining of the clusters. In K-means clustering, for example, the clusters are described by means of the data points contained in them. These means, which are in general not equal to any of the actual data points, are the cluster centroids (Scikit-learn developers 2007–2023b). For many clustering algorithms, the number of clusters to learn has to be set as a parameter value.This is the case with K-means, spectral clustering, Ward hierarchical clustering, or agglomerative clustering (Scikit-learn developers 2007–2023a). Some aspects of clustering methods remind us of prototypically organized categories. Each cluster can be interpreted as a category, and their centroids as the ideal prototypes, which represent (but are not equal to) the most typical exemplars of the category. Clusters are formed based on continuous distance relationships between data points and cluster centroids. This corresponds to the idea of prototype categories where some members are central and others marginal. A difference is, however, that clustering results in separate categories which have clear boundaries. However, for single data points, it is possible to calculate the distances to several centroids. That way the cluster boundaries can be crossed when the model is interpreted.

How could clustering algorithms such as K-means be applied to analyze literary texts and their genres as prototypical structures? The clusters that are found can be interpreted as literary text types that have a prototypical core. The texts that are members of the clusters can be seen as members of the text type with certain proximity or distance to the prototype. In contrast to statistical classification, though, the clustering algorithms determine the text types only through the analysis of feature values and distributions and not by optimizing the relationship between text types and conventional literary genres, because no genre labels are used in the clustering process. Once there is a model of clusters, genre labels can be applied afterward to evaluate to what extent the text types overlap with the conventional genres. The resulting intersections can be interpreted as textual genres. An expectation is that the conventional genres will tend to appear as textually more fragmented when the models are not forced to take into account the traditional genre labels. Therefore, the unsupervised categorization is better suited to detect discrepancies between conventional and textual genres than the supervised approach. However, it is less useful to learn which textual features constitute the common basis for specific conventional genres. Considering the status of the prototypes that are determined through the calculation of cluster centers, they are not to be confounded with literary works that have been described as prototypical and especially influential for a genre by literary historians or that were considered as such by contemporary writers. Instead, these are statistical means of text types and can be viewed as representing the most normal textual characteristics for a literary type. It can be concluded that clustering algorithms are closer to the idea of prototype structures than classification methods, but of course they are also limited to the aspects of categories that are detectable through the textual surface and text style.

Besides prototypical structures, another alternative approach to conceptualizing genre categories is discussed here: family resemblance networks. The idea to describe categories in analogy to family resemblances goes back to Wittgenstein. In his linguistic-philosophical work, the concept serves to describe linguistic activities involving the use of the same word for phenomena which only have partial and indirect similarities, like the resemblances that can be seen between different members of a family: Consider, for example, the activities that we call ‘games’. I mean board-games, card-games, ball-games, athletic games, and so on. What is common to them all?—Don’t say: ‘They must have something in common, or they would not be called ‘games’’—but look and see whether there is anything common to all. For if you look at them, you won’t see something that is common to all, but similarities, affinities, and a whole series of them at that [...]. And the upshot of these considerations is: we see a complicated network of similarities overlapping and criss-crossing [...]. I can think of no better expression to characterize these similarities than ‘family resemblance’; for the various resemblances between members of a family—build, features, colour of eyes, gait, temperament, and so on and so forth—overlap and criss-cross in the same way.—And I shall say: ‘games’ form a family. (Wittgenstein 2009, paras 66–67)

Here, the relationships between members of a group are conceptualized as a network of overlapping – but not in all cases common – similarities. The family resemblance analogy was adopted in literary genre theory already in the 1960s. It became popular because it allowed for more open definitions of genre, able to overcome the necessity that the features relevant to a genre should be present in all literary works attributed to it. Such an open concept of genre categories is especially useful for capturing changes in genres over time. These often entail that the same generic term is applied to literary works with quite different textual characteristics between one period and another. The gradual process of change can be understood as a chain of links between individual works in a network, which can be followed. While immediate neighbors in the network share some features, this must no longer be true for distant relationships. However, the details of applying the notion of family resemblance to literary genre theory are not always laid out, even if the approach is mentioned as a possible option by literary theorists. Often, the term is invoked briefly as a counter-concept to logical classes, which are seen as insufficient for the theoretical representation of historically variable genres, for example by Jauß: The continually new ‘widening of the genre’, in which Croce saw the supposed validity of definitional and normative genre concepts led ad absurdum, describes from another perspective the processlike appearance and ‘legitimate transitoriness’ of literary genres, as soon as one is prepared to desubstantialize the classical concept of genre. This demands that one ascribe no other generality to literary ‘genres’ [...] than that which manifests itself in the course of its historical appearance. By no means must everything generically general – what allows a group of texts to appear as similar or related – be dismissed [...]. Following this line of thought, literary genres are to be understood not as genera (classes) in the logical senses, but rather as groups or historical families. (Jauß 2014, 131)

Jauß’s statement shows that the recourse to the family resemblance concept also serves to defend the idea of literary genres per se against the critics who question it entirely. Critics make reference to the constant processes of transformation which are at work when individual works refer to genre conventions and when these individual works contribute to the genre’s modification at the same time. More recently, there has been some consensus that literary genres are not to be equated with logical classes. Fowler gives an overview of the genre concepts that prevailed in the history of genre theory. He takes the position that understanding genres as a means of classification is an error. He sees a practical value in using genres for taxonomic efforts but argues that the principal value of genres is functional: “genres have to do with identifying and communicating rather than with defining and classifying. We identify the genre to interpret the exemplar. [...] If we see The Jew of Malta as a savage farce, our response will not be the same as if we saw it as a tragedy” (Fowler 1982, 38).

Here, a less critical view of the potential textual reality of genres is maintained. To view “The Jew of Malta” either as a savage farce or as a tragedy would mean to view it in light of two different conventional genres connected with different discursive expectations towards the text. In a stylistic analysis, the text would be grouped together with other instances of the conventional genre in question. Depending on the perspective formed by the selection of other texts and features, different textual clues and “normative facts” can be found as traces of the conventional genres. This does not mean that the chosen generic perspective must lead to the definition of a textual genre based on necessary common features and a definition that fits all exemplars associated with the corresponding conventional genre. However, it is also not excluded from the outset that there can be textual similarities in addition to to the conventional genre’s functional and non-textual communicative characteristics. Fowler goes on to say that the missing necessary common features of genres make their definitions impossible: “Either defining characteristics are absent altogether, or they are limited to meager distinctions that do no more than subdivide the genre. In short, genres at all levels are positively resistant to definition” (Fowler 1982, 40). The theory of family resemblance then appears as a solution that allows for describing connections between literary texts participating in the same genre without needing to define what their common characteristics are: “It promises to apply not only to close-knit connections within subgenres [...] but also to far-flung resemblances between widely divergent works [...]. Genres appear to be much more like families than classes” (Fowler 1982, 40). It turns out that Fowler puts definitions on a level with logical classes. He states that descriptions of genres in terms of family resemblances should not be seen as preparations of genre definitions as classes but as descriptions of “a different sort of grouping, not reducible to a class” (Fowler 1982, 42). Although Fowler advocates for conceptualizing genres as families, he sees a need to modify Wittgenstein’s concept of family resemblances. According to Fowler, in its original form, it would entirely impede generalizations about literary forms. Fowler criticizes that Wittgenstein focuses on “directly exhibited resemblances” and suggests instead focusing on the traditional links between literary works: “What produces generic resemblances, reflection soon shows, is tradition: a sequence of influence and imitation and inherited codes connecting works in a genre. As kinship makes a family, so literary relations of this sort form a genre” (Fowler 1982, 42). Here Fowler clearly addresses the conventional level of genres. Traditional references between the works are a prerequisite to treating them as members of a family, which limits the possibilities of the network to cover just everything. However, it also means that the conventional genre precedes literary text types. Comparing Fowler’s concept to Fricke’s definition of genre, the level of convention is not dependent on the textual level but the other way around. Still, it is not entirely clear how Fowler delimits a common tradition (and a common generic convention) because he cautions: In its modified form, the theory of family resemblance also suggests that we should be on the lookout for unexhibited, unobvious, underlying connections between the features (and the works) of a genre. As with heredity, with generic tradition too we have to expect quite unconscious processes to be at work, besides those that readers are aware of. It would be strange if genre did not in part operate unconsciously, like other coding systems within language and literature. (Fowler 1982, 43)

His remark on the possibility of unconscious processes that lead to shared features and resemblances between works again points to the textual level because conventional markers of genre are unlikely to be unexhibited. So the common literary tradition must be understood in a wide sense, for example, as a common temporal, geographical, or cultural context. Usefully the common literary tradition can also be understood as a common major genre in which the works that are part of the family network participate, but then the traditional links are quite loose again.

By analyzing literary works in such a broad frame of common tradition, for instance, by building a corpus that represents it and categorizing the texts based on textual features, digital genre stylistics can contribute to uncovering the underlying connections between the texts. In order to formalize the family resemblance concept, this needs to be done not by presupposing necessary common features or feature values and distributions that are similar for all the texts in a group at once but by allowing indirect and overlapping similarities. This is possible with network analysis. A proposal for such a textual “family resemblance analysis” is made in chapter 4.2.2.2, where it is applied to subgenres of nineteenth-century Spanish-American novels. The families are subgroups in the entire network, and they can be delimited because some regions in the network are interconnected more closely than others. With that, it is possible to define literary text types based on family resemblance relationships. These text types can be related to conventional literary genres so that non-classificatory textual genres become definable. The recourse to stylistic networks as an alternative method for category building solves several problems: the idea of family resemblance networks can be applied in the analysis of textual relationships between literary works, and there is no need to limit the concept to tradition, convention, and non-textual communicative functions (to which it is of course also applicable), the excessive openness of a network that never ends can be remediated by first choosing a corpus of literary works that represents their common broader traditional and/or generic background and second, by identifying subparts of the entire network that represent (sub)families, the inability of classification and clustering algorithms to build categories on top of indirect similarities is overcome.

Using the family resemblance concept not only as an abstract counterpart to logical classification but also to formalize it can help enhancing its reputation in literary genre theory. It has been criticized there for being a handy slogan and also because the boundaries between different categories are not defined sharply.An early critique was formulated by Vivas already in 1968. Vivas discusses several aesthetic, social, and philosophical reasons for the unpopularity of the idea of genres as classes. He criticizes the family resemblance notion as a loophole to avoid the question of the genre’s nominalistic or realistic status: “Taken seriously, nominalism involves the notion that structures have no status in being whatever. But how a totally invertebrate world is possible I have never been able to understand. ‘There is,’ you may say, ‘a new solution to this old problem.’ ‘Yes, I know,’ I reply, ‘there is a newfangled one: It is the evangel of Saint Ludwig.’ According to these glad tidings the members of a class share among themselves, not identities but family resemblances. Obviously I cannot stop to analyze this newfangled solution here. Let me merely lay it down that between two members of a family the resemblance is that of shared identity. We are therefore not farther along than we were before” (Vivas 1968, 101). Nevertheless, Vivas defends the idea of genres as open concepts. Fishelov, for example, describes the takeover of the family resemblance concept by genre theorists as the attempt to find a philosophical foundation for “the dominant trend in modern critical theory, with its stress on the flexible and dynamic nature of literary genres” (Fishelov 1993, 54). He criticizes the concept as too open because it cannot explain readers’ relatively high consensus about boundaries between different genres. Subsequently, Fishelov again looks for necessary conditions for differentiating genres that others have described in terms of the family resemblance notion, such as the tragedy or the novel (Fishelov 1993, 55–68).Hempfer, in turn, criticizes that the common features that Fishelov finds may be necessary but not sufficient (Hempfer 2014, 409–410). Fricke instead proposes to use combinations of necessary and optional features in “flexible” genre definitions. These should overcome both the definitions of genres as combinations of necessary features, which he sees as too rigid to capture historical change, and definitions in terms of family resemblances, which in his opinion, are too loose. Instead, he proposes structures such as, for example, “[1] + [2] + [3] + [4a u/o 4b] + [5a u/o 5b u/o 5c]” (Fricke 2010, 9) for the definition of the anecdote, meaning that the first three features are necessary, plus at least one of the variants of the features 4 and 5, respectively. So Fricke also goes back to necessary features but complements them with additional sufficient or “alternative necessary” features.

Nonetheless, Hempfer emphasizes the potential value of both the prototype categories and the family resemblance networks for genre theory: “I believe that genre theory within literary studies can, on the basis of the concepts of family resemblance and prototypes, manage to realign key questions, especially those arising from the polysemy and historicity of genre concepts” (Hempfer 2014, 414).As an example of the application of the family resemblance concept, Hempfer describes the history of the elegy, a genre that was originally only identifiable metrically and later by a number of other traits, i.a., intertextual references, and motifs (Hempfer 2014, 416–417). Hempfer concludes: “The diachrony of the genre can best be represented as a synchronic network of relations, in which each individual text or epochal version of the genre is linked to other historical versions through common features. [...] The genre identity, then, is not produced by a single trait but by the entirety of all relations among their historical versions” (Hempfer 2014, 419). For an application of the family resemblance concept to genre theory, see also Strube, who interprets a definition of the novella set up by Seidler in that way (Strube 1993, 21–25). This optimistic line is followed here. To start with, all three concepts of genres as categories – genres as logical classes, prototype categories, and family resemblance networks – are seen as heuristic tools that can be employed in digital stylistic analyses of genre. They can be used to look for literary text types and their connections to and overlaps with conventional genres in the form of textual genres. In the previous chapter on the system and history of genres, the issue was raised about where to draw a line between different literary text types and different textual genres in relationship to conventional genres. This question can have different answers depending on the chosen categorization method.

With a classificatory approach, text types and textual genres will be more closely attached to the conventional genre expressed by the genre labels so that some internal variation in the category might be covered. With unsupervised clustering, the expectation is that if the number of clusters is optimized and not fixed beforehand, there will be more text types than conventional genres. The text types are expected to represent the genre-internal or cross-genre textual variation more closely but are more distant from conventional genre groupings. The same is assumed for a family resemblance network analysis, only that this approach is even more flexible in allowing for partial similarities, which might lead to text types that can be related to conventional genres in a more meaningful way. However, these assumptions need to be checked and substantiated by empirical work to determine which of the concepts suits which historical genres best on the textual level. Even if it is expected that the more open concepts of prototypes and family resemblances are, in general, better suited to describe the internal structure and variability of historical, literary text types, classification has its value. It can help to find a good feature basis by which the text types can be linked to conventional genres and use that basis as a starting point for the more open procedures. Classification (see chapter 4.2.2.1) and a family resemblance network analysis (chapter 4.2.2.2) are employed in the empirical part of this study to analyze the subgenres of nineteenth-century Spanish-American novels.

All the three techniques that were proposed here as possible implementations of the three genre category concepts – statistical classification, clustering, and network analysis – have been employed in digital genre stylistics. Even other, non-categorizing approaches, for example, Principal Component Analysis (PCA), are common to explore how literary texts appear as groups in features spaces.Principal Component Analysis (PCA) is a technique for dimensionality reduction that projects the data points onto so-called “principal components”, which aim to preserve as much variation of the data as possible. The number of dimensions that the data has can be reduced by only considering the resulting principal components further. In digital genre stylistics, it has, for example, been used by Schöch to visualize how French classical tragedies, comedies, and tragicomedies distribute over principal components based on topic features (Schöch 2017c, paras 33–41). Schöch groups the PCA analysis under the heading “Clustering”, as does Oakes in his general introduction to statistics for corpus linguistics, because the data is grouped based on similarity (or distance) relationships. Oakes also uses the term “categorization” for clustering methods (Oakes 2003, 95). Here, clustering and classification are considered categorization methods (in the general sense of category building). However, PCA is not because the data points as a whole are not assigned to separated text categories. A related method is Factor Analysis, which Biber used to find groups of feature distributions that serve as the basis for defining functional text types (Biber 1993b). Even so, the relationship between these methods and the literary theoretical considerations about genre categories has not been analyzed much yet. Thoughts about prototypicality, the historical variability, and social embedding of literary genres have been expressed and connected to statistical text analysis,See, for instance, Underwood’s approach to genre via the history of reception or Schröter’s proposal to apply machine learning methods to reconstruct the historical change of disordered genres such as the German Novelle (Underwood 2016; Underwood 2019, 34–67; Schröter, forthcoming). but the family resemblance concept, for example, has not been explicitly formalized in digital genre stylistics so far. It is important to link key discussions of literary genre theory to digital genre stylistics so that research results of the two areas are confronted more deeply. This can increase the interest of the two disciplines in each other and also challenge the findings achieved in both of them. As with the definitions of text types and genres, also the concepts of genre categorization and the possibilities for their implementation are crucial points of contact between literary genre theory and digital genre stylistics.

Characteristical of the historical novel is the crossing-over of fiction and history. Historical events, places, persons, and conditions of a past epoch are represented and arranged narratively (Álamo Felices 2011, 84). While factual history aims to reconstruct a past period by identifying and recording significant personalities and events and by tracing the development of social institutions, historical novels serve to revitalize historical characters and the world they lived in (Read 1939, ix). As a literary subgenre, the historical novel gained considerable popularity during the nineteenth century through the novels of Walter Scott, also in the sphere of the Romance languages and cultures and in the Spanish-American countries (Dill 1999, 131; Janik 2008, 64–67; Lefere 2013, 17; Maxwell 2009). However, in the Spanish tradition, historical fiction also has remote origins in the historical narratives of the Spanish conquests in the Americas (Read 1939, 1–28). In general, several characteristics have been pointed out to be constitutive of the historical novel (Fernández Prieto 1996; Lefere 2013, 17–62; Spang 1998): temporal distance between the writing, publication, and reception of the novel and the past in which the narrated events take place co-occurence of invented and historical personages, places, and events localization of the narrated events in a precise historical past

According to the bibliographical data that was collected for this study, the historical novel was the most frequent thematic subgenre in the nineteenth century. Furthermore, many historical novels were explicitly marked as such in series or subtitles, which shows that the subgenre was well established on a conventional level.See chapter 4.1.5.3.1 for an overview of the proportions of thematic subgenres in the bibliography and the corpus and chapter 4.1.5.1 for a list of the most frequent explicit subgenre labels. In the Spanish-American historical novels, typical topics are the period of conquest, the colonial era, and the wars of independence. Especially the latter are not part of a remote historical past, but in the case of Mexico and Argentina, they took place in the early nineteenth century itself, and for Cuba towards the end of the century. In some historical novels published in Argentina, Cuba, and Mexico, a European setting is chosen, for example, from Antiquity or the Middle Ages, but most are concerned with local history, which is temporally more immediate. This makes the Spanish-American historical novels special with regard to the first general defining aspect of temporal distance because that distance is not given in some cases. However, the novels are often explicitly declared historical, even if the events are contemporary or took place in a very recent past. Even if the more distant past of the conquest or colonial period is treated, this is done differently than in European historical novels: “Scott, for instance, tried to escape from his century and return to a spirit of the past; Mexican historical novelists who dealt with the distant past attempted to interpret that past in terms of their own nineteenth-century [liberal] thought” (Read 1939, 58). Nevertheless, nineteenth-century Spanish-American historical novels can be divided into two groups: romantic historical novels dealing with the conquest and colonial times on the one side and the ones dealing with contemporary historical events on the other (ix). Read calls the latter “novels of contemporary history” (x) and Molina “prospectively historical novels” (“novelas prospectivamente históricas”, Molina 2011, 285–312).

Another typical element of romantic historical novels is a sentimental plot. The most prominent example is the novel “Amalia” (1851–1855) by the Argentine José Mármol, which tells the story of a group of resistance fighters against the dictatorship of RosasJuan Manuel de Rosas (1793–1877) was a governor of the province of Buenos Aires who established a dictatorial system marked by repressive measures that lasted between 1829 and 1852 and that enforced a political and economic hegemony of Buenos Aires over the other provinces. and of the protagonist’s tragic love relationship (Dill 1999, 127). As Dill states: “Amputierte man den genannten Romantypen [politischer Roman und historischer Roman] ihre politisch-sozialen Teile, würde der private Part samt Gefühlswelt der bürgerlichen Protagonistinnen, d. h. der sentimentale Roman, übrigbleiben” (Dill 1999, 138). However, the historical novel was not only present in the romantic period. It continued to be practiced towards the end of the nineteenth century and also in the twentieth century. It was, for example, influenced by Spanish realist and French naturalistic authors. The Mexican novel “Los bandidos de Río Frío” (1892) which was written by Manuel Payno, for instance, carries the subtitle “Novela naturalista, humorística, de costumbres, de crímenes y de horrores”. That it has been classified as a historical novel despite five other explicit subgenre labels in its subtitle marks it as one of the cases with a clear discrepancy between the generic convention, the textual form, and the critical tradition that the works have been seen in.Read (1939, 260) calls it “the best historical novel of the nineteenth century.” Another example of realist historical novels are the works belonging to the series “Episodios Nacionales Mexicanos” (1903) authored by the Mexican Victoriano Salado Álvarez (Fernández-Arias Campoamor 1952, 84–85; Read 1939, 293–303). Read characterizes them as follows: In the collection of material for his work Salado Álvarez exhausted accessible periodicals of the period, and gleaned carefully the memoirs of acquaintances and little known books and documents, in search of human aspects of the period and its chief characters, attempting to bring to light fresh information instead of revamping the threadbare stories that constituted the patriotic equipment of various partisan groups and that had been the chief source of of such writers as [the late romantic author] Juan A. Mateos. (Read 1939, 295)

In sum, several aspects can be noted regarding the expected textual coherence of nineteenth-century Spanish-American historical novels. First, the subgenre is long-lasting, there are numerous works attributed to it and it was recognized as a subgenre by contemporaries. These factors suggest that the subgenre is also stylistically distinguishable from other subgenres. What might make its textual classification difficult, though, is that it is often mixed with a sentimental plot and thus has characteristics in common with the sentimental novels, as well. Furthermore, the novels of contemporary history treat similar political and social subjects to other types of novels that are not designed and communicated as historical ones. In addition, stylistic differences can be expected between the early romantic versus the later realist and naturalistic historical novels. Moreover, the range of different special topics (conquest, colonial era, wars of independence) might lead to a subdivision of the textual genre into several subtypes, at least if semantic features are used to categorize the texts.

The novelas de costumbres (also called costumbrista novels, novels of customs, or novels of manners) derived from short pieces of prose, the so-called “cuadros de costumbres”, which had a French origin and were very popular in Spain before they reached Spanish America.In the Mexican case, a different view on the costumbrista tradition sees the origin of the novels of customs not in Spanish models but in the early works of the Mexican author Fernández de Lizardi (Calderón 2005, 316–317). These prose texts were mainly published in journals and periodicals. In them, life in urban or rural settings is described, and the male and female types of different social strata are portrayed with their habits. Furthermore, traditional festivities are a common topic in the texts. The cuadros evolved to full novels retaining the same characteristics. Observing the particular, the traditional, and the vernacular were important concerns of the novels of customs. They were, therefore, well suited to take up elements of the Spanish-American reality of life and contribute to the development of a national novel. Stylistically, they are characterized as creating a local color, for instance, by including colloquial speech and creating a picturesque effect, which links them to picaresque novels. Moreover, a humorous or critical-ironic attitude is typical for the novels of customs in their romantic form. In Spanish America, they were also used as a vehicle for political purposes (Janik 2008, 60–64).

As in the case of the historical novels, also the novels of customs were often explicitly designated as “novelas de costumbres”.“Novela de costumbres” is the second most frequent explicit label for the thematic subgenres. Independently of the explicit label, the novels of customs on the sixth rank of the most frequent primary thematic subgenres in the bibliography and on the third rank in the corpus. See chapters 4.1.5.1  and 4.1.5.3.1 for the corresponding overviews. The label was used by convention but also with specific, explicitly formulated purposes. The Mexican author José Tomás de Cuéllar, for example, was a prolific writer of novels of customs. He produced a whole series called “La linterna mágica” (published between 1871 and 1892). The first novel of the series is preceded by an introduction explaining Cuéllar’s literary program. The magical lantern is a symbol for the purpose of the novels of customs. Cuéllar claims that he first happened upon the expression in a small corner shop of some Mexican village. This is already indicative of the aim to represent local and rural customs. The lantern illuminates all kinds of spots the author visits, showing the multitude of ordinary people he wants to portray. It enlarges the view on vices and defects and, at the same time, minimizes the size of each person so that they come to the fore above all as social groups. The reader is invited to follow the lantern’s light which accepts the insufficiencies of the people with humor and also presents alternative models of virtue (Cuéllar 1890, xii–x). The consciousness and the frequency with which the term “novela de costumbres” was employed show the strength of the historical generic convention. However, the novels of customs are not detached from other prominent thematic subgenres. Novels can be at the same time historical and of customs, as for example “Calvario y Tabor.  Novela histórica y de costumbres” (1868, MX) by Vicente Riva Palacio or “Julia” (1868, MX) by Manuel Martínez de Castro, which had the subtitle “novela de costumbres mexicanas” in the first edition and “novela histórica y de costumbres” in the second one. There is also a close relationship between novels of customs and sentimental novels, as Janik points out: Die literarische Darstellung von Natur, Landschaft, landwirtschaftlicher Arbeit, von Gutshäusern und Gutsverwaltung, ebenso von städtischem Leben und städtischen Sitten sollte die gekannte Wirklichkeit ansichtig machen, doch eingebunden in eine tragende Handlungsstruktur. [...] Die Hauptaufgabe, die sich jedem Autor stellte, war die erzählerische Entwicklung eines komplexen Geschehens mit dramatischen Peripetien. Im Gefolge der genannten französischen Autoren [Balzac und Stendhal] und ihrer ‘romantischen’ Wegbegleiter (Rousseau und Chateaubriand) bildet in einer ganzen Reihe von Werken die Entstehung einer tiefen existentiellen Liebesbindung [...] die Grundstruktur. (Janik 2008, 67–68)

A prominent example of such a combination of the novela de costumbres and the sentimental novel is the Cuban work “Cecilia Valdés o La Loma del Ángel” (1839/1882) by Cirilo Villaverde, which contains the opposing tendencies of a romantic love story and a political stance focusing on the Cuban reality in the 1830s (Janik 2008, 75–77). Although they form mixed types with other romantic subgenres, the novels of customs still represent early forms of realistic texts. In the later nineteenth century, they entered into similar relationships with the realist and naturalistic novels. As Kohut notes, the delimitation of Romanticism, Realism, and Naturalism is difficult in Spanish America for several reasons, one of them being the Costumbrismo as an element between the different literary currents: Zum Realismus gehört die Zuwendung zur Gesellschaft, zur Romantik die häufig idyllisierende Perspektive. [...] Wichtiger als der Costumbrismo als eigenständige literarische Richtung ist die entsprechende Einfärbung zahlreicher realistischer bzw. naturalistischer Romane. So gab der Chilene Alberto Blest Gana seinem Roman Martín Rivas (1862) den Untertitel Novela de costumbres político-sociales, der Argentinier Lucio Vicente López seinem Roman La gran aldea (1884) den Untertitel Costumbres bonarenses. (Kohut 2016, 196)

Like the historical novel, also the novela de costumbres is a primarily thematic subgenre that was vital in the whole nineteenth century, although it was most popular in the Romantic period. Because one of its aims is to depict local habits realistically and faithfully, including linguistic peculiarities, there may be stylistic traits that help to distinguish this type of novel from the other thematic subgenres. However, in the literary-historical characterizations, this subgenre appears as one that enters into fusions with quite distinct other subgenres, which might make it difficult to classify it on the basis of features related to themes.

The third thematic subgenre that was chosen for text analysis is the sentimental novel proper. Contrary to the historical novel and the novela de costumbres, sentimental novels are usually not marked with explicit subgenre labels by authors or editors. In the whole bibliography and corpus that were compiled for this study, there is just one work carrying the subtitle “novela sentimental” – the novel “El canto del cisne (Novela sentimental)” (1910, AR) by Roque C. Otamendi. From the point of view of historical genre convention, the status of sentimental novels is, therefore, different from that of historical novels and novels of customs. It was a subgenre that was signaled in more subtle ways, for example, by using female first names as titles, e.g., “Soledad” (1847, AR) by Mitre, “Esther” (1851, AR, Cané), or “Clemencia” (1869, MX) by Altamirano. Furthermore, a sentimental plot is present in the majority of the nineteenth-century Spanish-American novels, which makes this subgenre one that is part of a general generic repertoire of the time. It can appear in a pure form or as a basic element to provide a narrative structure for novels that have other, additional, or superordinate thematic objectives. It can, for example, motivate descriptions of historical and contemporary conflicts and settings. Because of the omnipresence of sentimental elements in different kinds of novels over time, the sentimental novel has also been called a “metagenre”, which might challenge the classification of novels into subgenres (Varela Jácome [1982] 2000, sec. 1.4; Zó 2015, 14–16).

In pure form, the plot of a sentimental novel is centered on a highly personal conflict with exceptionally sensitive protagonists whose development of emotions is shown in the novel. Usually, the conflict is resolved in either a tragic or a happy end (Molina 2011, 375–386). Typical structural elements of sentimental novels are letters or passages of diaries included in the text, which give insight into personal communication and the protagonists’ mental world and sensitivities. As the historical novel, also the sentimental novel has European, especially French models that the Spanish-American authors followed, for instance, the epistolary novel “Julie ou la Nouvelle Héloïse” (1761, FR) by Rousseau, “René” (1802) by Chateaubriand, or the autobiographic novel “Graziella” (1852, FR) by Lamartine, but also the German epistolary novel “Die Leiden des jungen Werhers” (1774) by Goethe. In the sentimental novels proper, the protagonists are described in idealizing terms, emphasizing their moral superiority and tending to a flowery and metaphorical style. The heroine of Mitre’s novel “Soledad”, for example, is described as “imagen escapada de las telas de Rafael”, “un serafín bajado del trono del Señor”, and “la estatua de la castidad meditando”. Not only the protagonists but also the settings are described from subjective perspectives (Varela Jácome [1982] 2000, sec. 1.4). The novels are often written in the first person.See the overview of thematic subgenres by narrative perspective in chapter 4.1.5.3.1. If they do not have a European setting, the action is commonly placed in the bourgeois milieu of the Creole middle or upper class and thematically focused on private life, concerned with the home, love, and family (Dill 1999, 138).

The prototypical sentimental novel should have a recognizable textual coherence, as it has its own set of structural elements, theme development, and stylistic means. The assumption is that it forms a separate text type besides the historical novels and the novels of customs that have a more realistic style. Moreover, the historical novels almost exclusively have a third-person narrator. This can have a significant influence on the novels’ style, at least if the textual features used for categorization include personal pronouns and declined verb forms. The novelas de costumbres that are part of the corpus, in contrast, are half narrated in the first and half in the third person, which supports the observations made by Janik and Kohut about their intermediate status. As for the sentimental novel in a wider sense, that is, novels with a primary sentimental plot, but, for example, a historical setting or social and political concerns, the textual distinction of the various subgenres is assumed to be more difficult.

In Spanish America, Romanticism was the dominant literary current in the second third of the nineteenth century and continued to influence the production of novels until the end of the century. Compared to the development of the romantic current in the European countries and North America, its acclimatization in Spanish America was delayed because of the wars of independence and subsequent political struggles such as civil wars and oligarchic regimes. The cultural sector was not fully developed yet, local cultural models were missing, and also the ideological climate was different. Nevertheless, foreign models of various kinds entered the cultural area of the Spanish-American countries discontinuously and were soon adopted, for example, French sentimental novels, or the historical novels of Walter Scott and James Fenimore Cooper (Dill 1999, 120; Lichtblau 1959, 65; Varela Jácome [1982] 2000, sec. 1.1.3). In Spanish America, the romantic current is connected to political ideas of liberalism and nationalism. The writers are “civic poets” and not primarily literati, and they use the novel to denounce social and political grievances, promote ideologies, and suggest new models of society (Dill 1999, 121).

Formally, romantic novels are characterized by looseness, unconventionality, and the rejection of formal rigor. The protagonists are idealized heroes, who are often solitary, rootless and driven by instincts. Emotions and motives of the characters tend to be represented in stereotypical ways, and there is a tendency to reduce conflicts and rivalries to simple contrasts between good and bad characters, whose role is underlined by employing a corresponding oppositional descriptive vocabulary (Dill 1999, 120, 127–128). In the novels, verisimilar elements can be combined with fantastic ones. Typical elements of the plot are turbulent and violent actions, deception, fraud, and surprising effects, for example, through disguises and discoveries (Varela Jácome [1982] 2000, sec. 1.4). Even passages of pure description, for example, those concerned with representing the local landscape, are characterized by the use of a subjective and idealizing style (Lichtblau 1959, 66).

The romantic current put forth several distinct types of subgenres of the novel, among them sentimental, historical, political, social, adventure, indianist novels, and also the novels of customs. That different subgenres of the novel played an important role in the period in which the romantic current was dominant is readily apparent from the structure of many literary histories concerned with Spanish-American nineteenth-century literature. Usually, the description of the romantic novel is subdivided into separate sections concerned with the different subgenres, which is not always the case for the later currents.Dill, for example, structures the chapter on the romantic novel into the subsections “Der politische Roman”, “Der historische Roman”, “Der indianistische Roman”, “Der kubanische negristische Roman”, and “Der sentimental Roman”. The chapter on the realist novel is not subdivided, and the one on the naturalistic novel only has a subchapter concerned with the city novel (“Der Großstadtroman”) as a special type of naturalistic novel (Dill 1999, 125–139, 159–166, 168–176). In the introduction to her book on the Spanish-American romantic novel, Suárez-Murias lists the sentimental novel, the indianist novel, the historical novel, the costumbrista novel, the Roman à thèse (novela de tesis), and the dime novel (novela de folletín) as subgenres of the romantic current (Suárez-Murias 1963, 12–13). Gálvez, who studies the Spanish-American novel up to 1940, structures the chapter on the novel of the romantic period into subchapters on the historical and the political, the indianist, and the sentimental novel. She dedicates another subchapter to the novel of the transition to realism. That chapter includes parts on the historical, social and costumbrista novel, the latter including the gaucho, indio, and antislavery novel. As she is concentrating on the novel alone, Gálvez’s account is more differentiated than the one of Dill. She takes up several subgenres again in the chapters on the later realist, modernist, and regionalist currents, for example, the historical novel and the novel of customs, showing that they did not cease to exist, but continued to be practiced under the influence of different literary currents (Gálvez 1990). Nevertheless, the later currents did not produce the same range of new, own, distinguishable, and widely recognized thematic subgenres as the romantic current did.

Regarding the textual coherence of the romantic novels, several different kinds of hypotheses can be formulated. Thematically, the inclusion of a sentimental story is a typical element of all kinds of romantic novels and might facilitate its classification against realist and naturalistic novels. Even if a sentimental plot also occurs in novels belonging to the other currents, it is less typical. Then again, the clearly distinguishable thematic subtypes of romantic novels lead to the assumption that the conventional romantic subgenre might fall into several textual subgroups if semantic features are used. In the narrow sense of linguistic style, though, it is probable that the romantic novel is coherent. A hypothesis is that specific parts of speech, vocabulary, and punctuation marks are used to express subjectivity and emotionality, for instance, many personal and possessive pronouns, qualifying adjectives and adverbs, interjections, or exclamation and question marks.

What makes a clear distinction of works by current difficult is that there are individual works that combine elements of several currents. Especially romantic and realist elements are often combined, which is an effect of the chronologically discontinuous integration, adaptation, and further development of foreign currents in the Spanish-American context.On the simultaneous presence of romantic and realist elements, see Lichtblau (1959, 66) and Varela Jácome ([1982] 2000, sec. 2). Furthermore, the presence of romantic works up to the end of the nineteenth century raises the question of whether there is a contrast between an early and a late romantic style which could become visible in the categorization of the texts.

In its artistic meaning, the term “Realism” emerged in the eighteenth century to designate an aesthetic model opposed to the idealism and individualism propagated by Romanticism (Álamo Felices 2011, 118). In Europe, it began to spread in the first half of the nineteenth century and dominated from the middle of the century onwards. In Spanish America, in contrast, numerous realist novels were only published after 1880 (Varela Jácome [1982] 2000, sec. 3). The realist movement was linked to the formulation of positivist philosophical theories and the development of the natural sciences. According to the positivist philosophy, new insights should derive from positive findings that are sensually graspable, actually observable, and verifiable. For the arts, this meant that their primary objective was to reflect the social reality aesthetically by accurately depicting the characters and describing the environments meticulously. In realist novels, dialogues are an important means to develop the action. They are represented in detail but without the need to put aside the omniscient perspective. In sum, the goal was to create an effect of verisimilitude, an effet de réel, which aimed to make the reader believe that what is described is the real world. By convention, a relationship of identity is assumed between the contemporary world and the mimetic fictional world. To what extent the author is understood as contributing to the creation of the fictional world with his constructive and imaginative work depends on how closely the ideal to reflect the external world directly and mimetically is followed (Álamo Felices 2011, 118–119).

That the depicted world is not shown through an idealizing filter in the realist novels entails that not only the bourgeois milieu is represented but also lower social strata. Social and political ills, such as corruption, intrigues, lobbyism, or poor education, are disclosed, for example, in the novels of the Mexican writer Emilio Rabasa. The portrayal of personal defects, such as greed, excessive materialism, or drive for recognition, is as well part of the realist repertoire, for example, in the Cuban novels “Un hombre de negocios” (1883) by Nicolás Heredia or “Mi tío el empleado” (1887) by Ramón Meza (Instituto de Literatura y Lingüística de la Academia de Ciencias de Cuba 1999, sec. Realismo). On the one hand, the Spanish-American realist novels described the rapid economic and social development and life in the big cities, for instance, in the novel “La gran aldea. Costumbres bonaerenses” (1884) written by the Argentine Lucio Vicente López. On the other hand, rural environments were also a topic, for example, in the novel “La parcela” (1898) by the Mexican author López Portillo y Rojas (Rössner 2007, 148, 188).For the defining characteristics and topics of the realist novel, see also Navarro (1955, 20–24).

Realism as a literary current prevailed in the second half of the nineteenth century. Besides the novels that are directly associated with this current, realistic elements are already present in earlier works. For instance, the novelas de costumbres also aspired to represent the local world and society in all its facets, but with a different motivation than the realist novels in the proper sense. Furthermore, especially Cuban novels are marked by a realistic orientation from early on and persistently in the nineteenth century because many authors were concerned with the social ills caused by the system of slavery. In terms of literary currents, this leads to works that mix romantic and realist(ic) elements, for instance, the novel “Francisco” (1839/1880) by Anselmo Suárez Romero: la idílica presentación de los desgraciados amores de Francisco y Dorotea contrasta con las escenas de la penosa vida de los esclavos en los barracones y los castigos inhumanos que les eran infligidos por parte de sus mayorales, descritas con gran crudeza. Esta coexistencia de elementos de ambas normas estéticas – la romántica y la realista –, que tan tempranamente se inicia, caracteriza buena parte de nuestra narrativa decimonónica y perdura hasta los inicios del presente siglo (Instituto de Literatura y Lingüística de la Academia de Ciencias de Cuba 1999, sec. Realismo)

The aesthetic closeness of the early Cuban novels with the later realist and naturalistic novels has led Molina to consider some nineteenth-century novels as instances of Naturalism before its time. Her study covers works written and published between 1830 and 1927 (Molina 2001).

The presence of realistic aspects in some of the novels of the first half of the nineteenth century is expected to complicate the textual separation of romantic and realist novels. On the other hand, the simultaneity of romantic, realist, and naturalistic works in Spanish America after 1880 and the existence of novels that included elements of several currents at once is another factor that could render the classification of the works by literary current more difficult.In literary-historical accounts of the nineteenth-century Spanish-American novel, some works are described as transitional between the romantic and realist currents, for example, “La Calandria” (1890, MX) by Rafael Delgado, or the historical novels of the Mexican writer Juan Antonio Mateos (Varela Jácome [1982] 2000, sec. 2.1.3; Gálvez 1990, 96, 105). Furthermore, literary historians come to different conclusions regarding the status of some works in relationship to literary currents. The novel “Cecilia Valdés o La Loma del Ángel” (1839/1882, CU) by Villaverde, for instance, is described as primarily romantic with realistic elements by Suárez-Murias and Varela Jácome, as realist by Dill, and as transitional between both currents by Gálvez (Dill 1999, 160–161; Gálvez 1990, 115–117; Suárez-Murias 1963, 36–40; Varela Jácome [1982] 2000, sec. 1.3). However, the assumption can also be made that there are prototypical cores of romantic, realist, and naturalistic novels in the sense of distinctive genres following aesthetic models that primarily have a European origin. In the case of the realist novel, elements that could lead to a coherent text type are specific topics such as life in the city, business and economic speculation, and the portrayal of different social classes. On a stylistic level, the hypothesis can be formulated that there are many and long descriptive passages and dialogues which have specific linguistic characteristics, for example, regarding the use of many different adjectives and nouns. These could contribute to the identification of a realist text type. A second hypothesis is that besides a realist text type in the narrow sense, another text type could become visible, in which realist and romantic elements are combined and in which the many novels that have conventionally been described as mixtures of romantic and realist(ic) elements take part.

Like the first Spanish-American romantic and realist novels, also the naturalistic ones started from a foreign model that was successively adapted to the particular needs of expression in Argentina, Cuba, and Mexico. In the case of the naturalistic novel, the model was primarily supplied by the French author Émile Zola, who propagated a scientific conception of man and a biological and medical way of thinking. In the naturalistic novel, realist techniques were carried forward to utilize the literary work as a means of experimentation and and for case studies of milieu, anthropology, genetics, or pathologies (Dill 1999, 168–169). Compared to the realist novel, the importance of the author’s creativity for producing an effect of reality is replaced by the aspiration for a pure documentary style. Nature shall be discovered “as is”. The development of the characters is a logical one, and inner reactions are subjected to mechanical processes. Reality is shown in its raw form, and no moral limits are set to the kinds of environment or human abysses that are studied. The goal is to stimulate the readers’ reflection by reproducing things as they are (Varela Jácome [1982] 2000, sec. 4). In terms of characters, the broad ensemble of the realist novel tends to be reduced, and the focus is on the protagonists whose psychological development is examined. The characters expose themselves through their way of speaking, either in dialogue or indirect speech, showing the regional or social determination of their means of expression. The influence of the milieu on the life of the characters is worked out (Dill 1999, 169).

In France, the naturalistic novel spread from the 1870s onwards, and in Spanish America it became popular in the 1880s. This shows that the adoption of this literary current happened with less delay than in the case of the romantic and realist novel. However, this led to the already mentioned superimposition of all three currents in the Spanish-American countries (Varela Jácome 2000, sec. 4). In Spanish America, naturalistic works were published in several different countries and in all of the three countries that are the focus here. The first and most naturalistic novels appeared in the Río de la Plata region, which includes Argentina.The great number of naturalistic novels that were written in Spanish America becomes evident in the comprehensive study that Schlickers published on the Spanish-American naturalistic novel. She includes 63 novels in her book and discusses almost every novel in its own chapter (Schlickers 2003). Authors whose works have been associated with the naturalistic current are, for example, the Argentine Eugenio Cambaceres, the Mexican Federico Gamboa, and the Cuban Martín Morua Delgado. For the authors, the naturalistic aesthetic provided an appropriate means to reflect the rapid economic upswing, as well as the technological and social processes of modernization that took place in the Spanish-American countries in the last decades of the nineteenth century, with all the challenges and difficulties that these brought for individual social groups or people (Schlickers 2003, 9–10). The range of topics covered in naturalistic novels includes aspects of biological and environmental determinism, immigration, mental ills, materialist neuroses, sexual rage, problems of alcoholism, the destructive force of prostitution milieus, and the representation of the urban Moloch. For example, the working conditions in factories and mines, but also the tedium of idle members of the upper class, are taken into account (Dill 1999, 169; Schlickers 2003).

Naturalistic novels share some features with realist novels because of the ambition to depict reality closely, and the different literary currents concurred temporally in the last decades of the nineteenth century. There are, however, some elements of naturalistic novels on the basis of which it can be hypothesized that they are distinguishable as textual genres. The crude style, usage of scientific or other specialized vocabularies, and the specific topics of this type of novel, such as the topics of immigration, mental and physical problems, and processes of social disintegration, are expected to facilitate their classification against realist and especially romantic novels. It is assumed that the novels that are conventionally labeled as naturalistic ones are also textually coherent to a relatively high degree. However, a difficulty might arise from the confusion and flexibility of the terms “realist” and “naturalistic” in the corresponding literary critical discourse and thus from the conventional class labels. If these terms are used with an overlapping or even identical scope, they are no longer a sign for different subgenres.On the difficulty to delimit the terms and concepts of costumbrismo, realismo, regionalismo, and naturalismo, see Navarro (1955, 12–19). Sánchez, for instance, only uses the term “novela naturalista” for works of both the realist and naturalistic types (Sánchez 1953, 257–259).

In a pretheoretical understanding, fictionality describes the property of a text (or other medium) to involve fiction, which means that it is about something imagined and invented. A novel, for example, is about events that did not actually take place, even if the author was inspired by the reality he or she knows and even if the author alludes to this reality in the text. Even so, theoretical considerations of fictionality show that it is not enough to assume that a text is fictional if it is about imaginary worlds (Klauk and Köppe 2014, 3).Nevertheless, earlier attempts to define fictionality that focused on reference semantics took this line. According to these views, fictional and factual texts are characterized by a respective specific mode of referentiality. Whereas a factual text could be defined as a text that references the empirical reality directly and is conceived and perceived as such by the author and its readers, a fictional text predominantly references invented places, characters, or events. Elements existing in the real world can also be part of a fictional text, but there should be elements in a fictional text that do not have any counterpart in reality. This means that these elements cannot be referred to anything existing before and outside of their linguistic formulation and creation in the text. However, it has been shown that the assumption of different modes of referentiality for fictional and factual texts is problematic, at least in the outlined narrow conception of the term (Weidacher 2017, 375–378). Recent approaches focus on pragmatic aspects to determine the fictionality of a text. According to the “institutional” theory of fictionality, for example, certain texts are considered fictional because of a coordinated and conventional social practice (an institution). A text is produced with the intention to be received according to the conventions of the fictionality institution. The sender and recipient of a fictional text enter into a contract establishing that questions of empirical referentiality and truth are not posed within the confines of the fictional text. The reader accepts the existence of the entities presupposed in the text and engages with them imaginatively if he or she recognizes the intention of the author to write a fictional text. For this, the authorial intention needs to be manifest in the text in some way, but ultimately, it is a pragmatic attribution to determine the fictional intention of a text (Köppe 2014, 35; Weidacher 2017, 378–381).

In accordance with this view, the fictionality of the texts to be included in the bibliography and the text corpus was assessed as follows. Statements of authors and readers regarding the fictionality of a text were taken into account. If it was indicated clearly that the text was conceived and received as fictional at the time and place of its publication, these signals were highly rated. In addition to explicit statements concerning the fictionality of the text, other paratextual and textual signals were evaluated. A comprehensive overview of potential signals of fictionality is given by Zipfel (Zipfel 2014, 97–119), who organizes them as follows: textual signals peritextual signals: place of publication, for example a specific collection or journal; book format; publisher; series; author; title and subtitle, possibly with genre labels; cover text; dedication; preface; etc. epitextual signals: publisher’s statements, reviews, interviews with the author; etc.

Of the various potential signals of fictionality, peritextual signals were especially useful to evaluate whether texts are to be considered fictional and if they should become part of the bibliography and the text corpus because they are very accessible.The peritext includes paratexts that are published as part of a work, e.g., prefaces and dedications, whereas the epitext involves paratexts that are published outside of the immediate context of the work (Genette 1987). Details such as author, title and subtitle, place of publication, publisher, and series are usually included in bibliographic descriptions of work editions and can, therefore, also be taken into account when the texts themselves are not available.For some of the works listed in bibliographies of the Spanish-American novel in the nineteenth century, no edition could be found neither in the WorldCat (a union catalog currently containing more than 430 million bibliographic entries from libraries worldwide, see OCLC 2023) nor in individual relevant library catalogs. Even if editions could be located, it was not in every case feasible to see them, especially if there were only print editions distributed over different American libraries. The issue of accessibility of the texts will be discussed in more detail in chapter 3.3.1 (“Selection of Novels and Sources”). A good indicator is a genre label in the title or subtitle of a work that refers to a fictional text type. Examples of such titles for Spanish-American narrative texts in the nineteenth century are: “novela”, “relato”, “narración”, “leyenda”, “romance”, “cuento”, or “drama”. There are other labels that are also common but less clear regarding the fictional status of the texts, for example: “historia”, “crónica”, “estudio”, “esbozo”, “cuadro”, “escenas”, “episodio”, “memorias”, “apuntamientos”, “anécdotas”. Sometimes labels refer to subgenres, such as “aventuras” or “costumbres”. To be able to decide whether a text is to be considered fictional or not in cases where labels are ambiguous, or where there are no explicit labels at all, other kinds of information were used. Where editions of a work were accessible, prefaces, introductions, and headings were consulted to see whether they clear up the issue of fictionality. Textual signals on the level of the story and on the level of the narration were also taken into account, but only in cases of doubt. A textual signal that is easy to recognize typographically and is typical for fictional narrative texts, though it is neither a necessary nor a sufficient criterion, is the reproduction of direct speech. Words or phrases that mark the end of a story or text can also be easily identified. Epitextual signals were not systematically researched. Especially for the bibliographical database, decisions were also based on information from existing bibliographies of fictional texts, literary histories, and other critical research literature.

In the case of Spanish-American novels, there are several factual text types that share characteristics with certain subtypes of the novel in terms of content or narrative mode. These are historiographic works versus historical novels, (auto)biographies versus (auto)biographical novels, travelogues versus travel novels, philosophical treatises versus philosophical novels, political treatises versus political novels, etc. That the boundaries between some kinds of fictional and factual texts are not always clear is influenced by several factors. Many of the authors in the nineteenth century who wrote novels were also authors of historiographic, political, journalistic, or philosophical works because there were still very few professional literary writers. Furthermore, many Spanish-American countries reached their political independence in the early nineteenth century, and there was a need to justify it and to contribute to the creation of a national identity not only through historiography but also by means of literary works (Kohut 2016, 171–172; Lindstrom 2004, 76–77; Sommer 1993). In his essay “Revistas literarias de México” from 1868, the Mexican author Ignacio Manuel Altamirano explains the ever more important role of the novel in this process: La novela es indudablemente la producción literaria que se ve con más gusto por el público, y cuya lectura se hace hoy más popular. Pudiérase decir que es el género de literatura más cultivado en el siglo XIX y el artificio con que los hombres pensadores de nuestra época han logrado hacer descender a las masas doctrinas y opiniones que de otro modo habría sido difícil que aceptasen. [...] la novela hoy ocupa un rango superior, y aunque revestida con las galas y atractivos de la fantasía, es necesario no confundirla con la leyenda antigua, es necesario apartar sus disfrazes y buscar en el fondo de ella el hecho histórico, el estudio moral, la doctrina política, el estudio social, la predicación de un partido o de una secta religiosa: en fin, una intención profundamente filosófica y trascendental en las sociedades modernas (Altamirano 1868, 17–18)

As long as they are either designated directly or indirectly as fictional in their paratexts or exhibit characteristics that are typical for fictional texts, these works were included in the bibliography and the corpus, even if they resemble factual texts because of their content or because of the way the narration is organized.

For example, the Mexican author Ireneo Paz wrote several historical novels that he labeled as such, but also a series of “leyendas históricas”. They are all centered on historical figures, as their titles suggest: “El Lic. Verdad”, “La Corregidora”, “Hidalgo”, “Morelos”, “Mina”, “Guerrero”, “Antonio Rojas”, “Manuel Lozada”, “Su Alteza Serenísima”, “Maximiliano”, “¡Juárez!”, “Porfirio Díaz”, and “Madero” (Pi-Suñer Llorens 2005, 386). They could also be interpreted as historical biographies, but because they are labeled as “legends” and contain direct speech, detailed descriptions of situations (e.g., weather conditions) and characters (e.g., behavior and appearance in specific situations), they are considered fictional texts here.

A work that is sometimes mentioned in critical works on the Spanish-American novel is “Vida de Juan Facundo Quiroga” (1845, AR) by Domingo Faustino Sarmiento.In the portal “Novela hispanoamericana del siglo XIX”, which is part of the “Biblioteca Virtual Miguel de Cervantes”, for example, “Facundo” is classified as “novela histórica argentina” but also as a biography of Juan Facundo Quiroga and as an autobiography of the author Sarmiento (Sarmiento [1845] 2000). In the first part of the work, the country, its inhabitants and their customs are described, followed by a biography of the Argentine caudillo Juan Facundo Quiroga. The last part contains considerations about Argentina’s political and economic future (Lichtblau 1959, 39–40). In a preface, the author refers to reactions by readers who missed certain details in the descriptions of historical events. Sarmiento defends himself by explaining how difficult the coordination of events that occurred in so many different places and at so many different points in time was challenging with the limited means he had (some reports of eyewitnesses, some simple manuscripts, some aspects recalled from his memory). He ends with the intention to improve his work in these aspects if time allows: Quizá haya un momento en que, desembarazado de las preocupaciones que han precipitado la redacción de esta obrita, vuelva a refundirla en un plan nuevo, desnudándola de toda digresión accidental, y apoyándola en numerosos documentos oficiales, a que sólo hago ahora una ligera referencia. (Sarmiento [1845] 2000, sec. Advertencia del autor)

In this authorial statement, there cannot be recognized any intention to write a fictional text. Moreover, the different parts of the work are not unified, and there are very few passages where direct speech is reported. “Vida de Juan Facundo Quiroga” is therefore considered a non-fictional text and excluded from the bibliography and the corpus.

Other borderline cases are descriptions of travels, for example, “La tierra natal” (1889, AR) by Juana Manuela Gorriti, “Mis montañas” (1893, AR) by Joaquín Víctor González, and “Una excursión a los indios ranqueles” (1870, AR) by Lucio Victorio Mansilla. All three texts also include autobiographical elements. For a factual travel narrative, three conceptual aspects are essential: the discourse is organized around a journey, for example, according to its itinerary or its chronology, the narrator can be identified as the author who recounts his or her experiences, which presupposes that there was an actual journey that took place before the narration, there is a general tendency towards description and objectivity. The travel narrative can furthermore be identified on the basis of paratextual signals (Alburquerque-García 2015, 19–29; Chávez and Urdapilleta 2015, 11). The discourse of a fictional travel narrative is equally organized around a journey, but the narrator can usually not be identified with the author, and no actual journey is needed as a basis for the narration. In addition, in a fictional travel narrative, narration tends to prevail over description and subjectivity over objectivity. It is likely that there will be paratextual signals pointing to the fictional status of the text. The second aspect (identification of the narrator as the author) is also relevant to distinguish an autobiography from an autobiographical novel. In the latter, the narrator and protagonist are not to be identified with the author.

When the three examples are examined, the following characteristics can be determined. In “La tierra natal”, the framing story is a railway trip from Buenos Aires to Salta. The text is structured into chapters that roughly correspond to stops of the journey. The traveler and first-person narrator gives an account of the journey and inserts conversations of fellow passengers, but also memories of her hometown. In a preface, Gorriti calls her work “páginas de lejanas memorias” (Gorriti [1889] 2001, 1). The end of the narration is marked with the word “Fin”.

In “Mis montañas”, the first-person narrator gives a report of a trip to the Sierra de Velazco in the Argentine province of La Rioja. The text is divided into 21 chapters which consist of landscape descriptions and impressions, historical background information and the imagination of historical events, the portrayal of local customs, the evocation of local characters and episodes, and personal memories. The work is prefaced by the Argentine writer Rafael Obligado, who gives several intertextual references. For example, he compares “Mis montañas” to the epic poem “La cautiva” by Esteban Echeverría. However, he does so not to stress its fictionality but the literary treatment of the Argentine landscape: “La propiedad artística de la cordillera argentina pertenece a Vd. de hoy para siempre, como la de la llanura al poeta de La Cautiva” (González [1905] 2001, X).Other similar references are made to the authors Alexander von Humboldt, Jacques Bernardin Henri de Saint-Pierre, William Wordsworth, Gregorio Gutiérrez González, François-René de Chateaubriand, Henry Wadsworth Longfellow, Domingo Faustino Sarmiento, and Olegario Víctor Andrade (González [1905] 2001, XII, XX).

“Una excursión a los indios ranqueles” begins with a letter written by the narrator, identified as “Lucio” and “coronel Mansilla”, just like the author, to his friend Santiago, in which he explains the circumstances of his expedition to the province of Córdoba where the indios ranqueles live. In 68 chapters, the narrator recounts his experiences in the form of letters to his friend. The work contains descriptive passages concerned with sociological, zoological, botanic, philological, and folkloristic facts, but also an intercalated novella and novelistic amatory and military scenes (García 1952, 132; cited by Lichtblau 1997, 609; Rössner 2007, 186–187).

In all three works, the discourse is organized around a journey that actually took place. All the texts are narrated in the first person, and the narrator can be identified with the author, either because of an explicit mention in the text (“Una excursión a los indios ranqueles”) or because of implicit formulations in the prefaces (“La tierra natal” and “Mis montañas”). In the paratexts, there is no clear evidence that the three travelogues were conceived or perceived as fictional. As to the third defining aspect of a factual travel narrative, the nature of the three works under consideration is less clear. All of them combine descriptive with narrative passages and objective representations with subjective perceptions to different degrees. One indicator for a narrative style, and hence a fictional text, that can be evaluated quantitatively is the amount of direct speech in the three texts. In figure 2, the travelogues are compared to other novels in the corpus regarding the proportion of paragraphs containing direct speech.The script that produced the box plot is available at https://github.com/cligs/scripts-nh/blob/master/corpus/direct-speech-travelogues.xsl and the result at https://github.com/cligs/data-nh/blob/master/corpus/direct-speech-travelogues.html. Accessed January 24, 2020. The group of novels that the travelogues were compared to is a subset of the whole corpus consisting of 92 novels in which direct speech has been marked up. As can be seen, the amount of direct speech in the three travelogues is less than in 75 % of the novels, so even if there are novels with an equal proportion of paragraphs containing direct speech, they do not represent the typical novel.

To conclude, even though the three travelogues resemble novels in some aspects (narrative,  subjective, and probably also fictional passages), they also share essential characteristics with factual travel narratives, and there are no indications that they were intended and read as fictional texts in their time. As a consequence, they were excluded from the bibliography and the corpus, even though they exhibit a certain generic ambiguity.The presence of narrative, subjective, and fictitious elements in travelogues has a long tradition in Spanish-American writings, going back to some chronicles of the Conquista (Anderson Imbert 1995, 17–48). This contributes to the literary character of these works but does not justify considering them plain fiction. The generic ambiguity of the three travel narratives discussed here is also evidenced by their inclusion or exclusion in other text collections and bibliographies. In the “Biblioteca Virtual Miguel de Cervantes”, for example, all three texts are part of the portal “Novela hispanoamericana del siglo XIX”. While “La tierra natal” and “Mis montañas” are also classified as novels in the general keyword system of the virtual library, “Una excursión a los indios ranqueles” is not. It is labeled with “Descripciones y viajes” (Gorriti [1889] 2001; González [1905] 2001; Mansilla [1870] 2001). In the bibliography of the Argentine novel authored by Lichtblau, “Una excursión a los indios ranqueles” is included as a borderline case, while “La tierra natal” and “Mis montañas” are not mentioned (Lichtblau 1997).

In contrast to the examples that were discussed in detail above, in the majority of cases, the fictional status of the texts that were candidates for the bibliography and the text corpus could be determined easily based on paratextual information, bibliographical and literary-historical sources. In the unclear cases, a reasoned decision was made, as exemplified above, whereby textual and paratextual information was preferred over critical discussions as far as possible.

According to Weber, narration is “[1] adressierte, [2] serielle, [3] entfaltete berichtende Rede [4] mit zwei Orientierungszentren [5] über nicht-aktuelle (meist: vergangene), [2] zeitlich bestimmte Sachverhalte (besonders: Ereignisse in zeitlicher Folge) [6] von seiten eines Außenstehenden” (Weber 1998, 63; cited by Zymner 2017, 365).The following clarifications of the definition of narration are a summary of Weber‘s more detailed explanations. The various elements of this definition will be briefly explained here. While Weber’s definition also holds for oral narration, it will only be applied to written narration in this context. That narration is addressed means that there is someone (a narrator) narrating and addressing someone else (a reader). It is the serial exposition of chronologically determined circumstances, facts, or events, which means that something is told in a specific order, which does not have to correspond to the underlying chronological order of the events. This may typically be an exposition, followed by the complication of events and a subsequent clarifying conclusion, but can, of course, also take other forms. As to the underlying circumstances or events, a narrow definition of narration presupposes that there are at least two propositions that involve development or shift. Narration can then be understood as the representation of a situational change. A broader definition would also include a series of discrete propositions, which do not necessarily have to be connected in the form of succession. “Entfaltete berichtende Rede” refers to the relationship between a report and a narration. The latter can be conceived as a detailed, stylistically evolved report. A narration is centered on two points of orientation because, on the one hand, there are the narrated circumstances, facts, or events that did already take place and the people involved in them (the first system of orientation). On the other hand, there is the moment of reporting the events including the presence of the narrator (the second system of orientation). Usually, what is narrated is past, at least from the point of view of the narrator. Because it is also possible that something imagined is narrated, either imagined as past or as possible in the future, it is more appropriate to say that what is narrated is not present. A narration is presented by someone external and distanced who can report the events without the necessity to stick to their immediacy, succession, chronology, or unity (Weber 1998, 11–63).

Although this definition was very useful for the decision to include texts into or exclude them from the corpus, provided that they were, in principle, available, its usefulness for the selection of entries for the bibliography was limited in the same way as for fictionality. Where editions of the texts could not be accessed, it was necessary to rely only on available metadata and on third-party information. In terms of metadata, mentions of narrative genres in book titles and subtitles or in titles of book series were especially helpful. Regarding third-party information, it had to be taken into account how narrativity was defined in each context (if it was defined at all). For example, Lichtblau discusses the selection criteria for his bibliography as follows: The problem of identifying those works that clearly belong in the classification ‘novela argentina’ beset me at every stage in the preparation of this bibliography. But I have attempted, within a certain arbitraryness inherent in all literary categorization, to be consistent in the selection or omission of the works cited. [...] In addition, I have included a few celebrated works of Argentina literature that, although not novels, retain many of the characteristics of that genre and are associated with its development and artistic expression. We may thus say that Echeverría’s El matadero, Cané’s Juvenilia, and Mansilla’s Una excursión a los indios ranqueles have been recruited for this bibliography without having the proper credentials as ‘novel’. I did leave out, however, Sarmiento’s Facundo, not wishing to stretch the point too much. (Lichtblau 1997, XV–XVI)

He does not provide an explicit definition of the novel and does not refer to the concept of narrativity. His criteria could only be inferred from the examples that he mentions.He is explicit in two aspects, though: What he considers an Argentine novel and how he distinguishes novels and short novels from short stories (Lichtblau 1997, XV–XVI). Therefore, wherever full texts were available, the information obtained from other bibliographies was checked before a work was included in the current bibliography. An example of a text that is included in Lichtblau’s bibliography (Lichtblau 1997, 309), but excluded here, is “La flor de las tumbas” (1866, AR), written by Santiago Estrada because the text has the form of a dramatic text instead of a narrative text. It starts with a cast list, is divided into acts and scenes, contains stage directions, and consists entirely of character speech. This does not fulfill the criteria established by Weber, especially that a narration should be addressed, reported by someone external, not be immediate, and have two centers of orientation. In the preface, the author explains how he conceived his work generically:

Este trabajo no es un drama en la acepción literaria de la palabra. Moriría en el teatro, para el cual no está dedicado. El artista puede revestir sus concepciones en la forma que mejor se avenga a su expresión espontánea.—Este trabajo es un romance. Dibujar los cuadros o pintarlos, eso queda al arbitrio del artista. ¿Quién me obligaría a prestarle el empaste de la narración?

¿Puedo esperar que una lágrima escapada del alma del lector, le de el colorido que yo le niego, dejándolo en la simplicidad elemental de sus líneas?... No lo sé.—Escribo para sentir, y nada más.

Su forma no carece de precedentes. Sin traer a recuerdo magistrales producciones literarias, que tomando la división y sencillez del drama, no han aspirado a la exhibición viva de la escena, citaré solamente los conocidos romances que un poeta francés ha llamado: comedias de sillón,—y las que el marqués de Varennes ha denominado: proverbios.

Esto por lo que respecta a la forma.

Estrada thus says that his work is not a drama because it is not intended to be presented on stage. Instead, he calls it “romance”. However, he also clearly says that it does not have the form of a narration. It is kept “simple” and “rudimentary”, without coloring, drawn, but not painted, which a narration in the sense of a detailed, stylistically evolved report would be.

In general, however, it was easier to determine the narrativity of the texts eligible for the bibliography and the text corpus than their fictionality. As to the borderline cases for fictionality, the historical biographies and the travelogues are, for the most part, narrative. Only Sarmiento’s “Vida de Juan Facundo Quiroga” is not predominantly narrative, but it would still have to be discussed how much narrativity a text needs in order to be interpreted as a narration. As Weber states, when he elaborates his definition further, normally, a narration does not consist entirely of narrative text. It can also contain other forms of presentation, for example, the report of direct speech, descriptions, argumentative passages, or comments (Weber 1998, 64–70). An example of a text containing scenic presentation is the historical novel “La loca de la guardia” (1896, AR), written by Vicente Fidel López. In chapter 40, the conversation between a judge and an accused person in a trial has the form of dramatic speech. Nevertheless, this passage amounts only to about 5,300 words, and the entire novel has a length of approximately 97,500 words, so it can still be considered a narrative text.

“Prose” can be defined as a form of text that is metrically not bound, as opposed to text in verse form (see, for instance, Kleinschmidt 2003, 168). This criterion concerns primarily the distinction between narrative prose and poetry. Many of the Spanish-American novels in the nineteenth century contain inserted poems. They may be quotations at the beginning of individual chapters or part of the narration, for example, if they are recited in public by a character or are part of a love letter that is represented in the text. In general, these insertions only make up a small part of the entire text and do not question that a work is written predominantly in prose. As for the selection of texts for the bibliography, caution is required when works carry the generic label “romance” or “leyenda” because they can either be novels written in prose (for example, “El romance de un médico” (1905, AR) by Cupertino del Campo and “Un santuario en el desierto. Leyenda original” (1890, MX) by José Francisco Sotomayor) or epic texts written in verse (e.g., “Perfiles de la conquista. Romance histórico. 1521–1887” (1887, MX) by Juan Antonio Mateos and “Un ángel desterrado del cielo. Leyenda religiosa” (1855, MX) by Niceto de Zamacois). The latter were excluded from both the bibliography and the corpus.Sometimes novels have also been versified by other authors, for instance, the novels of the Argentine Eduardo Gutiérrez, which have been reworked by Bartolomé R. Aprile, Silverio Manco, and Apolinario Sierra. See, for example, Gutiérrez and Aprile ([1944] 2015), Gutiérrez and Manco ([1948] 2015), and Gutiérrez and Sierra ([1944] 2015). There are also many texts without generic labels, which can be of any genre (novels, collections of short stories or poems, plays, other types of literary or non-literary texts) and be written in prose or verse. In these cases, the recourse to existing bibliographies of the novel and to library catalogs that include information about the genre was indispensable to finding the relevant texts.

The length of the text is one of the criteria that serve to distinguish the novel from other forms of fictional narration in prose, especially shorter ones such as the novella and the short story. However, usually, these genres are also differentiated according to other criteria because there may be exceptions, for example, very short novels and very long novellas, so that a novella might be longer than a novel in individual cases. Moreover, there is no consensus on the exact or approximate lower boundary of the length of a novel. Traditionally, the length of a fictional narration is expressed in page numbers which can only be a rough indicator because of differences in book format, layout, and typography from one edition to another.Especially editions with a layout in two columns lead to more words per page than single-column layouts. Some of Eduardo Gutiérrez’ novels were published with a two-column layout. See, for instance, Gutiérrez ([1893] 2016); Gutiérrez ([1880] 2016a); Gutiérrez ([1880] 2016b). It is more precise to measure the length of a text independently of the design of a print edition, for example, in the number of words or characters, but this is only feasible for texts which are available in electronic form and machine-readable.

In “Aspects of the novel”, a collection of literary lectures about the English language novel held in 1927, Forster claims: “Any ficticious prose work over 50,000 words will be a novel for the purposes of these lectures” (Forster 1927, 17), but without motivating the number. In the context of a German handbook on literary genres, Fludernik mentions the following page limits: She sets an upper limit of 40 to 50 pages for the short story and the novella and a lower limit of 80 pages for the novel, leaving a corridor of about 30 pages for unclear cases (Fludernik 2009, 632). Unfortunately, she also does not explain how she arrives at these numbers. A more detailed discussion about the extension of the short story, novella, and novel can be found in “La novela corta mexicana en el siglo XIX” by Mata, who is looking for pragmatic criteria allowing him to define the scope of his object of study. He points out that every proposal of an exact number can, at best, apply to a specific historical context but not to the novel in general. As to Forster’s suggestion, Mata states that the number of 50,000 words seems appropriate for the typical, extensive novels of the nineteenth century but not for many of the paradigmatic novels of the twentieth century, which are shorter (Mata 1999, 16). It should be added that also the geographical and the cultural context determine the characteristics of a historical genre. In the nineteenth century, the novel had a longer tradition in Europe than in Spanish America and was more stabilized as a genre (Fludernik 2009, 638–645),Fludernik traces the history of the novel from early modern precursors up to the twentieth century and states that the novel is a European genre which spread internationally in particular in the nineteenth and twentieth centuries. Nevertheless, the novel as a genre was not unknown in the Spanish-American colonies in earlier centuries. The circulation and reception of European novels in the colonies and the existence of precursors of the Spanish-American novel are set out, for example, in Sánchez (1953, 67–127). See also Lindstrom (2004, 47–77). so it can be assumed that more works complied with the established model of the time. The range of the texts considered novels in the nineteenth century in Spanish America was broad. In the early century, many of the novelistic narrative texts in prose were quite short,For example, the short novels written by José Joaquín Pesado, Ignacio Rodríguez Galván, Ramón de Palma y Romay, Félix Tanco y Bosmeniel, and Juana Manuela Gorriti. while European models – extensive historical, realist, and naturalistic novels – gained more ground towards the middle and end of the century.For instance, the historical novels of Ireneo Paz and Juan Antonio Mateos, the realist novels of Carlos María Ocantos, and the naturalistic novels of Eugenio Cambaceres and Federico Gamboa. Towards the turn of the century and in the twentieth century, many novels were shorter again, in correspondence, interrelation, confrontation, and also independence from European developments.Like the modernist novels written by Amado Nervo and Efrén Rebolledo, for example. Using the limit set by Forster, many texts that can be assigned to the genre novela would be excluded from analysis. The strategy followed by Mata is to consult calls for literary competitions to see which limits they pose for the length of texts belonging to different narrative genres. On that basis, he arrives at the following numbers: a maximum of 5,000 words for short stories, a minimum of 5,000 words and a maximum of 35,000 words for short novels, and more than 35,000 words for novels (Mata 1999, 16–17). Despite his remark on the historicity of genre lengths, Mata relies on modern literary competitions in order to establish the length of novellas or short novels in the nineteenth century, which he analyses. It can only be speculated why he did not use information about literary competitions in the nineteenth century – maybe because of the scarcity of sources?

An important question is whether it would be more appropriate to distinguish the novel from other, shorter forms of narrative prose not on the basis of text length but using structural and content-related criteria. Usually, the novel is described as a complex form of narration, while the shorter text types are characterized as simpler, single-stranded forms. According to general definitions, the novella, for example, is said to present an exemplary story with one central event, with a closed structure and only a minor elaboration of the characters’ life. The short story is characterized by a relative unity of place, time, and plot. The latter is usually limited to the representation of single events and has an abrupt ending. The characters tend to be typified. In the novel, in contrast, several parallel storylines and subplots, changes of place and time, and fully elaborated characterizations are more common. These structural and content-related aspects are, of course, also induced by the extent of the form (Fludernik 2009, 632; Strube 1993, 21; Zymner 2017, 371–380). Ultimately, the complex interplay of the different factors would have to be taken into account to determine to which genre a narrative prose text belongs because none of the criteria is in itself sufficient. The use of general generic definitions is problematic, though, because they do not take into account the cultural and historical context.

It is questionable whether the novella, for example, was a common genre in literary production in Spanish America in the nineteenth century at all, and even if it was, it is doubtful whether the above-mentioned characteristics would have applied. While novels and short stories can often be distinguished based on the works’ subtitles (“novela” versus “cuento”)There are exceptions here, too. “La loca de la guardia” (1896, AR) by Vicente Fidel López has the subtitle “Cuento histórico” and is a work published as a book with almost 500 pages. There are also some texts of intermediate length which are called “cuento”, for example, “María, la perla de la Diaria. Cuento cubano” (1866, CU) by Rafael Otero, published independently with 119 pages, and “El hogar en la pampa (Cuento)” (1866, AR) by Santiago Estrada with 133 pages. They can all be considered novels., there is no distinctive term for short novels in Spanish. They are often called “novela”, as well, and sometimes “novelita” or “novela corta” (Mata 1999, 32–33).Mata mentions a whole range of denominations for the short novel in nineteenth-century Mexico. Many short novels were produced in Argentina, Mexico, and Cuba in the nineteenth century. Some were published independently in book formFor example, “María del Consuelo. Novela” (1894, MX) by Alberto Leduc with 39 pages, “Comunidad de nombres y apellidos. Novela original” (1845, CU), and “Teresa. Novela original” (1839, CU) by Cirilo Villaverde with 63 and 93 pages, or “Un ángel y un demonio, o el valor de un juramento (Novela original)” (1857, AR) by Margarita Ochagavia with 104 pages., some as part of collections of several shorter narrative textsThe “Panoramas de la vida. Colección de novelas, fantasías, leyendas y descripciones americanas” (1876, AR) by Juana Manuela Gorriti, “Tardes nubladas. Colección de novelas” (1871, MX) by Manuel Payno, and “Mesa revuelta. Colección de artículos de amena literatura, opúsculos, juicios críticos, historietas, novelas, folletines, revistas viejas y otras muchas cosas” (1860, CU) by Francisco Calcagno, for instance. and the majority in journals (Mata 1999, 29; Molina 2011, 58–59). In his account of the nineteenth-century short novel in Mexico, Mata states that short novels were among the first kind of narrative texts which were published a lot in journals shortly after the country’s independence. He characterizes them as generally not having much literary value and not having been designated with the term “novela corta”, which was practically unknown in the early nineteenth century. Many of the terms that were used in the titles of the texts point to the preliminary character of the works: “pequeña novela”, “esbozo de novela”, “proyecto de novela”, “esquema de novela”, “tentativa de novela”, “ensayo de novela”, “apuntes para una novela”, etc. (Mata 1999, 32–33). Mata relates these titles, as well as the fact that many shorter novels were simply called “novela”, to the problem of the missing term for the intermediate narrative genre, which on the other hand, already existed in other languages. According to him, the term “novela corta” only became common in the Iberian Peninsula and Mexico towards the end of the nineteenth and the beginning of the twentieth century, an observation which can be confirmed by analyzing the works consulted for the bibliographic database (Mata 1999, 33).Assessing materials for the bibliographic database, the following works and collections with the subtitle “novela(s) corta(s)” were found: “La manigua sentimental. Novela corta” (1910, CU) by Jesús Castellanos, “Otras vidas. Novelas cortas” (1909, MX) by Amado Nervo, “Gil Luna, artista. Novelas cortas” (1908, CU) by Luis Rodríguez Émbil, “El enemigo. Novela corta” (1908, MX) by Efrén Rebolledo, “Thespis (Novelas cortas y cuentos)” (1907, AR) by Carlos Octavio Bunge, “Voces perdidas (Novelas cortas y cuentos)” (c. 1907, AR) by Jorge Lavalle Cobo, “Sucesos y novelas cortas” (1903, MX) by José López-Portillo y Rojas, “Novelas cortas de varios autores” (1901, MX) – a compilation of earlier short novels – and “La capilla de los álamos. Colección de novelas cortas” (1892, MX) by Manuel Covarrubias y Acevedo. Furthermore, there were volumes of collected works entitled “novelas cortas”: “Obras del Sr. D. J. María Roa Bárcena. Novelas cortas” (1910, MX), “Obras de Don Florencio M. del Castillo. Novelas cortas” (1902, MX), “Obras de Don Manuel Payno. Novelas cortas” (1901, MX), “Obras del Lic. D. J. López-Portillo y Rojas. Novelas cortas“ (1900, MX). They were all published in the late nineteenth and early twentieth century. An exception is the earlier collection “Horas de tristeza. Novelas cortas” (1849, MX) by Florencio M. del Castillo. Towards the end of the century, short novels gained prestige, especially in the context of the Modernismo current (Mata 1999, 143). Mata argues that all these texts of intermediate length should be treated as “novelas cortas”, understood as a genre between the short story and the novel, which existed from the early nineteenth century on but has been neglected by literary critics and historians (Mata 1999, 139). When defining this short novel in the first chapter of his book, he refers to Walter Pabst’s study “Novellentheorie und Novellendichtung”, an account of the origins of the European novella in Romance languages (Mata 1999, 11–12). From a taxonomic perspective, this may make sense, as all of these narrative texts are of intermediate length, but if genre is understood as an historico-cultural phenomenon, it would have to be analyzed if there is a direct relation between the early “novelitas” and the European novellas at all. Mata’s argument that the early short novels were the protagonist of the initial period of the Mexican (national) narrative (Mata 1999, 141) – in their capacity as first attempts towards the genre “novela”, fostered and popularized by the press  – seems more likely. Nevertheless, it would have to be examined in detail to what extent authors, readers, editors, and critics of the time understood the early short novels as representatives of the genre novella. For the later short novels, this link would equally have to be discussed, although there is certainly more awareness for the “novela corta” because the term is used more often. Even so, novels, in general, tended to be shorter again, making it difficult to differentiate between “novela” and “novela corta”.There are other approaches to the Mexican short novel besides Mata‘s. In particular, the portal “La novela corta. Una biblioteca virtual” has been developed by a research project hosted at the Universidad Nacional Autónoma de México (2008–2023). The portal is accompanied by critical approaches to the short novel published in five volumes, among them Chaves (2011). Like Mata, Chaves (115–119) links the history of the short novel in Mexico to European traditions (German, French, and English). In a compilation of Mexican romantic short novels, Ruedas de la Serna concludes: “Sin embargo su interés radica precisamente en que fueron los primeros ensayos narrativos de nuestros escritores en que surge una clara conciencia de la expresión literaria. Cierto que estos avanzaban penosamente en el dominio de esta nueva técnica de representación de la realidad, de la que, como de tantas otras cosas, se nos había privado. Cuánto, sin embargo, no habrían contribuido estas obritas en la batalla de nuestros intelectuales del siglo pasado por transformar su sociedad, y cuánto no deben a estas primicias los novelistas posteriores” (Cárabes and Ruedas de la Serna 1998, 71–72), thus evaluating the early short novels as first narrative attempts, the view preferred here. For Argentina and Cuba, no comprehensive studies of the short novel in the nineteenth century could be found.

To conclude, the short novel is not easily recognizable as an independent genre with a certain coherence in Argentina, Cuba, and Mexico in the nineteenth century. Furthermore, there are reasons to consider many of the shorter novels as novels, as well.In accounts of the nineteenth-century Spanish-American novel, shorter novels are often included, especially the early ones published in the 1830‘s and onwards, which are mentioned as first novels of their kind, for example, “Netzula” (1837, MX) by José María Lacunza as the first indianist novel (Brushwood 1966, 71; Sánchez 1953, 546; Varela Jácome [1982] 2000, 4). Suárez-Murias mentions various “novelitas” when tracing the development of the Cuban romantic novel (Suárez-Murias 1963, 22–27). Molina includes several short novels in her book about the early Argentine nineteenth-century novel (Molina 2011, 405–489). Therefore, in this dissertation, neither the lower limits for the novel set by Forster (50,000 words) nor by Mata (35,000 words) are used. Instead, an own limit of words was deduced from bibliographic descriptions of novels, taking into account the extent of the texts in conjunction with historical subgenre labels in order to approximate the minimum and the typical length of a novel for contemporary authors and editors. Of course, not all the novels were labeled as such, but the majority were, which makes it possible to arrive at a better understanding of the extent of the texts belonging to the genre in their time. The term “novela” is understood as designating novels, not novellas, despite exceptional cases where it is clearly used in the latter sense.For example, in the title “Fru Jenny. Seis novelas danesas” (1915, AR) by Carlos María Ocantos, a cycle of six novellas that share the same geographic setting and are published together in one volume. This work is out of the scope of this dissertation anyway because of its publication after 1910. Works with the subtitle “novela corta” or “novelita” were excluded from the calculation.

In principle, it would have been possible to also use structural and content-related criteria to select texts for the corpus, but this would not have been very efficient because an application of these criteria would have presupposed either access to detailed summaries of the texts or a close-reading of all the texts. To be able to decide upon the inclusion of texts into the bibliography, again, either detailed summaries or the full texts of all eligible works would have had to be accessible, which was not the case. Furthermore, the use of structural and content-related criteria would have presupposed established definitions of the various narrative genres, which, especially for the Spanish-American short novel, are not available. The extent of the text, in contrast, is usually part of bibliographic descriptions of the works and is a piece of information that is easy to access. It is therefore used as a proxy here to distinguish between novels and other shorter types of narrative prose texts.

The unit chosen here to measure the extent of the texts is the number of words. For each eligible text that is accessible in a full-text format of good quality,The quality of texts that have been digitized using OCR without being corrected afterward is usually not sufficient. See, for example, the full-text versions of texts uploaded to the Internet Archive (e.g., Ramírez [1868] 2008). this number was accessed with a simple regular expression counting all the tokens separated by non-word characters (such as white space or punctuation marks).tokens = re.split(r"\W+", text, flags=re.MULTILINE) With this approach, complex linguistic structures like compounds or words with clitics are not assessed, but this is acceptable because the focus is on the comparability of text length and not on the linguistic characteristics of the texts. For the entries in the bibliography, the number of pages was used and converted to an estimated number of words. One hundred pages were selected randomly from 50 different nineteenth-century Spanish-American novels to identify an average number of words per page and to balance out differences in layout, typesetting, and font. The words on these pages were then counted.If a page turned out to be a blank page or a page containing an index or only an image, it was replaced by the next (or preceding) regular text page because these are considered exceptional page types. On the other hand, chapter beginnings and endings with fewer words than full text pages were kept because they appear regularly in the novels to a certain extent. Parts of words occurring at the beginning or the end of a page were counted as whole words. Figure 3 shows the distribution of the number of words per page for the random sample.The script used to generate the list of random pages and the box plot are available at https://github.com/cligs/scripts-nh/blob/master/corpus/words-per-page.py. Lists of the random pages and the corresponding editions of novels can be accessed at https://github.com/cligs/data-nh/blob/master/corpus/random-pages.csv and https://github.com/cligs/data-nh/blob/master/corpus/pages-novels.xml, respectively. The full texts of the selected pages are collected in the file https://github.com/cligs/data-nh/blob/master/corpus/pages-text.xml, and the resulting box plot file can be found at https://github.com/cligs/data-nh/blob/master/corpus/words-per-page.html. Accessed January 25, 2020. The number of words per page ranges from 50 to 475, with a median of 191 words. In the following, this median is used to estimate the number of words of a text with a known number of pages.

To examine the range of lengths of nineteenth-century Spanish-American novels, 129 full texts and 252 bibliographic entries of works carrying the label “novela” either directly in the title or subtitle or in the title or subtitle of a series to which the work belongs were analyzed.The data that was used here were preliminary corpus and bibliography, which were successively refined. The works chosen were written by Argentine, Cuban, and Mexican authors (see chapter 3.1.2 for an explanation of how the country to which an author belongs was determined) and published between 1830 and 1910 (see chapter 3.1.3 explaining the chronological limits used here). See chapters 3.2.1 and 3.3.1 for details about the sources of the full-text corpus and the bibliographic database, respectively. All bibliographic references of novels missing page numbers were left out. Furthermore, works contained in the following collections were not considered for the calculation of the typical length of a novel: “Mesa revuelta. Colección de artículos de amena literatura, opúsculos, juicios críticos, historietas, novelas, folletines, revistas viejas y otras muchas cosas” (1860, CU) by Francisco Calcagno, “Leyendas, novelas y artículos literarios” (1877, CU) by Gertrudis Gómez de Avellaneda, and “Panoramas de la vida. Colección de novelas, fantasías, leyendas y descripciones” (1876, AR) by Juana Manuela Gorriti because it is unclear which of the genres mentioned in the titles apply to which texts; “Horas de tristeza. Colección de novelas” (1850, MX) by Florencio María del Castillo and “Tardes nubladas. Colección de novelas” (1871, MX) by Manuel Payno because they were published in other editions with the title “Novelas cortas”; and “Novelas en germen” (1900, MX) by Emilio Bobadilla because the texts are all shorter than 50 pages and the title “Novelas en germen” can be interpreted as designating short novels. In the case of the full texts, the words were counted. For the bibliographic entries, the number of pages was converted to a number of words using the median number of words per page.Obviously every work can have multiple editions. In the case of several different editions, the mean of their respective number of pages was used to balance out differences regarding the number of words per page. A work with several editions was considered eligible as a novel if at least one of the various editions published between 1830 and 1910 carried the label “novela”. For the full texts, only one available edition was used to count the words. Another factor of uncertainty when using page numbers of bibliographic entries is that they usually refer to the pagination of the books and not to the number of pages of the work, so prefaces, indexes, appendices, etc. might be included, which means that the works themselves are possibly shorter than calculated here. The results for the full texts, the bibliographic entries, and both combined are displayed in figures 4, 5, and 6, respectively.The script that was used to select the works from the preliminary corpus and bibliography, to calculate the numbers of pages and words, and to create the box plots for figures 4, 5, 6, and 7 is available at https://github.com/cligs/scripts-nh/blob/master/corpus/words-novelas.xsl. The corresponding data and results can be viewed at https://github.com/cligs/data-nh/tree/master/corpus/words-novelas. A list of all the works that were used for the calculation of the word (and page) limit is given at https://github.com/cligs/data-nh/blob/master/corpus/words-novelas/novelas-length.csv. Accessed January 27, 2020. All the distributions have a pyramidal form which means that they are right-skewed: the higher the number of words, the fewer works carrying the label “novela” there are, or, in other words, most of the “novelas” are rather short.A right-skewed distribution is one with many low and few high values and with a mean that is higher than the median. Looking at the numbers, the shortest novel in figure 5 has 3,438 words, and the longest one 334,441, which is almost a hundred times as long, so the spectrum of lengths is very large. The median is at 44,000 words, the first quartile at 25,000 words, and the third quartile at 73,000 words.The medians and quartiles are rounded to the nearest thousand. For the full texts alone, the median is at 61,000 words, and for the bibliographic entries alone, at 36,000 words. That the numbers are higher for the full texts is probably due to the selection of the texts: many of the digitized nineteenth-century novels available in full-text format are long novels, which tend to be considered paradigmatic. Furthermore, many historical novels were chosen for the corpus, and these have a tendency to be longer than novels of other subgenres. In addition, very short novels were avoided in the collection of the full texts, but in the bibliographic entries, they were included. With a lower limit of 50,000 words as proposed by Forster, more than half of the “novelas” would be left out, and with Mata’s limit of 35,000, still more than one-fourth of them would be considered short novels.

Based on these results, the question remained where to make a cut-off. It did not seem reasonable to include all the texts with the same length as the shortest “novelas”, as these are only about 20 pages long, so they clearly overlap with novellas and longer short stories.For example, the novels “Dos niñas hechiceras (novela original)” (1874, AR), published independently under the pseudonym “Guindilla”, and “Una sanjuanina, o sea Carolina. Novela de costumbres” (1881, MX) by Guillermo Quiroga, which are both only 18 pages long. In these cases, a recourse to structural and content-related criteria would have been indispensable to be able to differentiate between the genres. It was helpful to look at the length of texts explicitly labeled as “novela corta” to define a lower word limit. Figure 7 shows the distribution of word lengths of 65 “novelas cortas”.The word length for the “novelas cortas” was determined in the same way as for the “novelas”. Three works were available as full texts so that their words were counted with a regular expression. For the other 62 works, the number of pages was converted to a number of words using the mean number of words per page calculated above. Again, the shorter texts dominate, with a few outliers of greater length. The median for the short novels is around 7,300 words, the first quartile at 4,900, the third quartile at 10,400, and the upper fence at 16,800 words.The values were rounded to the nearest hundred. The upper fence defines a limit between values that can still be considered typical in a distribution and those that can be considered outliers. It is set at the largest sample that is larger than the third quartile (Q3) but still lower than Q3 + 1.5 * IQR (the interquartile range, which is Q3 - Q1).

Cutting off the “novelas” at the first decile – meaning that the shortest 10 % are left out – leads to a value of 16,000 words as a minimumRounded from 16,044., which is very close to the upper fence of the “novelas cortas”. That way, exceptionally long “novelas cortas” are included, while “novelas” of the same length as typical “novelas cortas” are excluded. Reformulated in page numbers, the limit amounts to 84 pages.Rounded from 83.8. Interestingly, this limit is very close to the 80 pages assumed by Fludernik (2009, 632). In this dissertation, the word limit was used to select texts for the corpus, and the page limit for the selection of entries for the bibliographic database.The number of words in the full texts was rounded to the next thousand before the limit was applied. To be independent of the naming conventions again, all fictional narrative texts in prose of this length were included.

Of course, this cut-off is still arbitrary to a certain extent – why should a “novela” with 15,000 words or 79 pages be excluded, but a “novela corta” with 16,000 words or 84 pages be included? It is nevertheless a limit deduced on the basis of empirical data from the same cultural-historical context as the works to be analyzed, which makes it probable that it approximates the generic conventions of the time. Furthermore, no clear cut could be seen in the data, the transition from very short to longer novels being rather fluent so that every other limit would have led to a similar arbitrary split. In addition, a numeric criterion is directly usable in a quantitative study without the need for extensive close reading.

In some definitions of the novel, an independent publication as one or sometimes several books is mentioned as one of the characteristic traits of the texts belonging to the genre (Fludernik 2009, 627; Steinecke 2007, 317). However, an independent publication will not be required here in order to select texts for the bibliography and the corpus for several reasons. First, the publication of a work as one or several independent books depends to a certain extent on the length of the text. As discussed in the previous subchapter, many of the nineteenth-century Argentine, Cuban, and Mexican novels were quite short and were sometimes published in a volume together with other works, especially when the authors wrote a whole series of novels, for example, the “Entretenimientos literarios” (1843–1844, CU) by Virginia Felicia Auber de Noya or the “Episodios nacionales mexicanos” (1902–1903, MX) by Victoriano Salado Álvarez. Shorter novels were also published in collections of works of various narrative genres, such as the “Panoramas de la vida” (1876, AR) by Juana Manuela Gorriti. Second, the publication in book form corresponds to a particular model of distribution for literary works, which was not the only one in nineteenth-century Spanish America. A large part of the novels was published in journals and literary magazines, many of them in serial form.Two forms of serial publication were common: the novela por entregas, where parts of the novel were delivered loosely, as a booklet accompanying a newspaper, or included in a literary magazine, and the novela de folletín, where the novel was published subsequently in specific columns of a daily newspaper (Molina 2011, 27; Villegas Cedillo 1984, 12–15). Not all of these novels were also published in book form afterward. Whether a contemporary or modern monographic publication exists also depends on the degree of canonization of a work. As the present study aims to include as many novels as possible so as to broaden the empirical basis for the description and analysis of subgenres of nineteenth-century Spanish-American novels, no restrictions are made regarding the form of publication of a work.In spite of this, a bias towards the selection of the more canonized works can hardly be avoided because they are the ones that are better transmitted and more often digitized, especially as full texts. Moreover, bibliographies of the novel refer mainly to monographic publications. See chapters 3.2.1 and 3.3.1 on the selection of texts for the bibliography and the corpus for details.

However, an independent publication in book form is also not just a practical matter related to text length and modes of distribution. Although the question of a novel’s unity and delimitation is not easily answered by requiring it to be published independently, this still emphasizes its autonomy as a work of art. As discussed in the section on length above, very short novels published in book form existed. On the other hand, there are also novelistic works which are so long that they do not fit into one physical volume. These are often published in several books called “tomos”, for example, the first book editions of “El fistol del diablo” (1859–1860, MX) by Manuel Payno with four or “Amalia” (1855, AR) by José Mármol with eight volumes. In the case of sequels and cycles published as several books, it is less obvious if each part should be considered its own novel or if they form one novel altogether. Often, the connection between the texts is indicated in titles and subtitles, as the following examples illustrate: ”Libro extraño”, “Libro extraño. Genaro. Tomo II”, “Libro extraño. Don Manuel de Paloche. Tomo III”, “Libro extraño. Méndez. Tomo IV”, “Libro extraño. Hacia la justicia. Tomo V” (1894, 1895, 1899, 1897, 1902, AR) by Francisco SicardiIn this case, it is strange that the third part has a later publication date than the fourth part. There must be an earlier edition of “Don Manuel de Paloche”, but this could not be verified. “Dramas militares. El Chacho”, “Dramas militares. Los montoneros. Continuación del Chacho”, “Dramas militares. El rastreador (Continuación de Los montoneros)”, “Dramas militares. La muerte de un héroe. Continuación y fin de El Chacho, Los montoneros y El rastreador” (all 1886, AR) by Eduardo GutiérrezGutiérrez was a very productive writer who wrote 34 novels, many of them organized in cycles. Besides the “Dramas militares”, he also wrote “Dramas cómicos”, “Dramas policiales”, and “Dramas del terror”. “Entre dos luces” and “El candidato. Segunda parte de Entre dos luces” (1892, 1893, AR) by Carlos María OcantosOcantos wrote a whole series of 20 novels called “Novelas argentinas”, published between 1888 and 1929, to which also “Entre dos luces” and “El candidato” belong (Ianes 2018, 19). “Las dos tragedias. Primera parte de Pepa Larrica”, “La confesión de un médico. Segunda parte de Pepa Larrica”, “Religión o muerte. Tercera parte de Pepa Larrica” (all 1899, AR) by Rafael Barredo

In the first case, some aspects point to the unity of the work (that the first volume has the same title as the whole cycle, “Libro extraño”, and that the volumes are called “tomo” like different physical volumes of the same novel in other cases). In contrast, others emphasize the independence of the different parts (that the parts have their own title from the second volume on and that they were all published, and thus probably written and finished, in different years). In the second case, all the parts have a common “supertitle”, “Dramas militares”, they are all published in the same year, and each sequel refers to the previous part(s). Even so, all the parts also have their individual title. In the third case, the title of the first novel does not convey any information about a superordinate work, but the subtitle of the second novel indicates that it is a sequel to the first one. These two works were published in subsequent years. In the last case, all the books are numbered parts of the common superordinate title “Pepa Larrica”, and they were all published in the same year, suggesting a united work. A factor complicating the decision in all of these cases is that none of them includes the label “novela”.

As a rule of thumb, a work is considered an independent novel here if it has its own title (and optionally a subtitle indicating the genre) that is not a subtitle of a part (such as “Primera parte: El prólogo de un gran libro”, “Segunda parte: La víspera de un gran día”, etc.), if it has its own structure starting with a first chapter and optionally ending with a trailer indicating the end of the work (e.g., “Fin”, “Fin de la obra”), and if it is optionally published in one or several independent books. These parameters are easy to determine not only for texts that are eligible for the corpus but also for bibliographic entries because viewing the table of contents is enough to decide, and no close reading of the full text is needed.If content-related criteria would be considered, they could be: Is the set of main characters identical in the different parts? Is the setting the same? Is the plot a direct continuation or predecessor of another part‘s plot? The relationships between the various parts of “Libro extraño”, for example, are discussed by Gnutzmann (1998, 183–185).

Following this rule, the parts of the first three cases above are all considered individual novels, while the fourth case as well as the different parts of a work published in several volumes but all carrying the same title, such as “El fistol del diablo” o “Amalia”, are considered one novel. Thereby, the decision of an author (or editor) to publish a novel with its own title in an independent book is, by and large, respected. The relationship between different parts of a novelistic cycle should, however, not be ignored because it can be expected that there are similarities in content and style that influence the results of an analysis of a whole corpus of novels: it is very probable that these works are closer to each other when compared to other independent works. It can also be assumed that the degree of similarity varies according to the closeness of the parts. The books of “Libro extraño” probably have a stronger stylistic relationship than the different parts of a more extensive and looser series such as the ten novels of “La linterna mágica. Colección de pequeñas novelas / Colección de novelas de costumbres mexicanas” (published between 1871 and 1892, MX) by José Tomás de Cuéllar or the thirteen “Leyendas históricas de la independencia” (published between 1886 and 1913, MX) by Ireneo Paz. The existence of cycles and series of novels with different degrees of connectivity is another factor contributing to the great variance of the genre novel in terms of extent which also a quantitative analysis has to deal with. With the decisions made here, a short novel of around 15,000 words is compared to a novel of several hundreds of thousands of words and both to individual parts of sequels of varying length. If text length is not taken into account in the calculations, several shorter parts of a sequel have more influence on the results than a very long novel considered as one. This must be remembered when analyzing the results of the stylistic analysis.

Applied to texts not published independently, the rule of thumb leads to the following decisions: a novel published in a journal, possibly in serial form, is considered one work if it has its own title and structure. Such a work is considered finished if all the existent parts are included, and if there is no obvious interruption of the structure.That way, clearly unfinished and unpublished works are excluded (for example, “Beatriz” (MX) by Ignacio Manuel Altamirano) but works where only a self-contained part was realized and published or transmitted are included (for example, “Ambarina. Historia doméstica cubana. Tomo I” (1858, CU) by Virginia Felicia Auber de Noya). Likewise, shorter novels included in collections are treated as individual works if they fulfill the above criteria.For example, “El pozo del Yocci” (1876, AR) by Juana Manuela Gorriti, which was published as part of the collection “Panoramas de la vida”, or “Las ranas pidiendo rey. Confesiones de una afrancesada (1861–1862)” and “La corte de Maximiliano. Nuevas confesiones de una afrancesada (1863–1867)” (1903, MX) by Victoriano Salado Álvarez, published as parts of two different volumes of the “Episodios Nacionales Mexicanos”. On the other hand, collections of short stories published independently are excluded because each work contained in them has its own title and, eventually, its own structure.Usually, novels can be easily distinguished from collections of short stories regarding their structure because, typically, novels have numbered chapters. Cases that need a closer look are books containing narrative, fictional text, a title not mentioning the genre, and various parts with headings but without numbering because they could either be novels or collections of short stories. A special case in this regard is the work “Pago Chico” (1908, AR) by Roberto Payró. Originally, its parts were published individually and at different times in a literary magazine. In the monographic form, the parts are connected as numbered chapters. The work is characterized as follows by literary historians: “Pago Chico, loser Kranz von Erzählungen mit gemeinsamem Protagonisten, der Stadt Pago Chico” (Dill 1999, 210); “Payró, der in der Provinzstadt Bahía Blanca selbst Opfer politischer Repression geworden war, zeichnete im Verlauf seiner journalistischen Karriere das satirische Bild dieses Systems sowohl in Artikeln als auch in einer Serie von Erzählungen, die 1908 und 1928 unter den Titeln Pago Chico und Nuevos cuentos de Pago Chico in Buchform zusammengefasst wurden” (Rössner 2007, 347–348). Even though the monograph was published during the lifetime of the author and in the time frame of this study, this work is excluded from the bibliography and the corpus because it was not primarily conceived as a novel. Generally, only novels published for the first time between 1830 and 1910 are included.See chapter 3.1.3 (“Limits of the Nineteenth Century”) explaining the chronological limits used here.

So far, only the very general formal criteria of fictionality, narrativity, prose, length, and form of publication were discussed to select texts for the corpus of novels. Although it is intended not to restrict the definition of the novel much further so as not to exclude texts of certain novelistic subgenres from the beginning, two additional criteria going beyond the form are discussed here. The first one refers to the target readership of the novels. In the bibliography and corpus used in this dissertation, only novels written for adults are included. There are also some novels written especially for children which were published between 1830 and 1910 in the three countries of interest here.E.g., the novel “El manantial” (1908, AR) by Emma de la Barra and the series of didactic and popular scientific novels “La ciencia recreativa” (1871–1879, MX) by Alberto F. Arriaga. Although small in number, these are not considered because it is assumed that the target readership influences the writing style, and if they were included, children’s literature would be another influencing factor that would have to be taken into account.

The second additional criterion is a realistic representation of characters and setting, which has been adduced as an important factor in the definition of the novel in order to distinguish it from epic narrative texts and romances. The latter are characterized by mythical heroes and vague and exotic mythical sceneries (Fludernik 2009, 628–629). This criterion does not necessarily hold for all subtypes of the novel, for example, historical, fantastic, and science fiction novels. Nonetheless, it is helpful to exclude some texts which are very far away from the prototypical realistic novel. In this dissertation, texts with non-realistic elements are included as long as these do not dominate the text and as long as the other selection criteria for novels are fulfilled. Two texts that are sometimes included in bibliographies and representations of the nineteenth-century Spanish-American novel are excluded here: “Peregrinación de Luz del Día o Viaje y aventuras de la Verdad en el Nuevo Mundo” (1871, AR) by Juan Bautista Alberdi and “Los dioses de la Pampa” (1902, AR) by Godofredo Daireaux.“Peregrinación de Luz del Día“ is included in Lichtblau’s bibliography of the Argentine novel (Lichtblau 1997, 15–16), whereas “Los dioses de la Pampa” is not, but it is labeled as a novel in the “Biblioteca Virtual Miguel de Cervantes” and in “Wikisource” (see Daireaux [1945] 2001; Daireaux 2006). The protagonist of “Peregrinación de Luz del Día” is the allegorical figure “Verdad” who travels to America to flee from the political and social conditions in Europe. This work has been characterized as a satire, a philosophical dialogue, a novelized allegory, or an allegorical novel (Lichtblau 1997, 16; Molina 2011, 403). It is excluded here because the protagonist is not realistic. In “Los dioses de la Pampa”, Apollo and the Muses travel to Buenos Aires hoping to find the “new Athens”. Disappointed because the arts are disregarded in this big city, they return to Greece. Before leaving, they only catch a glimpse of the Pampa, whose unbeknown, natural gods are presented in the main part of the book and are affiliated with the birth of the Argentine Republic. Because also this text has allegorical traits and, furthermore, no coherent plot, it is excluded, as well.

If one summarizes the selection criteria outlined in the previous sections, the following working definition of the novel can be set up for the present study:

A text is considered a novel if: it was conceived and received as fictional at the time and place of its publication and it is predominantly narrative and it is predominantly written in prose and it is at least 16,000 words or 84 pages long and it is published with an own title and structure, either independently, as part of a monographic collection of works, in a journal or a magazine and it is written for an adult readership and its characters and setting are predominantly realistic.

This definition of the novel is, on the one hand, general, because some of its elements (fictionality, narrativity, prose, realistic representation) correspond to characteristics mentioned in other general definitions of the novel, as well. On the other hand, it is context-specific because the length and publication criteria were derived from the pool of historical texts considered here. The adult readership criterion is one that is probably not critical in general definitions of the novel but that is included here to avoid stylistic outliers. However, as could be seen in the previous sections, even the general criteria need to be interpreted and broken down into specific paratextual and textual markers in order to be applicable to individual texts in a specific historical and cultural setting.

This definition is conceived as classificatory, which means that all the conditions should be met by a text to be considered a novel. That way, clear decisions can be made to include texts into a general corpus of novels, which in turn sets the frame for the analysis of subgenres. Inside this classificatorily defined corpus, alternative definitory concepts of (sub)genre(s) are examined.

Example 5 shows the work entry of the novel “Los casamientos del diablo” (1889, AR) by Enrique Ortega in the digital bibliography, which includes several subgenre labels.

The subgenre labels are encoded in term elements that are characterized further by the attribute type. First, explicit generic labels that occur in the main title, subtitle, or series title, are marked as "subgenre.title.explicit". In the above example, there is an edition of the novel with the subtitle “novela histórica americana”. Because the generic identity is determined on the work level, information about the subgenre is taken from all the work’s editions. Therefore, if there are several editions and only one carries the explicit subgenre label, it is nonetheless included here. If there are several editions with differing subgenre labels, all of them are considered. The second type of label is called "subgenre.title.explicit.norm" and contains a normalized version of the explicit subgenre label. The resp attribute indicates by whom the normalization was done. For the novel at hand, the label “novela histórica americana” is normalized to several individual subgenre labels: “novela histórica”, “novela americana”, and “novela”. The primary purpose of the normalization step is to make the explicit subgenre labels comparable in computational analysis. In the current case, the first label “novela histórica” refers to a subgenre predominantly determined by the theme of the novel. The second label “novela americana” points to the cultural-geographical and linguistic origin and identity of the novel. In the bibliography, there are also novels with the subtitle “novela argentina”, “novela cubana”, “novela mexicana”, “novela original”, etc. These kind of labels either refer to the continent (“americana”), to the country (“argentina”, “cubana”, “mexicana”), or to the fact that the novel was originally written in Spanish and not translated (“original”).The usage of the label “novela original” in the Spanish context is described by Botrel as follows: “Las normas/formas tipográficas bibliográficas permiten también observar cómo después de un período en el que se precisa el origen de la novela («novela escrita en francés por Mr.» o «Madama...», «en inglés por Mistress...» o «Sr...» y «traducida al castellano por...» iniciales) al preeminencia del título unida con la hispanización casi sistemática de los nombres de los autores traducidos (Javier de Montepín, Pablo Feval, etc. y la importancia numérica de las traducciones, con la desaparición de la mención del traductor, al menos en las referencias bibliográficas, hace que el género novela venga disociado de una por otra parte deseada hispanidad y asociado con una patronímica y toponimia extranjerizante, como producto extranjero o, más probablemente, asimilado. La mención «novela original» o «española» introducirá durante cierto tiempo una distinción poco decisiva, estadísticamente al menos” (Botrel 2001, paras 12, 13). The third label extracted from the example, “novela”, refers to the genre of the text. It is encoded as a subgenre label here, as well, because in the bibliography, only about half of the works carry the explicit label “novela”.Of 829 works in the bibliography, only 403 carry the explicit label “novela”. Because all the works that are included follow the selection criteria for novels defined in chapter 3.1.1 above, it is assumed that the explicit label “novela” points to a subtype of all the texts that can be considered as novels formally. Other labels usually designating genres, such as “cuento”, “drama”, “ensayo”, or “leyenda”, are treated in the same way.

After the explicit generic signals of the titles, implicit signals are evaluated and captured in terms of the type "subgenre.title.implicit". Here again, the attribute resp serves to indicate who made the interpretation. In this case, two implicit labels, “novela sentimental” and “novela romántica”, are recorded. The word “casamientos” in the title is interpreted as a reference to a sentimental plot, and the whole title “los casamientos del diablo” is interpreted as a sign of a novel of the romantic current.In romantic novels, it is common that issues and people are portrayed in black and white (e.g., the good and the bad), exaggerated, or dramatized, and the word “diablo” points in that direction.

The above example does not contain terms of the second group ("subgenre.litHist") because for the novel “Los casamientos del diablo”, no statements about its subgenre were found that were made by literary historians. Therefore, the terms of the third group ("subgenre.summary") only take up the values that were inferred from the title. The summary at the end of the different subgenre terms has the function of organizing the previous data into categories of generic information in order to enhance the comparability of the terms throughout the bibliography for further analysis. What kind of generic information is given, is indicated in the part of the type attribute after "subgenre.summary". The summary values have five subtypes: "signal", "theme", "current", "identity", and "mode". Terms of the type "subgenre.summary.signal" contain all the subgenre labels that were signaled by the work title either explicitly (marked as "subgenre.summary.signal.explicit") or implicitly (subgenre.summary.signal.implicit). In the above example, all the subgenre labels are derived from signals of the text. However, in other cases, there are further subgenre labels that were assigned to the work by critics, but that cannot be deduced from the work’s title. The second subtype of the summary values is "theme". Terms of this type contain all the labels that refer to subgenres defined primarily or in part by the theme of the text. In the current example, there are two thematic labels: “novela histórica” and “novela sentimental”. The first one was given explicitly and is therefore encoded as a term of the type "subgenre.summary.theme.explicit" whereas the second one was deduced from the title and is marked as "subgenre.summary.theme.implicit".

The first of the “theme” terms carries the attribute cligs:importance with the value "2". With this attribute, an order of priority is given for cases with several subgenre terms of the same type. It was decided to use this attribute only for “theme” and “current”, i.e., for subgenre labels belonging to these two categories. These are the types of subgenre labels that are at the center of interest of this dissertation. Furthermore, most literary histories and critical studies refer to novelistic subgenres of this kind. As to the priorities, in general, only one high priority ("2") is assigned, while the other terms without this attribute are interpreted as low-priority terms. Just as the normalization of explicit titles serves to enhance comparability, this prioritization has the pragmatic function of being able to select one value for each subgenre term of the types “theme” and “current” for cases where unique values are needed in an analysis. However, it is a simplification because it is ultimately not possible to map different subgenre assignments to a discrete numerical system as they usually represent different perspectives on the literary work.In the case at hand, “novela sentimental” refers primarily to the content and plot of the novel, while “novela romántica” points to the literary current. Nevertheless, there are overlaps between both terms. As rules of thumb, terms deduced from explicit signals are rated higher than those going back to implicit ones. Furthermore, signals that are stronger are valued higher, for example, if there are several signals pointing to a certain subgenre and only one signal points to another. In addition, terms that are mentioned by literary critics are valued higher than those that are not.

After the thematic terms, those referring to literary currents are listed ("subgenre.summary.current"). In the above example, there is only one term of this kind, the “novela romántica”. The term “novela americana” is encoded as a term of the type "subgenre.summary.identity". Finally, there are two subgenre labels grouped into the “mode” category: “novela histórica” belongs to the category "subgenre.summary.mode.reality" and “novela” to "subgenre.summary.mode.representation". The “mode” group contains labels that are not thematic and do not refer to literary currents or the cultural or linguistic identity of the works. Instead, these are labels indicating how the works relate to extratextual circumstances or to the way the text is organized and presented. In the example, "mode.reality" designates labels that involve the relationship of the text to reality. Usually, a historical novel intends to present settings and events of the past, but not the present reality. "mode.representation" includes labels that indicate how the novel is organized and presented linguistically. The term “novela” means that the text is presented in the narrative mode and not, for example, as a dramatic text. As can be seen in the example, some subgenre labels are repeated in the summary, in this case, “novela histórica”, which falls into the two categories "subgenre.summary.theme" and "subgenre.summary.mode.reality". On the other hand, each novel can have several subgenre labels of the same kind, as the two thematic labels “novela histórica” and “novela sentimental” of this example show. Finally, because all the values in the summaries are normalized, the summary terms also carry a resp attribute that shows who entered the values.

The system of the summary values needs to be explained further. Which categories were chosen, for which reasons, and which values can they take? The subgenre categories chosen (“theme”, “current”, “identity”, and “mode” with further subtypes) are not generally exhaustive from a genre theoretical perspective and not congruent to one specific theoretical model of genre. Instead, they reflect the generic signals that occur in the collection of novels represented in the bibliography and the corpus, as well as the terms with which the subgenres of these novels are described by literary historians. There are general models for describing the different levels to which generic labels might refer. Some of these models include more categories than the ones chosen here, and others have fewer or different categories. The categories chosen here are, for the most part, derived from a model developed by Wolfgang Raible (Raible 1980). In table 2, the different subgenre categories used in the present model are listed, exemplified, and commented on, and the levels of Raible’s model that correspond to the ones here or are similar to them are given. Types of summarizing subgenre labels. Kind of subgenre label Value of type Examples Explanation Level in Raible’s model signal subgenre.summary.signal novela histórica, novela naturalista, novela original, memorias subgenre labels that are derived from explicit or implicit signals of the novel - theme subgenre.summary.theme novela gauchesca, novela histórica, novela sentimental subgenre labels that refer to a main theme of the novel Objektbereich current subgenre.summary.current novela romántica, novela realista, novela naturalista subgenre labels that refer to the literary current of the novel - identity subgenre.summary.identity novela americana, novela mexicana, novela original subgenre labels that refer to the cultural-geographical and linguistic identity of the novel - mode subgenre.summary.mode novela epistolar, novela fantástica, novela humorística, cuadros, drama, memorias subgenre labels that refer to the mode the novel is narrated in / the form it is presented in Kommunikationssituation, Verhältnis zwischen Text und Wirklichkeit, Medium, sprachliche Darstellungsweise intention subgenre.summary.mode.intention novela cómica, novela moralista, novela de propaganda subgenre labels that refer to the aim the author/narrator pursues with the novel Kommunikationssituation attitude subgenre.summary.mode.attitude novela política, novela satírica subgenre labels that refer to the attitude the author/narrator has towards what is represented in the novel Kommunikationssituation reality subgenre.summary.mode.reality novela científica, novela fantástica, novela histórica, leyenda subgenre labels that refer to the relationship between the novel and reality Verhältnis zwischen Text und Wirklichkeit medium subgenre.summary.mode.medium novela epistolar, croquis, cuadros, páginas, panorama subgenre labels that refer to the medium that the novel uses (also in a figurative sense) Medium representation subgenre.summary.mode.representation cuento, drama, ensayo, episodios, novela subgenre labels that refer to the mode the novel is represented in linguistically (or narratively) sprachliche Darstellungsweise

Different types of subgenre labels were already introduced with the example “Los casamientos del diablo” above. In what follows, some more general considerations regarding the system of subgenre labels developed here are made, beginning with the category “signal”. It comprises subgenre labels that are derived from explicit or implicit signals of the novel, either in paratextual elements (in a title, subtitle, series title, preface, epigraph, etc.) or in the opening of the texts. These labels can be of any of the following kinds of labels (thematic labels, labels referring to literary currents, or other types of labels). For the bibliography, they were only derived from the titles because this is the only paratextual information directly available in the bibliographic records. For the corpus, other signals were evaluated as well. Apart from that, the approach to assigning the subgenres is the same for the bibliography and the corpus. Therefore, the general points are explained in this section, while only the additional corpus-specific aspects are explained below in chapter 3.3.4.

It is important to note that signals can be explicit subgenre labels, for example, the subtitle “novela histórica americana” above. Besides that, they can also be aspects of the title (and other paratextual and textual elements) that can be interpreted in terms of subgenre labels, for example, “casamiento” as pointing to a sentimental novel and “diablo” to a romantic novel. The evaluation of the signals thus involves a significant interpretive step, and it presupposes knowledge about possible subgenres. The knowledge is, on the one hand, derived from the bibliography and the corpus itself (which subgenres occur frequently and what are their characteristics?) and, on the other hand, from representations of the subgenres in literary-historical works. By encoding many steps of this interpretation process (starting from explicit labels, going on to normalized values, mentioning implicit signals, summarizing all in the categorized labels, and keeping their origin as “explicit” or “implicit”), it should be possible to follow the decisions made here closely for each of the novels in the bibliography. Nevertheless, another encoder might have reached other results. The position adopted here is that the genre or subgenre of a text cannot be determined unequivocally without presuppositions. To avoid the influence of the own previous knowledge or the necessity of previous definitions, one could opt for only referring to explicit generic labels. However, in the case of the bibliography and corpus at hand, this would have led to a very reduced setup because only some kinds of explicit subgenre labels are very frequent. Table 3 lists the top most frequent explicit labels, ordered by the frequencies of the normalized versions.The counts are based on the final bibliography. Novels carrying several different labels were counted twice, and the percentages were rounded. See also chapter 4.1.5 for an overview of the subgenres in the bibliography and the corpus. The normalized frequencies are higher because some novels carry explicit labels, e.g., “histórico” or “costumbres”, but not the whole label involving the word “novela”, i.e., “novela histórica” or “novela de costumbres”, or they contain additional elements. For example: “novela de carácter histórico”, “episodio histórico”, “leyenda histórica”, “historia novelada”; “novela original de costumbres”, “cuadro de costumbres”, “boceto de costumbres”, “ensayo de costumbres”, etc. All the subgenre labels that occur at least ten times in the normalized version are included. In addition, the number of novels without any explicit label is also given. The script used to determine the subgenre label counts given in this and the following tables in this section is available at https://github.com/cligs/scripts-nh/blob/master/corpus/frequencies-subgenre-labels.xsl and the full table of frequencies can be viewed at https://github.com/cligs/data-nh/blob/master/corpus/bibliography-subgenre-labels/frequencies-explicit-labels.csv. Accessed March 7, 2020. Top most frequent explicit subgenre labels in the bibliography. Subgenre label Frequency explicit Frequency explicit normalized novela 398 48 % 403 49 % novela histórica 73 9 % 133 16 % novela original 97 12 % 113 14 % novela mexicana 6 1 % 67 8 % novela de costumbres 25 3 % 57 7 % episodios 64 8 % 54 7 % memorias 49 6 % 54 7 % leyenda 42 5 % 44 5 % novela cubana 18 2 % 35 4 % drama 25 3 % 28 3 % novela nacional 0 0 % 26 3 % historia 22 3 % 25 3 % cuento 15 2 % 15 2 % novela argentina 5 1 % 15 2 % novela social 2 0 % 13 2 % novela americana 4 0 % 12 1 % escenas 11 1 % 12 1 % novela policial 0 0 % 11 1 % novels without any explicit label 207 25 % - -

As can be seen, only a few of the top most frequent explicit labels refer to subgenres of the novel in common sense, i.e., labels related to the themes of the novels: 16 % of the works carry the label “novela histórica”, 7 % the label “novela de costumbres”, 2 % the label “novela social”, and 1 % the label “novela policial”. The other frequent labels are either of a very general nature (“novela”, “leyenda”, “drama”, “historia”, “cuento”) or they refer to aspects of the novels that are usually not focused on in subgenre studies, such as the identity of the texts (“novela original”, “novela mexicana”, “novela cubana”, “novela nacional”, “novela argentina”, “novela americana”) or the way the text is structured and presented linguistically (“episodios”, “memorias”, “escenas”). Furthermore, it can be noted that even the topmost frequencies decrease sharply. Finally, not all the novels have explicit labels: 25 % of the novels in the bibliography do not convey any generic information explicitly.

For some of the subgenre labels that entered the top list, an author and series bias can be noted.Strictly speaking, this is no bias, though, because the novelistic production of the time is as it is: some authors wrote more novels than others, and some wrote whole series of novels of a certain subgenre. Most of the occurrences of the terms “drama” and “novela policial”, for example, stem from the numerous novels written by the Argentine author Eduardo Gutiérrez, which are organized in series and carry subtitles of the form “dramas policiales”, “dramas militares”, “dramas cómicos”, etc. The many “episodios”, “memorias”, “leyendas”, “novelas nacionales”, and “historias” are connected to series of historical novels, some of which are called “episodios nacionales” or “leyendas históricas”. It was decided not to keep the combined labels in the normalized form, though, because even if some of the combinations occur several times and lead to correlations in the frequencies of the labels, their components are also part of other kinds of subtitles. Furthermore, the combinations of individual subgenre labels in the subtitles are so varied and often individual that it would be impossible to compare them without any normalization step. The original combinations can still be reproduced because they are encoded in terms of the type "subgenre.title.explicit".

If only the thematically oriented explicit subgenre labels would be regarded, most of the 829 novels in the bibliography would have had to be considered general fiction because only 36 % of the novels have such labels.See https://github.com/cligs/data-nh/blob/master/corpus/bibliography-subgenre-labels/frequencies-explicit-thematic-labels.csv for a full table of frequencies of explicit thematic labels in the bibliography. Accessed March 7, 2020. However, many of the novels have been interpreted as belonging to certain subgenres, and many signal their subgenre(s) implicitly. Some well-known and also relatively frequent subgenres of the novel are rarely indicated explicitly, for example, sentimental novels. In the whole bibliography, there is only one novel with the explicit subtitle “novela sentimental”, but many more novels can be assigned to this subgenre. When also implicit signals and literary-historical assignments are included, the picture of the top most frequent subgenres changes, as table 4 shows.The table shows the top ten subgenre assignments and the number of novels without any assignment. See https://github.com/cligs/data-nh/blob/master/corpus/bibliography-subgenre-labels/frequencies-subgenre-labels.csv for the full table. Accessed March 7, 2020. Top most frequent subgenres in the bibliography. Subgenre label Frequency absolute Frequency relative novela 404 49 % novela romántica 269 32 % novela sentimental 252 30 % novela histórica 244 29 % novela social 177 21 % novela de costumbres 133 16 % novela realista 122 15 % novela original 113 14 % novela naturalista 81 14 % novela mexicana 67 8 % novels withouth any subgenre assignment 51 6 %

Many more novels in the bibliography are covered with this approach. So for the reasons given, it was decided to include interpretive subgenre labels, as well. Both implicit signals evaluated by the author of this dissertation and assignments made by other literary historians are considered as such. The difference between both is that for the other literary-historical labels, it is not known in detail on what bases they were assigned.This is usually only outlined explicitly in specific and comprehensive studies of certain subgenres, e.g., in Rivas (1990) for the anti-slavery novel or Schlickers (2003) for the naturalist novel. In contrast, in overview works such as general literary histories, the criteria for the assignment of novels to subgenres are normally not explained. Literary-historical labels are discussed in more detail below. The interpretive labels worked out here are derived from specific textual signals: the titles (in the case of the bibliography) and additional paratextual elements (in the case of the corpus). The decisions rest on a certain set of subgenres taken as the basis for interpreting the implicit signals. This set does not comprise all of the existing explicit labels, though. Instead, the focus is on subgenres related to themes and literary currents, as these are the kinds of subgenres most often referred to in literary histories and also because there are known concepts of these subgenres that can be used.See chapter 2.3 for a presentation of the subgenres related to themes and literary currents. In addition, the set contains some subgenres that repeatedly occur as explicit labels in the bibliography, and that can be inferred from textual signals in other cases, even if they are not part of the critical subgenre canon, for example, the “novela contemporánea”. Table 5 contains the set of subgenres used to interpret implicit signals. Like the list of kinds of subgenres, this set is also not exhaustive from a general perspective on the subgenres of the novel. Instead, it is based on the relevance of the subgenres for the bibliography and the corpus. Set of subgenres used as a basis for the interpretation of implicit signals. Kind of subgenre Subgenre labels theme Künstlerroman, novela abolicionista, novela biográfica, novela científica, novela contemporánea, novela criminal, novela de aventuras, novela de costumbres, novela de familia, novela de la ciudad, novela de misterio, novela de viajes, novela didáctica, novela doméstica, novela filosófica, novela gauchesca, novela histórica, novela humorística, novela indigenista, novela militar, novela moralista, novela picaresca, novela política, novela psicológica, novela regional, novela sentimental, novela social current novela romántica, novela realista, novela naturalista identity novela regional mode.intention novela didáctica, novela humorística, novela moralista mode.attitude novela abolicionista mode.reality novela científica, novela contemporánea, novela de misterio, novela histórica mode.representation novela filosófica, novela psicológica

Some of the subgenres of this set that are included in the thematic group also belong to other levels of the model defined above.Which subgenre labels are grouped in which subgenre category here is explained in detail below (see table 10). They are listed again in the lower part of the table for the sake of completeness. Nevertheless, when this set of subgenres was applied to interpret implicit signals, the focus was on the thematic aspects. Furthermore, the thematic subgenres are placed on different levels of generality. The types novela de familia, novela de la ciudad, and novela doméstica are more specific than, for example, novela social, and they could also be subsumed under the latter term. That terms of different levels of generality occur in the list is because there are signals in the bibliography and the corpus that can best be interpreted with these labels. For example, the novel “La familia de Sconner” (1858, AR) by Miguel Cané (father) is interpreted as a novela de familia and a novela social and the novel “La sociedad y sus víctimas. Escenas bonaerenses” (1902, AR) by Matías Calandrelli both as a novela de la ciudad and a novela social.Which subgenre labels are related to others is explained below in table 10.

Following up on the question of how many novels are covered when also interpretive subgenre labels are included, tables 6 and 7 show the most frequent subgenre labels related to themes and literary currents, including explicit as well as implicit signals and literary-historical assignments.The table shows the top ten plus the number of entries in the bibliography without this type of subgenre label. Full tables are available at https://github.com/cligs/data-nh/blob/master/corpus/bibliography-subgenre-labels/frequencies-thematic-labels.csv and https://github.com/cligs/data-nh/blob/master/corpus/bibliography-subgenre-labels/frequencies-labels-currents.csv. Accessed March 8, 2020. Top most frequent thematic subgenre labels in the bibliography. Subgenre label Frequency absolute Frequency relative novela sentimental 252 30 % novela histórica 244 29 % novela social 177 21 % novela de costumbres 133 16 % novela política 51 6 % leyenda 44 5 % novela criminal 37 4 % novela de la ciudad 27 3 % novela indigenista 27 3 % novela gauchesca 21 3 % novels withouth thematic label 134 16 % Frequencies of subgenre labels related to literary currents in the bibliography. Subgenre label Frequency absolute Frequency relative novela romántica 269 32 % novela realista 122 15 % novela naturalista 81 10 % novela modernista 8 1 % novela verista 5 1 % novela clasicista 3 0 % novels without label of literary current 424 51 %

The four biggest thematic groups are sentimental, historical, social novels, and novels of manners (novela de costumbres). For 16 % of the novels, no thematic label could be assigned. The literary current most frequently assigned are romantic novels, followed by realist and naturalist novels. In the case of the literary currents, more than half of the novels in the bibliography do not have any label of this kind (51 %). One reason for this is that the literary current is usually not given explicitly: there are only five novels in the whole bibliography with explicit signals for the “novela naturalista” and six for the “novela realista”. The term “novela romántica” does not occur at all. The second reason is that literary currents are mainly a concern of literary historians, and for 48 % of the novels in the bibliography, no assignments made by literary historians could be found. An important point to consider when looking at the numbers is that they do not mean that the novels that do not have a certain subgenre label do not possibly belong to that subgenre. The distribution of subgenre labels only indicates that these are the cases where information (explicit, implicit, literary-historical) is available.

Going on to the discussion of literary-historical labels, example 6 shows the entry of the work “Santa” (1903, MX) by Federico Gamboa. For this work, there are no labels from the first group (explicit or implicit labels inferred from the title) but many from the second (labels taken from literary-historical sources).

In various literary-historical works, “Santa” is classified as a naturalistic novel. The literary-historical labels are collected in terms of the type "subgenre.litHist", and the respective source is given in the attribute resp. All the different literary-historical assignments are summarized in terms of the type "subgenre.listHist.interp". In the case of “Santa”, the subgenre assignments made by critics are quite unanimous. Besides being classified as a naturalistic novel, “Santa” is also described as a realist and a social novel. Like explicit and implicit signals, literary-historical labels are also summarized and categorized further in terms of the type "subgenre.summary", following the procedures explained with the first example above. For “Santa”, the literary-historical labels result in three summary terms: “novela social” is encoded as a thematic label ("subgenre.summary.theme.litHist"), and “novela naturalista” and “novela realista” are grouped as labels referring to literary currents ("subgenre.summary.current.litHist"). The label “novela naturalista” is weighted higher than “novela realista” because it is mentioned more often by literary critics (@cligs:importance="2") and also because it occurs as an implicit signal in the paratext of the novel, as indicated with the terms "subgenre.summary.signal.implicit" and "subgenre.summary.current.implicit". This implicit signal is not derived from the title of the novel, though, but from other paratextual elements. In the case of “Santa”, this is possible because the novel is part of the corpus and was analyzed in more detail.See chapter 3.3.4 about the assignment of subgenre labels for the texts in the corpus.

In the same way that the explicit generic information occurring in the titles and other paratexts is normalized, also the subgenre labels collected from literary-historical works are interpreted and standardized because not all literary historians use the same terminology. Table 8 lists the different interpretive values contained in terms of the type "subgenre.litHist.interp" as well as the kinds of subgenres with which these values can be associated. Set of subgenres used as a basis for the interpretation of literary-historical subgenre labels. Kind of subgenre Subgenre labels theme Bildungsroman, crónica, Künstlerroman, memorias, novela abolicionista, novela científica, novela criminal, novela de aventuras, novela de costumbres, novela de familia, novela de la ciudad, novela didáctica, novela documentaria, novela de misterio, novela de viajes, novela fantástica, novela gauchesca, novela histórica, novela indigenista, novela moralista, novela picaresca, novela política, novela psicológica, novela regional, novela sentimental, novela social current novela clasicista, novela modernista, novela naturalista, novela realista, novela romántica, novela verista identity novela regional mode.intention novela didáctica, novela moralista mode.attitude novela abolicionista, novela política, novela satírica mode.reality novela científica, novela de misterio, novela fantástica, novela histórica mode.medium novela epistolar mode.representation crónica, memorias, novela documentaria, novela epistolar, novela psicológica

As for the set of subgenre labels used to interpret titles and other paratexts of the novels, the set of labels interpreted from literary-historical subgenre labels also focuses on thematic labels and labels referring to literary currents. The other kinds of subgenre labels are only of secondary importance in the interpretation process. A subgenre that is often mentioned in literary-historical works is the novela costumbrista (Gálvez 1990, 100–101; Remos y Rubio 1935, 57–109; Sánchez 1953, 227–256). It was decided to interpret this label as novela de costumbres because, historically, the novels carried the latter label. Costumbrismo is also often described as a literary current (Dill 1999, 155–157; Rössner 2007, 146–147), but this aspect is not highlighted here because the novelas de costumbres were written and published throughout the whole the nineteenth century. Some of them can be attributed to the romantic current, others to the realist current, and even some naturalistic novels carry labels including the word “costumbres”.Examples for romantic novelas de costumbres are “Ironías de la vida” (1851, MX) and “La hora de Dios” (1865, MX) by Pantaleón Tovar and the series of “novelas de costumbres” written by José Tomás de Cuéllar (Fernández-Arias Campoamor 1952, 63–65e been characterized as ). Novels of this type that havrealist are “La familia Quillango” (1880, AR) by José María Cantilo and “Antón Pérez” (1903, MX) by Manuel Sánchez Mármol (Fernández-Arias Campoamor 1952, 86; Gálvez 1990, 126–127). Naturalistic novels carrying the label “costumbres” in their title are “Quimera. Boceto de costumbres” (1899, AR) by José Luis Cantilo and “Fruto vedado (Costumbres argentinas)” (1884, AR) by Paul Groussac. On Costumbrismo as a longer lasting phenomenon, Fernández-Arias Campoamor writes: “Los novelistas románticos que fueron costumbristas constituyen el puente tendido entre el romanticismo y el realismo. Costumbrismo cultivado ocasionalmente, en realidad, lo hubo siempre en todas las literaturas [...] Pero el costumbrismo como inclinación extensa y generalizada se inicia en el romanticismo” (Fernández-Arias Campoamor 1952, 56). Kohut, too, points to the significance of Costumbrismo for several other literary currents: “Die Abgrenzung zwischen Romantik, Realismus und Naturalismus gestaltet sich schwierig. [...] Die Problematik wird durch den sogenannten Costumbrismo zusätzlich kompliziert, der wie in Spanien zwischen Romantik und Realismus steht. Zum Realismus gehört die Zuwendung zur Gesellschaft, zur Romantik die häufig idyllisierende Perspektive. [...] Wichtiger als der Costumbrismo als eigenständige literarische Richtung ist die entsprechende Einfärbung zahlreicher realistischer bzw. Naturalistischer Romane. So gab der Chilene Alberto Blest Gana seinem Roman Martín Rivas (1862) den Untertitel Novela de costumbres político-sociales, der Argentinier Lucio Vicente López seinem Roman La gran aldea (1884) den Untertitel Costumbres bonaerenses” (Kohut 2016, 196). Other standardizations were made. For example, novels related to gauchos were all normalized to novela gauchesca, novels related to cities to novela de la ciudad, novels about indigenous people to novela indigenista, novels about the system of slavery in Cuba to novela abolicionista, and novels related to crimes to novela criminal.Alternative terms are novela del gaucho, novela urbana, novela indianista, novela antiesclavista, and novela policial. In some cases, the alternative formulations do not involve differences in the meaning of the terms (this is assumed for novela gauchesca and novela del gaucho and for novela de la ciudad and novela urbana). In other cases, there are slight differences in the meaning. Novela criminal, for example, is more general than novela policial, and was therefore preferred here. Novela abolicionista was preferred over novela antiesclavista because it is the term that was in use historically. The term novela indigenista was preferred over novela indianista because it is more neutral. The novela indianista refers to romantic works that re-evaluate the historical past before the conquest of the Americas (Meléndez 1961; Rössner 2007, 144).

In Bib-ACMé, only a selection of literary-historical sources was used for the assignment of subgenre labels. The critical literature on Spanish-American novels is vast, so a choice had to be made. Works of different scopes were selected, preferably those where the assignment of a novel to a subgenre is explicit. The sources used are listed in table 9. Literary-historical sources for the assignment of subgenres. Scope Title Editor / Author Year Spanish-American literature Geschichte der lateinamerikanischen Literatur im Überblick Dill, Hans-Otto 1999 Spanish-American novel La novela hispanoamericana (hasta 1940) Gálvez, Marina 1990 Spanish-American novel Proceso y contenido de la novela hispano-americana Sánchez, Luis Alberto 1953 Nineteenth-century Spanish-American novel Evolución de la novela hispanoamericana en el siglo XIX Varela Jácome, Benito 2000 Spanish-American romantic novel La novela romántica en Hispanoamérica Suárez-Murias, Marguerite C. 1963 Nineteenth-century Argentine novel The Argentine novel in the nineteenth century Lichtblau, Myron I. 1959 Argentine novel (1838–1872) Como crecen los hongos. La novela argentina entre 1838 y 1872 Molina, Hebe Beatriz 2011 Cuban novel Tendencias de la narración imaginativa en Cuba Remos y Rubio, Juan J. 1935 Mexican novel Novelistas de Mejico. Esquema de la historia de la novela mejicana (De Lizardi a 1950) Fernández-Arias, Campoamor, José 1952 Nineteenth-century Mexican historical novel The Mexican historical novel. 1826–1910 Read, John Lloyd 1939 Spanish-American naturalistic novel El lado oscuro de la modernización: estudios sobre la novela naturalista hispanoamericana Schlickers, Sabine 2003 Cuban naturalistic novel El Naturalismo en la novela cubana Molina, Sintia 2001

In literary-historical works, assignments to subgenres sometimes occur in the text and also often through the structure and organization of a literary history if a novel is mentioned in a section carrying the title of a subgenre.An example of an assignment made in the running text is: “De este episodio tomó apuntes el joven subteniente [Heriberto Frías], dándole base para construir una novela histórica a la que dió el nombre de Tomochic” (Fernández-Arias Campoamor 1952, 83). In Dill’s literary history of the Spanish-American novel, several subchapters are entitled with subgenre labels so that the texts mentioned in them can be attributed to these subgenres. The chapter “Der Roman der Romantik”, for example, is subdivided into “Der politische Roman”, “Der historische Roman”, “Der indianistische Roman”, “Der kubanische negristische Roman”, and “Der sentimentale Roman”. Works mentioned in the chapter “Der politische Roman” are, for example, “Amalia” (1855, AR) by José Mármol or “Clemencia” (1869, MX) and “El Zarco” (1901, MX) by Ignacio Manuel Altamirano (Dill 1999, 125–139). However, clear assignments are not always made because many novels are rather described in their relationship to a certain subgenre and are also evaluated as mixtures or deviations.For example, Dill mentions Emilio Rabasa’s novels in the chapter on the naturalistic novel but designates them as anti-naturalistic (Dill 1999, 170). In her work on the Spanish-American naturalistic novel, Schlickers dedicates her own subchapter to each of the novels that she included in her corpus. In these detailed discussions of the works, she reasons about how each of the novels is in accordance with the criteria that she set up for a novel to be naturalistic and, in some cases, comes to the conclusion that they are not, e.g., for the novel “León Zaldívar” (1888, AR) by Carlos María Ocantos: “Resulta que León Zaldívar no es una novela naturalista, sino una mezcla entre novela rosa/folletinesca y costumbrista; Lichtblau [...] califica la novela de ‘happy combination of romantic and realistic elements’. [...] A nivel de la expresión, la distancia respecto a la poética naturalista se marca por los frecuentes comentarios del narrador que marca su hic et nunc y coincide ideológicamente tanto con el autor implícito como con el protagonista idealizado [...]. A pesar de una escritura por lo general ‘realista’, no se concretan ni el tiempo de la historia [...], ni se citan nombres –por ejemplo de los políticos que se critican. Así, la novela gana en dimensión alegórica lo que pierde en valor referencial, facilitanto así la transmisión y recepción masiva de la intención de sentido: la crítica del materialismo, la reivindicación de la sincera práctica de la religión católica y la idealización de la mujer abnegada y sumisa, para terminar con una lección moralizante: [...]” (Schlickers 2003, 200). In this way, Schlickers checks the novels that can be provisionally assigned to the naturalistic current because of their theme or certain generic signals against the more strict formal criteria that she set up for the subgenre. Regarding the discussion of novelistic subgenres, many literary-historical works tend to focus on the individuality of the works. Subgenres provide a frame for the description of groups of works. However, they are rarely understood as strict classes and more often as anchor points that help to analyze and represent a complex overall novelistic production in an ordered way.This can be seen, for example, in the monograph “The Mexican historical Novel” authored by Read. In order to give a comprehensive overview of the Mexican nineteenth-century fiction that can be considered historical, he also includes cases that are only historical on certain levels of the narration, e.g.: “López Portillo y Rojas pronounced Altamirano’s Clemencia the best Mexican novel of its time. [...] The setting of the story is the region around Guadalajara in December, 1863, the year of the French occupation of Mexico. The historical material serves only as a frame for the actions of two officers in the army of the republic [...]. A conflict developed over the affections of a beautiful woman, with a none too pleasant result” (Read 1939, 164–166) or “Martínez de Castro’s novel Eva, published in 1885, though not purely historical in nature, has enough of historical background to justify its inclusion in this study” (Read 1939, 256).

The subgenre labels assigned to the novels by literary historians have a different status than those occurring explicitly in the titles and other paratexts of the novels and those that are or signaled implicitly in the texts. Literary historical labels do not represent a contemporary perspective, and the agents who decided on the labels are different. They are scholars of the twentieth and twenty-first centuries aiming to provide systematic perspectives on the novelistic production of the nineteenth century and not authors, editors, or contemporary critics. Nevertheless, the labels do not behave differently as a whole. Like every author, editor, or contemporary might use the labels in a slightly different manner, also the approaches of scholars can differ. Definitions of subgenres and criteria for the composition of the corpus are more often given in studies that concentrate on one subgenre of the novel.See, for example, Rivas (1990, 121–154) and Schlickers (2003, 27–46). In general literary histories that are dedicated to a whole range of genres and subgenres, it is usually not explicitly discussed how the works were assigned to the subgenres. Comparative studies concerned with several types of subgenres are rare.An example of a study where several different subgenres of the novel are defined is Molina (2011). It is not a comparative study in a strict sense because rather than comparing the different subgenres, her goal is to describe and systematize the whole novelistic production in Argentina between 1838 and 1872. However, Molina finds that all the novels can be classified into one or several of four main types: “novelas históricas”, “novelas políticas”, “novelas socializadoras”, and “novelas sentimentales” (246–386). When no definitions of the subgenres are given, it can only be hypothesized how the assignments come about: they might be based on explicit historical labels, on previous assignments made by other literary historians, or on background knowledge and reading experience. As a result, the focus here is not on how the subgenre terms are defined in each case but the fact that they are signaled by literary historians. Together with the explicit and implicit signals found in the texts of the novels themselves, the subgenres emerge as categories that are collectively defined, and this includes a certain fuzziness.

Returning to the levels into which the subgenre labels – explicit, implicit, and literary-historical ones – are sorted in the summary, some more remarks are to be made. Regarding the relationship to the model proposed by Raible, it is evident that his model is a semiotic one in the linguistic sense of the term. Raible’s model, designed for literary and also non-literary genres, covers general aspects of the communication situation, the content and structure of the message, the medium, and the linguistic representation. With the level concerning the relationship between the text and reality, he addresses a point specifically relevant for literary texts. However, the aspects of the literary currents and cultural and linguistic identities of the texts are not covered by him.In the description of the level “Objektbereich”, he mentions the Old French terms “matière de Bretagne” and “matière de Rome” that designate legendary material concerned with the history of Brittany and Rome, but there is no own level for identity-related (culturally specific or national) texts (Raible 1980, 343). A model similar to Raible’s is the one developed by Jean-Marie Schaeffer, who also starts from the assumption that a literary work is a complex semiotic object and that generic terms can refer to different levels of this object. Broadly, Schaeffer distinguishes between the communicative act (“L’acte communicationnel”) and the realized discursive act (“L’acte discursif réalisé”). Raible’s level “Kommunikationssituation” overlaps with Schaeffer’s “L’acte communicationnel” and the “Objektbereich” with the “L’acte discursif réalisé”. The other levels defined by Raible can also be associated with Schaeffer’s two main levels. That is, also in Schaeffer’s semiotic approach, the levels that are named “identity” and “current” here are not included in the core model. Nevertheless, Schaeffer discusses these aspects as an aside:

Parmi les noms de genres que j’ai collectés, certains se réfèrent cependant à des déterminations qui sont irréductibles aux cinq niveaux de l’acte verbal que je viens de distinguer. J’ai indiqué plus haut que le modèle de la communication dont je me servais ne tenait pas compte du contexte, du lieu et du temps. Or, il existe de nombreux noms de genres qui sont composés à l’aide de déterminants de lieu ou de temps. Ainsi des termes comme tragédie élisabéthaine, tragédie classique, roman antique, sonnet baroque, etc., délimitent des traditions dans le temps, c’est-à-dire se réfèrent à des genres historiques au sense le plut fort du terme. [...]

La modification selon le lieu se rencontre sous deux formes. La première est celle de la spécification d’un genre selon les communautés linguistiques, mais à l’intérieur d’une sphère culturelle historiquement plus ou moins solidaire. Le phénomène est trés répandu en Occident: nous parlons ainsi de l’épopée grecque et de l’épopée romaine, du roman français et du roman anglais [...].

Following Schaeffer’s explanations, generic terms referring to literary currents could be subsumed under the temporal context, and terms related to the linguistic and cultural identity under the spatial context. Schaeffer is aware that these aspects are not covered by his model, but also that most of the generic terms are not reducible to a single discursive level, neither regarding aspects of the verbal message (“the text”) nor contextual factors: L’existence de se modifications temporelles et spatiales des noms de genres pose la question de la contextualization historiques des déterminations génériques, question que le schéma communicationnel que j’ai retenu occulte [...] elle ne peut évidemment que renforcer la conclusion qu’imposait déjà la prise en compte de la multidimensionnalité du message verbal, à savoir que les noms de genres, loin de déterminer tous un même objet qui serait « le texte » ou même un ou plusieurs niveaux invariants de ce texte, sont liés, selon les noms, aux aspects les plus divers des faits discursifs. (Schaeffer 1983, 119)

This is not only true for the terms associated with the levels of theme and mode here, as could be seen in table 5 above, but also for the groups of current and identity. Generic terms referring to literary currents do not only localize the subgenre temporally and historically. They also entail preferences regarding the themes of the novels as well as stylistic properties. Similarly, the terms subsumed under identity do not only relate to the cultural, geographical, and linguistic localization of the novel as a discursive object but can also point to thematic aspects of its content when American, Cuban, Argentinian, or Mexican matters are treated. The assignment of terms to the levels thus sets focuses for their analysis but is not to be understood as exhaustive or exclusive. This is also the reason for the reduced modal level here compared to the models of Raible and Schaeffer. With the modal subtypes of intention, attitude, reality, medium, and representation, the model used here contains aspects of the communicative situation and the textual message that revealed themselves to be relevant for the nineteenth-century Spanish-American novels analyzed here because they are implied by explicit subgenre labels. Figure 10 situates the categories of subgenre labels used in the encoding model of the bibliography here in a more general communicative model mainly based on the one proposed by Schaeffer.

The figure shows that the subtypes of intention, attitude, reality, medium, and representation can be grouped under the aspect of how the literary text is communicated and presented (mode). The category of theme stands for what is communicated, and the catetories of identity and current point to the context in which something is communicated.In the figure, the levels of medium and syntactic are connected because, in the context of novels, generic terms that refer to a medium are usually to be understood as indications of how the text is represented structurally, e.g., in the form of letters in epistolary novels. Many of the labels relating to medial aspects are more vague, for example, those originally referring to painting and drawing: “cuadros”, “bocetos”, “esbozos”, and “impresiones” describe figuratively how the text is represented. On the one hand, the model used for the bibliography and the corpus here has to be understood as an application, adaptation, and selection of the general semiotic models of (literary) genres. On the other hand, it is a bottom-up approach. Only those aspects of the general models that occur in the generic signals and literary-historical assignments of genres to the works in the bibliography and corpus of nineteenth-century Cuban, Argentine, and Mexican novels were selected. It is thus an empirically driven discursive model for generic terms of the novel in a specific cultural-geographical and historical context. It would be interesting to see which levels of the general models are activated for other corpora of the novel in order to find out which kinds of subgenre labels are typical for the novel in general and which ones are determined by contextual factors. When one looks at the relevance of the different levels for the whole set of novels, it can be observed that the levels of theme and current are the two main levels used in literary-historical approaches. Themes are also frequently involved in explicit historical subgenre labels, but regarding the quantitative relevance, only the novela histórica and the novela de costumbres stick out. The level of identity is very present among the top most frequent explicit labels but is only indirectly discussed in the critical literature.Of course, there is research on the Cuban/Argentine/Mexican/Spanish-American novels, and also “national novels” (Sommer 1993) are discussed, but these approaches are usually not related to explicit generic terms. The level of mode.representation gains quantitative relevance because of the distinction between novels explicitly labeled as “novela” and those that are not. More specifically, there are also many “episodios” and “memorias”. The level of mode.reality plays an important role because many of the terms that are thematic also point to the relationship between text and reality (“novela histórica” and “leyenda”, for instance). The other modal categories (mode.intention, mode.attitude, mode.medium) are less important in terms of numbers but are also present. Intention and attitude play a role in terms that are also thematic, for example, in political novels. Terms related to the medium are creatively used by the authors in a range of different generic labels for the novels.See also chapter 4.1.5, where a series of charts that display the distribution of subgenre labels is included. Table 10 contains an alphabetically ordered list of all the different subgenre labels found for the novels in the bibliography and the corpus. In the table, it is indicated to which levels of the model the subgenre labels were assigned.The whole table is also available at https://github.com/cligs/data-nh/blob/master/corpus/bibliography-subgenre-labels/overview-subgenres.csv. Accessed March 28, 2020. Set of subgenres occurring explicitly or implicitly in the bibliography. Subgenre label Kind(s) Supplement Explicit occurrence apuntamientos mode.medium, mode.representation - yes apuntes mode.medium, mode.representation - yes auto-novela mode.reality - yes Bildungsroman theme - no boceto mode.medium, mode.representation - yes bosquejo mode.medium, mode.representation - yes capricho mode.representation - yes cinematógrafo mode.medium, mode.representation - yes comedia de carácter theme, mode.intention comedia yes confesiones theme, mode.representation - yes contornos mode.representation - yes croquis mode.medium, mode.representation - yes crónica theme, mode.representation novela histórica yes cuadros mode.medium, mode.representation novela de costumbres yes cuento mode.representation - yes drama mode.representation novela romántica yes elegía theme, mode.medium, mode.attitude - yes ensayo mode.representation - yes entretenimientos mode.intention - yes episodios mode.representation novela histórica yes epopeya theme, mode.representation novela histórica yes esbozos mode.representation, mode.medium - yes escenas mode.medium, mode.representation - yes estudio mode.representation, mode.intention novela social, novela realista, novela naturalista yes fragmentos mode.representation - yes historia mode.representation - yes impresiones mode.representation - yes juguete mode.intention - yes Künstlerroman theme - no lecturas mode.intention novela didáctica yes leyenda theme, mode.reality novela histórica, novela romántica yes medallones mode.representation - yes memorias mode.representation (theme) - yes narración mode.representation - yes notas mode.medium, mode.representation - yes novela mode.representation - yes novela abolicionista theme (mode.attitude) novela social yes novela americana identity (theme) - yes novela analítica mode.representation (mode.intention) - yes novela andaluza identity (theme) - yes novela anecdótica mode.representation - yes novela argentina identity (theme) - yes novela azteca identity (theme) (novela mexicana) yes novela biográfica theme - no novela bonaerense identity (theme) (novela argentina) yes novela camagüeyana identity (theme) (novela cubana) yes novela científica theme, mode.reality - (yes) novela clasicista current (theme) - no novela cómica mode.intention (mode.attitude) novela humorística (yes) novela contemporánea theme, mode.reality novela social and/or novela política yes novela corta mode.representation - yes novela criminal theme - (yes) novela criolla identity (theme) (novela americana) yes novela cubana identity (theme) - yes novela curiosa mode.intention - yes novela de actualidad theme, mode.reality novela contemporánea, novela social and/or novela política yes novela de aventuras theme - (yes) novela de costumbres theme (current) (novela social) yes novela de crímenes theme novela criminal yes novela de familia theme novela social no novela habanera identity (theme) (novela cubana) yes novela de horrores mode.intention - yes novela de la ciudad theme novela social no novela de misterio theme, mode.reality - no novela de propaganda theme, mode.intention novela política and/or novela social yes novela de Tabasco identity (theme) (novela mexicana) yes novela de viajes theme - (yes) novela didáctica theme, mode.intention novela social yes novela documentaria theme, mode.representation novela social and/or novela política no novela doméstica theme novela social yes novela en acción theme novela de aventuras yes novela enciclopédica theme, mode.intention novela didáctica yes novela epistolar mode.medium, mode.representation - yes novela espiritista theme, mode.reality novela científica yes novela fantástica theme, mode.reality - yes novela festiva mode.attitude - yes novela filosófica theme, mode.representation - yes novela franco-argentina identity (theme) (novela argentina) yes novela gauchesca theme - no novela histórica theme, mode.reality - yes novela humorística mode.intention (mode.attitude) - yes novela india identity (theme) - yes novela indigenista theme - no novela jurídica theme novela criminal yes novela kantabro-americana identity (theme) (novela americana) yes novela mexicana identity (theme) - yes novela militar theme novela histórica yes novela mixteca identity (theme) (novela mexicana) yes novela modernista current (theme) - no novela moralista theme, mode.intention novela social no novela nacional identity (theme) - yes novela naturalista current (theme) novela realista, novela social yes novela original identity - yes novela patriótica theme, identity - yes novela picaresca theme - no novela policial theme novela criminal yes novela política theme, mode.attitude - yes novela popular theme novela social yes novela porteña identity (theme) (novela argentina) yes novela psicológica theme, mode.representation - no novela realista current (theme) novela social yes novela regional theme, identity - yes novela romana identity (theme) - yes novela romántica current (theme) - no novela satírica mode.attitude - yes novela sentimental theme - yes novela siciliana identity (theme) - yes novela social theme (mode.intention) - yes novela suriana identity (theme) - yes novela tapatía identity (theme) (novela mexicana) yes novela verista current (theme) novela realista, novela social no novela yucateca identity (theme) (novela mexicana) yes panorama mode.medium, mode.representation - yes perfiles mode.representation - yes páginas mode.medium, mode.representation - yes recuerdos mode.representation, theme - yes reflexiones mode.representation novela filosófica yes relación mode.representation - yes relato mode.representation - yes reseña mode.attitude, mode.representation - yes romance theme, mode.representation novela sentimental, novela histórica yes silueta mode.representation - yes tradición theme, mode.reality - yes tragedia theme, mode.representation - yes

The generic terms consist of nouns (“Bildungsroman”, “croquis”, “episodios”) or nouns that are characterized further by attributes (“novela argentina”, “novela de costumbres”, “novela histórica”). One or several kinds of subgenre labels that are considered the most important are given for each term. Kinds in parentheses mean that these are also possible assignments but that they were not encoded in the bibliography because they were not considered crucial and in order not to mix the categories too much. In general, it is assumed here that generic terms are complex signs and that many of the terms refer to different levels of discourse. For example, the term “novela naturalista” refers to the literary current of Naturalism, but also to the themes preferred by that current, i.e., social topics, including the account of the situation of outsiders and lower classes, and tabooed subjects such as adultery or prostitution. Furthermore, it refers to certain representational techniques used in naturalist novels and so on. The generic terms are loaded semantically through the characteristics of the works that carry the terms over time. However, it is also assumed here that there are differences in how relevant the different levels are for the terms and that it is possible to determine primary levels. For example, the term “memorias” is assigned to the level of mode.representation here and only secondary to the level of theme. Even though memories can imply certain themes (a life story, for example), the themes are not very specific and the presentation of the text in the form of memories, looking back on what was experienced and is remembered, is more important. Other examples are the terms referring to literary currents, for instance, “novela realista” or “novela romántica”. Although they are also connected to certain themes, this aspect is considered subordinate here. The same was decided for terms referring to a certain city, region, nation, or people, such as “novela habanera”, “novela de Tabasco”, “novela mexicana”, or “novela azteca”, for which the level of identity is taken as the primary level and thematic aspects as secondary. That way these subgenres are differentiated from the ones that are primarily thematic, for example, “novela sentimental”, “novela histórica”, or “novela de costumbres”.

The third column in the table indicates subgenres that are implied by the term in the first column. For example, the term “crónica” often includes a historical theme, and the terms “novela naturalista” and “novela realista” often include social themes. The supplements can serve to normalize terms, to generalize them, and to make other levels that are implied by such terms explicit. The supplements are not automatically assigned, though. They are assigned depending on whether they make sense in the individual case. Some of the possible supplements are given in parentheses in the table. These are not assigned in the bibliography and the corpus to make the subgenres more distinguishable. The “novela de costumbre”, for example, could also be understood as a form of social novel but would then have the same thematic label as naturalistic and realist novels because the term “social” has so many facets.

The last column of the table indicates whether the generic term occurs explicitly in the bibliography or not. A “yes” in parentheses means that the term occurs, but only in a normalized explicit form and not verbatim. Examples of terms that never occur explicitly in the bibliography but are often used by critics to characterize the novels are “novela gauchesca”, “novela indigenista”, and “novela romántica”. Others that occur primarily in the explicit form are, for example, “novela jurídica” or “novela contemporánea”.

The goal of the systematization of subgenre labels by normalizing terms and assigning them to different levels of discourse is to put the whole range of generic terms associated with the novels through explicit or implicit paratextual elements and critical assignments in a certain order to be able to analyze groups of the novels quantitatively. Effectively, the set of generic terms with which novels are designated is open and endless, full of variants and nuances, as already the terms in the table show. If one would take the terms as they are, in many cases, it would be difficult to form groups. For instance, there are “novelas jurídicas”, “novelas policiales”, and “novelas de crímenes”, which are subsumed under the term “novela criminal” here because it is assumed that these terms can be interpreted as referring to the same subgenre. On the other hand, it would be challenging to analyze subgenres that are defined on entirely different levels of discourse together, for example, a “novela romántica” compared to a “novela mexicana” or a “novela fantástica”. The whole structure of subgenres is not organized hierarchically and it does not consider historical change.

To summarize, the assignment of subgenres to the novels discussed in this section follows a strategy that is, for one thing, historically oriented, because all the explicit labels occurring in the titles of the novels in the bibliography (and also in other paratexts for the novels in the corpus) are collected, and furthermore, also implicit signals are evaluated. Explicit signals are normalized in order to make them comparable, and the interpretation of implicit signals requires prior knowledge about the subgenres, but all the steps are documented to make the process transparent. In general, historically adequate terms are preferred over ahistorical critical ones. Then again, the information available from historical signals is complemented by subgenre assignments made by literary critics in order to also open up those novels to an analysis of subgenres that would otherwise have to be considered general fiction because there are no clear signals available from the paratexts. However, because the sources of the generic information are encoded in detail for each entry in the bibliography (and corpus), it is possible to conduct analyses only on one or the other kind of information or to backtrack the statements about subgenres when the results of a combined analysis are interpreted.

No previous selection of certain types of subgenres was made. Instead, within the frame of the general, formal working definition of the novel provided in chapter 3.1.1.7 above, all kinds of generic information were collected. The information was then systematized in summarizing terms, following a discursive model based on other general, semiotic models of generic names. That way, it was captured which (types of) subgenre labels are the most frequent in the bibliography. Regarding explicit labels, the distinction between works carrying the label “novela” and those that do not is quantitatively relevant. Regarding thematic aspects, “novelas históricas” and “novelas de costumbres” are frequent. Furthermore, there is a considerable group of novels that have a label related to the cultural and linguistic identity of the text (“novela original”, “novela mexicana”, etc.)  as opposed to a bigger group of novels that does not have such a generic signal. When also implicit signals and literary-historical labels are considered, the most frequent kinds of subgenre labels are thematic or refer to the literary current(s) of the texts. Even if other kinds of subgenre labels are not so frequent, for example, labels concerning the representational mode of the text (other than the general term “novela”) or labels pertaining to the groups of intention, attitude, reality, or medium, this generic information is still valuable as background information when other subgenres are analyzed, and when the results need to be interpreted. The great variety of subgenre labels found for the works in the bibliography shows how open the genre novela is in general and how extensive the network of generic references is. Even labels referring to other major genres (e.g., “drama”, “tragedia”, “comedia”, “epopeya”, etc.) are used to mark novels. On the other hand, only a few subgenres were very frequent in the period and countries examined in this study.

In general, in TEI, the metadata is encoded in the TEI header, which contains descriptive and declarative metadata associated with the digital resource. Of the five principal components that are available for the TEI header, four were used in the TEI model here: the file description, which contains bibliographic information, the encoding description, in which it is documented which kind of information was encoded when the digital file was created based on one or several other source files, the profile description, which includes information about non-bibliographic aspects of the texts, and the revision description, in which the revision history of the file is given.For general documentation of the TEI header, see the corresponding chapter in the TEI guidelines (Text Encoding Initiative Consortium 2023c). In the following, each of these parts is presented, using the TEI file of the novel “Adoración” (1894, CU) by Álvaro de la Iglesia as an example. Particular focuses are the declaration of rights for the corpus files in chapter 3.3.3.1.2 and the text classification with keywords in chapter 3.3.3.1.6.

Besides the TEI header, the second main part of each corpus file is the transcription and encoding of the novel in the text element. It is further subdivided into three parts: front, body, and back. While the body is present in all the TEI files of the corpus, the other two parts are optional. The front part may contain “any prefatory matter (headers, abstracts, title page, prefaces, dedications, etc.) found at the start of the document, before the main body” (Text Encoding Initiative Consortium 2023g). The back part can contain appendices of any kind (Text Encoding Initiative Consortium 2023e). In the corpus, the front part was used to encode title pages, dedications, and prefaces of available historical editions of the novels because they often provide information about the subgenres of the texts. Such front matters were included in 231 files of the corpus. For the other 25 novels, no historical editions could be accessed, so no front matter is available.Editions are considered “historical” here if they were published within the chronological scope of the bibliography and corpus (1830–1910). Front matters of modern editions were not transcribed. In example 17, an excerpt of the front matter for the novel “Adoración” is shown.

In the example, the front matter of one historical edition of the novel from 1906 is transcribed. It includes a title page, a short dedication, and a longer preface, of which only a part is shown here. In other cases, there may be several front matters. Each front matter is enclosed by a division element, indicating its source edition in the attribute source (source="#PS"). This attribute contains a reference to the edition’s bibliographic description in the source description in the TEI header. The year of the corresponding edition is encoded in the attribute n (n="1906") on the div element. Inside the main division for each front matter, its different parts are encoded in further subdivisions (e.g., div type="titlepage", div type="dedication", div type="preface"). Although the TEI offers specialized elements for the encoding of front matter, e.g., titlePage, byline, docImprint, etc. (Text Encoding Initiative Consortium 2023k), only the general elements div, ab, head, p, and seg are used here to keep the overall TEI model for the corpus simple and because there is no special interest in the semantics of the front matter structure here. Instead, the front matters are transcribed with the primary goal of interpreting their contents with regard to the subgenres of the novels.See the next chapter 3.3.4 on the assignment of subgenre labels to the novels in the corpus for details.

A back matter is included in 140 of the 256 TEI files. In general, back matters are less relevant for the subgenre assignment. In most cases, they only contain a phrase marking the end of the novel (“Fin”) or a dateline documenting where and when the novel was written (e.g., “Buenos Aires, Agosto 27 de 1858”). Only rarely notes or comments by the authors are appended, such as, for example the following remarks made by Ignacio Manuel Altamirano about the length of his novel “Clemencia” (1869, MX), as shown in example 18.

The encoding of the main body of the novel’s text is kept simple, as well. Above all, the markup is used to represent how a novel is structured into parts and chapters to be able to use this structural information in the analysis of the texts. Example 19 shows how the beginning of the novel “Adoración” is encoded.

In general, divisions are marked with the element div, using the attribute type to characterize the kind of division further into "part", "subpart", "chapter", or "subchapter". Headings and paragraphs are also encoded. In general, no difference is made between the main and subheadings. Only longer descriptions of the content of a following chapter are marked as head type="argument". Regarding the structure inside of the main textual divisions, it was decided to generally encode blocks separated by line breaks or blank lines with the element p, following a typographic definition of a paragraph. The only exceptions made are for verse lines, which are encoded with l, and dramatic speech, encoded with sp because these are considered important distinctions from the point of view of genre analysis. It follows from this that the content of a p element does not always correspond to the structural linguistic definition of a paragraph as a sequence of semantically related sentences or as a thematic building block of a written text.For a characterization of the paragraph from a linguistic point of view, see Rinas (2015). A ubiquitous phenomenon in the novels, for example, is blocks of direct speech. These are also marked with p and additionally with said, as explained further below.

The TEI standard includes many different elements for the encoding of text blocks, for example, the neutral element abThe anonymous block element is described as “any arbitrary component-level unit of text, acting as an anonymous container for phrase or inter level elements analogous to, but without the semantic baggage of, a paragraph” (Text Encoding Initiative Consortium 2023d). or special elements for structures like openers and closers in letters, for list or table entries, etc. In principle, such alternative elements would be a better choice to encode blocks in the novels that are not paragraphs in the linguistic-semantic sense. However, a detailed analysis of the text bodies would be required to identify such structures. In addition, more specialized markup would require advanced scripts for querying the XML structure. Furthermore, it can be estimated that non-paragraph blocks that are not verse lines, dramatic speech, or direct speech are few in number in the novels. The p element was therefore preferred here as a general solution for the markup of typographic blocks in the text body.

Besides this general structure of divisions, headings, and paragraphs, some more phenomena were encoded, as summarized in table 17. Encoding of textual phenomena in the main body of the novels. Type of phenomenon TEI element(s) used Typographically marked subdivisions of the text (e.g., with a line or asterisks) milestone with unit and rend Typographically highlighted words or phrases seg with rend Gaps gap Verse lines lg, l Dramatic text castItem, castList, sp, speaker, stage, said Representations of written text writing with type Quotations quote Direct speech or thought said Text contained in quotation marks that is not a representation of written text, not a quotation, and not direct speech or thought q Embedded texts interrupting the surrounding text floatingText

The encoding of the first two phenomena (typographically marked subdivisions of the text and typographically highlighted words or phrases) aims to preserve minor structural information that was already contained in a structured way in the editions used as sources for the corpus. These typographic details may be interesting when individual sections of texts are analyzed, but they can hardly be used for comparative analyses of all the texts in the corpus. They depend highly on how a specific source edition of a novel was typeset and also on how much of possibly existing typographic information in the sources was kept when editions were digitized. Gaps were encoded to get a quantitative overview of how much text is missing in the novels. The other phenomena that were encoded in the body focus on how the narrated text is presented in terms of genre (prose versus poetry versus drama), medium (written versus spoken), voice, and perspective (quotations, narration, and the representation of speech and thought). In the remainder of this subchapter, examples of the different phenomena that were encoded in the main body of the novels are given, with a special focus on the detection of direct speech and thought.

The encoding of the novels is controlled by a RELAX NG schema, which in turn is based on a more abstract ODD file.The TEI standard offers many different modules for encoding all types of texts. Normally, projects dealing with a specific type of text define a subset of the TEI’s modules to work with. Such a subset can be documented in a so-called ODD (“One document does it all”) file, a format that is used to express TEI customizations and that is independent of other formal schema languages. Schemas in different languages can be generated from the ODD in a second step. RELAX NG is one schema language that is suitable for XML (Murata 2014; Text Encoding Initiative Consortium n.d.b) The RELAX NG schema makes sure that the general TEI vocabulary and structure of the corpus are consistent. It is complemented by a Schematron file that serves to check the structure and content of the metadata in the TEI header in a more detailed way (see chapter 3.3.3.1.6 on text classification in the TEI header above). Links to both schema files are included as processing instructions in each of the corpus files, as shown in example 37.

The Schematron file (“keywords.sch”) is corpus specific and is therefore kept inside of the same repository.See https://github.com/cligs/conha19/blob/master/schema/keywords.sch on GitHub. Accessed August 19, 2020. The RELAX NG schema (“clings.rng”) and the underlying ODD file, on the other hand, are designed more generally for all the corpora developed in the CLiGS project and are therefore stored in a separate repository called “reference”.See https://github.com/cligs/reference. Accessed August 20, 2020.

The CLiGS TEI schema includes elements that are basic for the encoding of literary narrative, dramatic, and poetic texts. However, it avoids other specialized block-level and inline elements. Its definition was kept as restrictive as possible, only allowing for elements and attributes that are actually in use in the different corpora. Compared to other established TEI customizations such as TEI Lite, TEI Simple, or the DTA-Basisformat (DTABf), the CLiGS schema is more restrictive, although it is not an exact subset of either of them. On the other hand, a few attributes have been added to the schema in the project-specific namespace “https://cligs.hypotheses.org/ns/cligs” (Schöch et al. 2019, paras 14–18). In the main TEI corpus files, the only additional attribute is cligs:importance, used to assign degrees of importance to metadata category values, for example, of different subgenre assignments.See the examples in chapter 3.2.3 above. Second, several custom attributes have been added to the schema to hold linguistic annotations produced with the NLP package FreeLing.The output of the FreeLing tagger is integrated into the TEI structure of the main corpus files, and the results are stored as a derivative TEI format, as explained further in chapter 3.3.5 below.

A Python script is used to test the validity of the corpus files against the RELAX NG schema. It produces a log file reporting the success or failure of the validation process for each TEI file.The Python file for validation is accessible at https://github.com/cligs/scripts-nh/blob/master/corpus/metadata_encoding/validate_tei.py. The resulting log file can be viewed at https://github.com/cligs/conha19/blob/master/schema/log-rng.txt. Accessed August 20, 2020. Validating the TEI files against the Schematron file requires a different strategy. In principle, Schematron validation is possible with the Python module “lxml”, but only if the queries used in the Schematron file conform to the XSLT 1.0 standard (Behnel 2022). To check that the metadata keywords in the TEI header conform to the keyword taxonomy, however, it was necessary to also use XSLT 2.0 expressions in the Schematron file. An alternative way for validation without Python is to compile the Schematron file as XSLT and apply this transformation script to all the TEI files in the corpus using Saxon directly from the command line. The error output of this transformation process is stored in a log file.Here, the compilation of the Schematron file as XSLT was done with the software Oxygen by choosing the preconfigured Transformation Scenario “ISO Schematron to XSLT (compile)”. The resulting XSLT file can be viewed at https://github.com/cligs/conha19/blob/master/schema/keywords-compiled.xsl. The subsequent command line call for saxon is: java -jar /home/ulrike/Programme/saxon/saxon9he.jar -s:/home/ulrike/Git/conha19/tei/ -o:/home/ulrike/Git/conha19/tei-checked/ -xsl:/home/ulrike/Git/conha19/schema/keywords-compiled.xsl > /home/ulrike/Git/conha19/schema/log-schematron.txt. The resulting Schematron log file is accessible at https://github.com/cligs/conha19/blob/master/schema/log-schematron.txt. If the file is empty, now errors were found. Accessed August 20, 2020.

To summarize this chapter on the encoding of the text corpus, it can be said that this thesis focuses on the encoding of detailed metadata about the novels rather than on a very detailed encoding of the texts themselves. This is due to the kind of resource that the corpus is intended to be: it is an edited text collection aimed to serve as the basis for quantitative genre analysis where metadata about the authors, source editions, the form and content of the texts and, above all, about the subgenres that the novels have been assigned to plays an important role. For the encoding of the textual body, a special emphasis was put on the markup of direct speech in a subset of the novels. In the next section, the assignment of subgenre labels to the novels, which was bypassed in this general chapter about metadata and text encoding, is set out in more detail.

829 works are registered in the bibliography, of which 256 (31 %) are contained in the corpus. The previous chapter discussed how many works were published per author. In this chapter, the first focus is on the number of works published over time, using the publication years of the first known editions of the works.See the figures 107 to 109 in the appendix of figures.

First, it is analyzed how many works were published per year between 1830 and 1910. The first work in the bibliography is from the year 1832, and in the corpus, there are two works first published in 1839. The last works in both Bib-ACMé and Conha19 were published in 1910. Apart from the 1830s, when only a few works were published, almost all the years are covered in the bibliography. Exceptions are the years 1849, 1852, and 1867. As the numbers were generally low in the 1840s and the early 1850s, it is possible that no novels at all were published in 1849 and 1852. In the year 1867, however, it is surprising. It may be the case that the political situation in the three countries made it difficult for authors to write or publish novels in that year. During the time, Argentina was involved in the War of the Triple Alliance, Cuba stood at the beginning of a period of internal wars, and in Mexico, the emperor Maximilian was overthrown by liberal troops. However, verifying that this had an effect on the number of novels that were published would require more research into the personal circumstances of the authors and the history of the publishing sector. Other years that are not represented in the corpus are some of the years in the 1830s, 1845, 1850, 1853, 1863, and 1878. Apart from the 1830s, for which it is more difficult to access the few novels that were published, this is interpreted as the effect of random selection. The number of works published increased considerably towards the end of the century. From 1880 on, at least ten works were published per year.

Summarizing the values for decades, the coverage of works in the corpus in relationship to the bibliography becomes clearer.In figure 108 in the appendix of figures, the percentages indicate the proportion of works contained in the corpus for each decade. From the 1850s to the 1890s, the share of works in the corpus is about one-third, which corresponds to the overall proportion of works included in the corpus. The 1860s are slightly overrepresented with 39 %. In the margins, i.e., the early and late decades, the numbers deviate more. The 1830s, 1900s, and 1910s are underrepresented in the corpus, and the 1840s are strongly overrepresented. Apart from the 1900s, such deviations are more likely in these decades because the overall number of works is much lower than in the central decades.

Summarizing even more and comparing the period before 1880 to the period in and after that year,See figure 109 in the appendix of figures. this results in a better representation of the earlier period. 37 % of the works in the bibliography that were published before 1880 are also contained in the corpus. In the later period, 28 % of the works in the bibliography are also part of the corpus. This means that the corpus contains proportionally fewer works in the period after 1880, although in total, more works were published in the later decades of the nineteenth century.

Another perspective on the number of works over time is obtained by differentiating by country.See figure 110 in the appendix of figures. When the development of the number of works published over the decades is observed that way, different patterns become visible. In Argentina, the number of works exploded in the 1880s. According to the bibliography, 90 works were published in that decade, compared to around 20 works each in the three decades before. Apart from the 1830s, for which only one work is included in the bibliography, all the decades are also represented in the corpus. For Argentina, of the central decades, the 1870s are overrepresented and the 1900s underrepresented. In Mexico, the number of works published rose earlier and not so sharply. Considerably more works were published from the 1860s onwards, and their number increased towards the end of the century, whereas in Argentina, the number decreased again after the 1880s. Regarding the decades with many works, the 1860s are overrepresented for Mexico in the corpus. As for Argentina, no work from the 1830s is included in the corpus. In contrast to Argentina and Mexico, the number of Cuban works published between 1830 and 1910 does not show significant growth. Actually, most works that are included in the bibliography were published in the 1850s (28 works), followed by the 1890s (21 works). Compared to Argentina and Mexico, more Cuban works were published in the early decades, from the 1830s to the 1850s. It is very probable that this different development of the number of published novels in the course of the nineteenth century is due to Cuba’s status as a colony, which only ceased in 1898, and which prevented the growth of the literature marked. On the other hand, Cuba was colonized early and had a close connection to its motherland Spain, which could explain why relatively more novels were published in the early decades by Cuban-Spanish than by Argentine and Mexican authors. However, the quality of the bibliographic sources used can also play a role, as discussed above in chapter 3.2.1. Comparing the overall number of works by country, most novels in Bib-ACMé and also in Conha19 are Mexican, followed closely by Argentine novels. The Cuban novels make up the smallest part, with 16 and 19 %, respectively.See figure 111 in the appendix of figures.

Looking not at which countries the novels are generally associated with in the bibliography and the corpus but in which countries they were first published, the role of Spain becomes visible: 7 % of the novels in the bibliography and 9 % of the novels in the corpus were first published in that country.See figure 112 in the appendix of figures. Apart from the lowest numbers, there is no difference between the bibliography and the corpus concerning the ranks of the countries where the novels’ first editions were published. Most novels were first published in Mexico, followed by Argentina, Cuba, Spain, France, and the USA. Comparing the numbers of the publication places to the general numbers by country, it becomes clear that not only part of the Cuban novels were first published in Spain, but also Mexican and Argentine novels.

Comparing the works in Bib-ACMé and Conha19 revealed that both are proportionately quite congruent when the distribution of works over time and the share of works by country are considered. Nevertheless, on a level of detail, also some differences became visible. In the corpus, especially the 1860s are overrepresented, and the 1900s are underrepresented. As a result, the period before 1880 is covered to a higher degree in the corpus than the one after this year. Regarding the countries, there are relatively more Argentine and Cuban and fewer Mexican works in the corpus than in the bibliography, but these differences range only between 2 to 6 %. As to the overall distribution of works over time, almost all the years between 1830 and 1910 are covered in the bibliography and the corpus, with the exception of some early years from the 1830s to the early 1850s, plus the exceptional year 1867, in which no works were published. Especially from the 1880s onwards, the number of works published increased considerably. However, this development is not the same in all three countries. In Argentina, most works were published in the 1880s; in Mexico, the number of works grew already in the 1860s; and in Cuba, no significant growth over time can be recognized at all.

Besides the metadata that can be evaluated for both the bibliography and the corpus, some informative aspects about the novels are only available for the corpus. They depend on more specific metadata that has only been collected for Conha19 or on the full texts of the novels that are only available in the corpus. Such aspects are analyzed in this chapter. Part of the metadata that was gathered for the corpus refers to technical and administrative aspects, such as the type of source medium, the kind of source edition, and the institution that held the source. Summaries of these data were already given in chapter 3.3.1 (“Selection of Novels and Sources”) above and are not discussed here any further.

One metadata item that was only collected for the novels in the corpus is their status as high- or low-prestige novels.See chapter 3.3.3.1.6 (“Text Classification with Keywords”) above for details on how prestige was modeled. In Conha19, 174 novels (68 %) are classified as high prestige and 82 novels (32 %) as low prestige. There is no difference in the proportion of high- and low-prestige novels from Cuba, but from Mexico there are more high-prestige novels, and from Argentina more low-prestige ones.See figure 113 in the appendix of figures for an overview of the proportions of high- and low-prestige novels by country. There are several probable reasons for this. Surely, the quality of the bibliographic sources used as a basis for selecting novels for the corpus is an influencing factor. The bibliography of the Argentine novel authored by Lichtblau is very comprehensive and also includes many authors and works that are not well-known. Furthermore, the state of digitization and access to digital sources plays a role. The collection of Argentine novels published on Wikimedia Commons by the Argentina Academy also contains many works written by lesser-known authors. Independently of the reasons, the corpus has a certain bias towards low-prestige Argentine and high-prestige Mexican novels.

The analysis is now deepened by considering the distribution of novels by prestige over time.See figures 114 (by decade) and 115 (before versus in or after 1880) in the appendix of figures. Over the decades, most low-prestige novels were included in the 1890s, 1880s, and 1860s. On the other hand, low-prestige novels are underrepresented in the 1840s, 1870s, and the 1900s. The decades 1830 and 1910 are not really informative because the number of works in them is so small. It makes sense that in the decades in which the overall production of novels increased considerably, more low-prestige novels were produced and were also selected for the corpus. Regarding the 1900s, they are, in general, underrepresented in Conha19The mean proportion of novels in the corpus compared to the bibliography is 31 %, and the 1900s are only represented with 20 %. See the overviews in the previous chapter., so the probability of selecting high-prestige works is higher. The 1840s, in contrast, are generally represented quite well in Conha19Comparing the corpus to the bibliography, in the 1840s, 54 % of the works are covered., so it can be assumed that there were not many works in that decade that are considered low-prestige today, or if there were, they are not known. Why low-prestige works are underrepresented in the 1870s is not clear. Summarized to the two periods before and in or after 1880, the proportion of low-prestige works is higher in and after 1880.

Another metadata item that is only available for the corpus is the narrative perspective of the novels. In general, there are 44 novels (17 %) with a first-person narrator and 212 novels (83 %) with a third-person narrator, so the latter clearly prevails. Regarding the distribution of the narrative perspective by country, it is interesting that the proportion of Cuban novels is much lower for the novels narrated in the first person than for those narrated in the third person. A hypothesis is that the individual, personal perspective was not so suitable for novels published in the colony. The first-person novels are mainly Argentine, closely followed by the Mexican novels, and the third-person novels are above all Mexican.See figure 116 in the appendix of figures.

When analyzed over timeSee figures 117 and 118 in the appendix of figures., it becomes visible that the proportion of novels written in the first person was highest in the 1870s, 1890s, and 1880s. No first-person novel is included from the 1850s, and also the 1840s and 1900s are mostly represented with novels narrated in the third person. The drop of first-person novels in the 1900s is surprising because otherwise, they became more frequent after the middle of the nineteenth century. Again, this decade is generally underrepresented in the corpus, which might be a reason. Comparing the period before 1880 to the one in and after 1880 shows that narrations in the first person are relatively and absolutely more frequent in the latter period, although the difference between the two periods only amounts to 6 %.

Further information that was collected for the novels in the corpus is the continent and country of the setting. 9 % of the novels are primarily set in Europe, and just one novel is set on another continentSee figure 119 in the appendix of figures. The novel with another setting is the science fiction novel “Viaje maravilloso del Señor Nic-Nac” (1875, AR) by Eduardo Holmberg, which tells an imagined trip of the protagonist to the planet Mars., so the great majority of 90 % is set in America. Looking at the country of the setting, Mexico, Argentina, and Cuba are most frequent, corresponding to the countries of origin of the novels. Other places the novels are set in are the European countries Spain, Italy, France, Greece, Switzerland, the USA, and other South-American countries (Peru, Chile, Bolivia, Brazil, and Uruguay). Together, American countries other than Mexico, Argentina, and Cuba make up for 4 % of the cases.

An evident question is if the preference for a European or American setting was influenced by the country of origin of the novels, i.e., if there is a difference between the Argentine, Mexican, and Cuban novels in this aspect. Analysis of the metadata shows that this is indeed the case.See figure 120 in the appendix of figures. While the proportions of Mexican, Argentine, and Cuban novels set in America correspond largely to the general significance of these countries in the corpus, the numbers are quite different for the novels set in Europe. Here, the majority is Cuban and the minority Mexican, suggesting that Cuba’s status as a colony during most of the nineteenth century had an influence on the setting of the novels. In addition, the relationship between Argentina and Europe was closer than that between Mexico and Europe as regards the choice of setting for the fictional texts. However, in absolute numbers, only 24 novels are set in Europe, so these trends should also not be overinterpreted. Following the proportions of works set in Europe over the decades and comparing their share in the period before 1880 and in or after that year makes clear that the number of works set in Europe decreased over time. Moreover, at least in the corpus, the 1860s were already a decade in which an American setting was clearly preferred.See figures 121 and 122 in the appendix of figures.

Besides the continent and country of the setting, also the time period covered by the novels is registered in the metadata. Three time periods are distinguished: contemporary, recent past, and past. For each novel, two values were encoded: the time period relative to the author’s birth year and relative to the year of the first known publication of the novel.See chapter 3.3.3.1.6 (“Text Classification With Keywords”) above for details. The proportions of works set in the different time periods are visualized in figure 123 in the appendix of figures. From both points of view, a contemporary setting is the most frequent one: related to the authors’ birth years in 82 % and relative to the publication date in 73 % of the cases. Novels set in the past are the second most frequent group, with 13 % and 16 %, respectively. The recent past is treated in 4 % of the novels when the authors’ birth years are concerned and in 11 % of the works when the publication year is decisive. The differences between the two approaches show that more novels treat a period that is past in relation to the publication date but still part of the contemporary experience of the authors. Fewer novels treat a period that lies in a more distant past. In the following, only the approach of comparing the publication date to the time period covered by the novels is considered further.

How do the proportions of the three time periods covered in the novels relate to the three different countries that the novels are associated with in the corpus?See figure 124 in the appendix of figures. As can be seen, the Argentine and Mexican novels cover most of the contemporary perspective. As the general proportions of works by country were 42 % Mexican works, 39 % Argentine works, and 19 % Cuban works,See figure 111 on the works by country in the appendix of figures, which was discussed in previous chapter. a contemporary setting is a bit overrepresented in the Argentine novels and slightly underrepresented in the Mexican and Cuban novels. A setting in the past is overrepresented in Mexican and Cuban novels and underrepresented in Argentine novels. Here the differences range between 3 and 6 % of the novels. The preferences are most striking concerning a setting in the recent past. Here, Mexican novels are overrepresented by 8 % and Cuban novels by 10 %, while Argentine novels are underrepresented by 18 %. All in all, the past is more a topic in the Mexican and Cuban novels, and the recent past is relatively most important in the latter ones, while the contemporary perspective is favored in the Argentine novels. This might be explained by the fact that the colonial history of Cuba and Mexico is longer than that of the Argentine region. In the case of Cuba, another factor is the difficulty of broaching contemporary issues in a country that was still under the control of the mother country, which may have led to a preference for representing the recent past. On the other side, Argentine society and economy developed rapidly in the nineteenth century, which supplied much material for the novels treating the contemporary period.

Did the preference for setting the novels in a certain time period change in the course of the nineteenth century? An analysis of the distribution of time periods per decade and a comparison of the period before 1880 to the one after that year suggests that there is no clear chronological trend but that there are some intermittent preferences instead.See figures 125 and 126 in the appendix of figures. The contemporary period was always dominating as a setting for the novels. Interestingly, the past was favored more in the 1860s and then again in the 1900s. Leaving out the first and last decades with very few works, also the recent past reaches the top positions in these two decades. The lowest proportions of novels treating the past and recent past can be seen in the 1880s and 1890s. So a first trend of treating historical issues came up shortly after independence was reached in most Spanish-American countries, probably as a way of contributing to writing their own history in literary terms. Then, in the decades of significant social and economic development, contemporary issues were more relevant. A return to a greater interest in the recent and further past at the beginning of the new century marks a new phase.The occupation with events that belong to the distant or recent past or are contemporary is not necessarily to be equated with the novels being historical novels or not. As Read states in his study of the Mexican historical novel: “It will be apparent as this study progresses that the works involved fall naturally into two groups, the romantic historical novels that deal with the conquest period and colonial times, and the novels that deal with historical events of the nineteenth century. The first group is essentially romantic, corresponding to the type developed by Walter Scott but with a distinctly local ‘middle ages’. Instead of turning to medieval Europe for exotic material, Mexican writers of this type of fiction sought out characters and institutions of their own dim past. Their hostility to the Spanish regime was still fresh enough to inspire them with a feeling of spiritual kinship with the Amerinds who had been the traditional enemies of the Europeans. [...] In this same group of romantic historical novels belong those fictional works that deal with colonial times. The Inquisition [...] is naturally the center of interest of these poetic interpretations of life in the colony. The second class of Mexican historical novels is that which finds its material in the history of the nineteenth century, the epoch in which the writers themselves had been actors in the dramas they presented. Such works may properly be called novels of contemporary history. Many of them were patterned after the Episodios nacionales of Pérez Galdós and the various historical romances of Erckmann-Chatrian. But though these two groups of works deal with materials from widely separated periods, they have much in common. Whatever the period involved, it was interpreted in terms of the ideals of the nineteenth century when Mexico was attempting to constitute itself a new nation [...]. Patriotism, a new sense of national identity and zeal for liberty and justice were the emotive forces that determined the trend of interpretation in both groups of historical novels to which attention has been called” (Read 1939, ix–xi). Both types of novels also existed in Argentina: historical novels in the strict sense, which broached the issues of the Conquest and colonial times, and novels that treated contemporary historical events or those of the very recent past. Many of the early novels of the latter kind had the Rosas regime as their subject. Molina subsumes them under the group of political novels and calls them “novelas prospectivamente históricas” (Molina 2011, 246–249, 285–311). A contemporary setting was also predominant in the realist and naturalistic novels of the later nineteenth century: “With reference to Argentina, it is significant that the development of the realistic novel should coincide with the extraordinary growth and progress which the Republic manifested during the years 1880 to 1900. As this economic and material transformation took place, greatly affecting every facet of the nation’s life [...] eager writers sought to mirror that rapid change and portray the new society that was surging forth” (Lichtblau 1959, 138). Regarding the Argentine naturalistic novel, Lichtblau remarks: “Not only did Argentina produce the first naturalistic novelist in Hispanic America in the figure of Cambaceres, but that country displayed as well the greatest over-all development of the naturalistic current in the nineteenth century. The tremendous material advancement, the great influx of immigrants, the changing social pattern, and the growing industrialization of the Republic–all these things writers used to advantage in applying the tenets of Zola to Argentine fiction.” (176–177). Condensing this development to the phase before and after 1880 results in relatively more novels treating the contemporary period in the latter and some form of the past in the former.

A characteristic of the novels in the corpus that goes beyond metadata is the length of the texts. In the context of the definition of boundaries of the novel, the minimum length of novels was discussed in detail in chapter 3.1.1.4 above. Therefore, regarding the novels that were included in the corpus, the question remains how long these actually are in terms of the number of tokens.This is summarized in the box plot in figure 127 in the appendix of figures. The shortest novel in the corpus has about 16,000 tokens, the longest one has about 331,000 tokens, and the median length is approximately 53,000 tokens.The numbers are rounded to the next thousand. The shortest novel is “Gubi Amaya” (1865, AR) by Juana Manuela Gorriti, and the longest one is the historical novel “El mendigo de San Ángel” (1865, MX) by Niceto de Zamacois. Surprisingly, the shortest and the longest novel of the corpus were first published in the same year. Despite the lower limit being 16,000 tokens, two novels with approximately 15,800 tokens, among them “Gubi Amaya”, were included because the number of tokens changed in the course of the preparation of the corpus, amongst other things, due to the correction of spelling errors. So in a strict sense, these two novels only fulfill the minimum length criterion when the number of tokens is rounded to the next thousand. It is interesting that the medium length of the Spanish-American novels in this corpus corresponds almost to the minimum length for novels set by Forster to 50,000 words (Forster 2016, 17), so the nineteenth-century Argentine, Mexican, and Cuban novels, as defined here, tend to be shorter than the typical English novel that Forster had in mind. 25 % of the novels are between 16,000 and 35,000 tokens long, the next 25 % between 35,000 and 53,000 tokens, the third quarter is between 53,000 and 96,000 tokens, and the last one between 96,000 and 331,000 tokens, so the spread of lengths increases considerably for the upper 50 % of the novels and the longest novels are clearly outliers.

Analyzed by country, the distribution of lengths is very similar for the Argentine and Cuban novels but different for the Mexican novels.See figure 128 in the appendix of figures. Again, in the following, the numbers are rounded to the next thousand. The median Argentine novel is 48,000 tokens long, and the median Cuban novel has 50,000 tokens, so they are both shorter than the overall median novel. The longest Argentine novel has 231,000 tokens, and the longest Cuban novel has 198,000 tokens. Compared to that, Mexican novels are longer. The medium Mexican novel is 67,000 tokens long, and the three longest novels with over 300,000 tokens are also Mexican. The three longest novels are historical novels, so it should be examined if there is a correlation between the length of the novels and their subgenre, which is done for thematic subgenres and literary currents in chapter 4.1.5.3 below. Testing for statistical significance, it turns out that the difference in length between the Mexican and the Argentine, as well as between the Mexican and the Cuban novels, is indeed significant.The script used for the significance tests is available at https://github.com/cligs/scripts-nh/blob/master/analysis/sign.py. Accessed January 3, 2021. Because the data is not normally distributed, the Mann-Whitney U test was used (instead of a t-test, for instance). The p-value that was calculated for the text lengths of Mexican versus Argentine novels is 0.001, for Mexican versus Cuban novels 0.04, and for Argentine versus Cuban novels 0.2, which means that there is no significant difference in the latter case.

How does the novels’ length develop over the decades?See figure 129 in the appendix of figures. First, in the decades 1830 to 1850, the median length drops from 110,000 to 36,000 tokens. In the 1860s, it jumps to 87,000 tokens, and after that, it raises from 47,000 in the 1870s to 99,000 in the 1910s. The works in the 1830s and 1910s are very few, though. Regarding the median, it is especially interesting to see the exceptional decade of the 1860s. As was found out in the evaluation of the time periods of the novels’ settings, in this decade, representations of the past were relatively favored. In addition, they were preferred in Mexican novels, and these were more numerous in the 1860s than Argentine and Cuban novels.See figure 110 (“Works by decade and country”) in the appendix of figures. Together with the observation that the longest novels of the corpus are historical novels, this might explain why there were more long novels in this decade than in the others. A test for statistical significance reveals that the text lengths can be considered significantly different in the following constellations of decades: 1860s versus 1870s, 1860s versus 1880s, and 1880s versus 1900s.The script used for the calculation of significances and variance ratios is available at https://github.com/cligs/scripts-nh/blob/master/analysis/sign.py. Accessed January 3, 2021. The data is not normally distributed, so the Mann-Whitney U test was used. The following p-values resulted: 0.02 for 1860s versus 1870s, 0.003 for 1860s versus 1880s, and 0.045 for 1880s versus 1900s. The other constellations had p-values above 0.05. The 1830s and 1910s were not included in the calculations because there are only 2 and 3 works for these decades, respectively. It is also noteworthy to see that the variability of the texts’ length (in terms of the spread of length in the two central quartiles) is lower from the 1870s to the 1890s than before and after that decade. In the last three decades of the nineteenth century, very long novels are the exception.Calculating the ratio of text length variances for different pairs of decades shows that the differences in variance are biggest for the 1860s versus 1890s, the 1860s versus  1880s, and the 1860s versus 1890s. The difference in variance can be considered significant for the 1850s versus 1880s-1900s,  the 1840s versus 1880s-1900s, and for all the constellations of the 1860s versus later decades. The differences in text length variance between the 1870s and the 1880s as well as the 1870s and 1890s are also significant, but the remaining ones are not. Variance ratios between 0.5 and 2.0 are considered similar, and values below 0.5 or above 2.0 as significantly different. The three biggest ratios of variance are 5.4 for the 1860s versus 1890s, 4.8 for the 1860s versus 1880s, and 4.0 for the 1850s versus 1890s. The 1830s and 1920s were not included in the calculations because of the low number of works in these two decades.

In the bibliography, 622 novels (i.e., 75 % of all the novels) carry an explicit (sub)generic signal of any kind, while 207 novels do not carry any explicit signal at all. In the corpus, 204 novels (80 % of all the novels in the corpus) have an explicit signal. The explicit label “novela” is carried by 404 (49 %) of the novels in the bibliography and by 134 (52 %) of the novels in the corpus. How can this be interpreted? Either the novel, as defined here, is a genre that is so self-evident that its representatives do not need the explicit denomination to be recognized, or it is so vaguely defined that as many other texts are covered by it. However, as information about almost all the works in Bib-ACMé and Conha19 was retrieved from relevant bibliographies and literary histories of the novel, the former aspect is more plausible.In a future analysis, it could be interesting to analyze if the presence or absence of the label “novela” corresponds to different subtypes of the genre. In what follows, the proportions of works in the bibliography with and without the explicit label “novela” are analyzed by decade.See the corresponding figure 136, in which a series of donut charts is given, in the appendix of figures. Up to the 1870s, more than half of the works in the bibliography carry the label “novela”. From the 1880s on, this label becomes rarer, suggesting a change in the conventions of labeling the works over time.The analysis is based on the information if the works first published in the respective decades ever carried the label “novela” between 1830 and 1910 because the works are dated according to their first known edition, but their subgenre labels are collected for all the editions that were published in the chronological frame of this study. This introduces a certain fuzziness concerning the anchoring of subgenre labels in time, so the effect of change might even be stronger. However, both types of works, those with and without the explicit label “novela”, were present in all the decades.

In total, 108 different explicit subgenre labels are found in the bibliography. Although these labels are called “explicit” here, they do not correspond exactly to the historical denominations used to mark the novels because the values were normalized in order to be comparable. Part of this normalization is that compound labels were split up, and each part was marked up separately.See chapter 3.2.3 above for details. The top 20 of these regularized explicit labels in the bibliography are analyzed here and compared to the corresponding labels in the corpus.See figure 137 in the appendix of figures. The top 20 positions were calculated from the bibliography’s point of view. The general label “novela” is the most frequent one. The other top explicit labels are of different kinds. Some labels are directly related to the themes of the novels and are recognizable as subgenres of the novel: “novela histórica” (on rank 2 in the bibliography), “novela de costumbres” (rank 5), “novela social” (rank 13), and “novela policial” (rank 17). Labels referring to the linguistic and cultural identity of the novels also recur in this top list: “novela original” (rank 3), “novela mexicana” (rank 4), “novela cubana” (rank 9), “novela nacional” (rank 11), “novela argentina” (rank 15), and “novela americana” (rank 16). Of the remaining labels, several are (not exclusively, but often) related to different kinds of historical novels: “episodios” (rank 6), “memorias” (rank 7), “leyenda” (rank 8), and “historia” (rank 12). The labels “escenas” (rank 18) and “cuadros” (rank 19) are often connected to novels of customs. Interestingly, also labels designating other genres, such as “drama” (rank 10), “cuento” (rank 14), and “ensayo” (rank 20), are among the top labels for the novel.

For the corpus, the ranks of explicit labels are similar, but there are also a few differences.See figure 138 in the appendix of figures. Labels that are in the top 20 for the corpus but not for the bibliography are “estudio” (rank 15), “novela realista” (rank 18), “crónica” (rank 19), and “novela militar” (rank 20). On the other hand, labels that are in the bibliography top 20 list but not in the corresponding corpus list are “novela nacional”, “novela policial”, “escenas”, and “ensayo”. When the first ranks of the bibliography and corpus lists are compared, differences are, for example, that the number of “novelas históricas” is almost equal to the “novelas de costumbres” in the corpus, whereas it is around twice as high in the bibliography. Furthermore, “novela cubana” is on rank 6 in the corpus compared to rank 9 in the bibliography.

Analyzing the top explicit subgenre labels brings to light several characteristics of the novels in the bibliography and the corpus. First, the most prominent explicitly marked subgenre is the historical novel, both according to the number of occurrences of the literal label and also based on several other subgenre labels related to it, which leads to the conclusion that the most prominent subgenres are historical novels. Second, there was an evident need to explicitly mark the linguistic, cultural, or national identity of the novels. As there are so many different kinds of identity labels, it is of interest to check how many novels carried such labels. In the bibliography, 272 novels (33 %) had an identity label, and in the corpus, 100 novels (39 %). As with the general label “novela”, also here the question arises if the use of identity labels depends on the period of publication of the novels.See figure 139 in the appendix of figures. A trend becomes visible over the decades, as the number of novels carrying identity labels decreases continuously. A third point that can be drawn from the top frequent explicit labels is which subgenres are more important in the corpus than in the bibliography. These are the novels of customs, realist, and naturalistic novels linked to the labels “novela realista” and “estudio” and also Cuban novels.

If implicit signals are included in the evaluation, the range of subgenres broadens because some subgenres are never marked explicitly. This also means that the assignment of subgenre labels gets more interpretive. For 511 works (62 %) in the bibliography and 207 works (81 %) in the corpus, implicit signals were found.That more implicit subgenre signals were found for the novels in the corpus also results from the fact that more paratextual information was evaluated for them (title pages, prefaces, etc.). If this is added to the explicit information, subgenre signals were recognized for 738 novels (89 %) in the bibliography and 254 novels (99 %) in the corpus. In the following, the top 20 subgenre labels for Bib-ACMé and Conha19 are analzyed, taking explicit and implicit signals into account together.See the figures 140 and 141 in the appendix of figures.

When also implicit signals are included, more subgenres related to the themes of the novels and to literary currents enter the top positions: in the bibliography, rank 2 is still occupied by the “novela histórica” followed by the primarily thematic labels “novela sentimental” (rank 3), “novela social” (rank 6), “novela de costumbres” (rank 7), “novela política” (rank 14), “novela criminal” (rank 17), and the “novela de la ciudad” (rank 20). Labels relating to literary currents are “novela romántica” (rank 4), “novela naturalista” (rank 12), and “novela realista” (rank 15). This shift could be expected because the assessment of implicit signals requires an interpretation frame, and subgenres focusing on theme and literary currents are the dominant perspectives in literary histories.

Comparing the top 20 signals of Bib-ACMé to Conha19, again, some subgenres gain more weight in the corpus: the “novela sentimental” moves from rank 3 to 2, the “novela de costumbres” from rank 7 to 4, the “novela gauchesca” enters the top list on rank 20, the “novela naturalista” from rank 12 to 8, and the “novela realista” from rank 15 to rank 10. The “novela romántica”, on the other hand, gets less important and moves from rank 4 to 6, the “novela histórica” switches from rank 2 to rank 3, and the “novela de la ciudad” is not part of the top list anymore.

When only statements made by literary historians are evaluated, the picture changes even more. Literary-historical assignments were recorded for 433 works (52 %) in the bibliography and 224 works (88 %) in the corpus. That the proportion of works with literary-historical labels is much higher in the corpus is certainly because the corpus contains more works that are better known and researched. There are 34 different literary-historical labels in the bibliography and 32 different ones in the corpus. As with the explicit paratextual signals, also here, the labels were homogenized to be comparable and do not correspond to literal statements in every case.

In the bibliography, the label most often assigned is “novela romántica”, followed by “novela histórica” and “novela realista”.See figures 142 and 143, which display the top 20 subgenre labels assigned to the works by literary historians. Labels that were not included in the top ranks of generic signals, but are top literary-historical labels, are “novela indigenista”, “novela abolicionista”, “novela modernista”, “novela de aventuras”, “novela verista”, “crónica”, “novela científica”, and “novela satírica”. Here, different critical perspectives on nineteenth-century Spanish-American novels are introduced, for example, specific topics and socio-cultural concerns (“novela indigenista”, “novela abolicionista”) or particular literary currents (“novela modernista”, “novela verista”), but also general generic subcategories that are not specified culturally (“novela de aventuras”, “crónica”, “novela científica”, “novela satírica”).

The top ranks of literary-historical labels in the corpus are of a similar kind but in part differently ordered. The “novela romántica” is also most important in the corpus. Labels with more weight than in the overall bibliography are “novela social”, “novela de costumbres”, “novela naturalista”, and “novela abolicionista”, for instance, and labels that are less relevant in the corpus are, for example, “novela histórica”, “novela realista”, “novela criminal”, and “novela gauchesca”. The different top ranks illustrate where there have been shifts in the composition by subgenre due to the selection of works for the corpus. Some were made on purpose, for example, the inclusion of more novels of customs as a counterbalance to the great majority of historical novels or a preferred inclusion of Cuban novels to strengthen the smallest country group. Others depend on the availability of the novels. Many crime and gaucho novels can be classified as low-prestige novels and are not yet readily available in digital format.

The differences between explicit subgenre signals, implicit signals, and literary-historical subgenre assignments underline that the views on what a subgenre of the novel is differ considerably, depending on the practice and the purpose of the labeling. There are only some intersections. The labels assigned to the historical editions probably served a number of functions: to clearly mark a novel as an instance of a known and popular subgenre of the time (e.g., “novela histórica”, “novela de costumbres”), to associate a novel with a certain literary current or school by using terms that were loaded with such semantics (e.g., “estudio”, “cuadros”), to signal that the language and content of a novel are autochthonous (e.g., “novela original”, “novela americana”), or to give the novel other attributes inscribing it in the general history of genres, playing with terms, or attracting attention towards its particular way of (re)presentation (e.g., “leyenda”, “episodios”, “memorias”, “drama”), among other possibilities. Literary-historical assignments of subgenre labels, on the other hand, usually have the primary function of classifying the works in question according to established critical schemes. They are much more uniform and concentrate on aspects of style and content that allow for connecting the novels with subgenres that are rich in tradition (e.g., “novela histórica”, “novela sentimental”, “novela de aventuras”), that are associated with certain literary periods and currents (“novela romántica”, “novela realista”, etc.), or that are of special critical interest regarding the novelistic production of a certain cultural-historical space (“novela gauchesca”, “novela indigenista”, “novela abolicionista”). Only in part do these two practices of labeling individual novels with (sub)generic terms overlap. The top lists of subgenre labels show that the labeling is more systematic in the literary-historical approach: there are not just a few thematically oriented labels but a whole range of them, and not just one or two labels relating to literary currents but labels for all the currents that have been recognized as relevant in the nineteenth century. The historical practice, on the other hand, is much more selective. “Novelas históricas” and “novelas de costumbres” were commonly explicitly named, but “novelas sentimentales”, for example, only exceptionally. In addition, there is historical evidence in the bibliography for the labels “novela realista” and “novela naturalista”, but not for the “novela romántica”. This does not prove that there was no awareness of the latter, but apparently, some types of novels were more part of a general, given generic pool than others.

Even though literary-historical approaches to subgenres aim to systematize the field, in sum, the resulting set of labels is still a conglomerate of different perspectives on the novels, even if to a lesser degree than the historical labels. This becomes very clear when the same novel is labeled with several terms at the same time. The works of the Mexican writer Victoriano Salado Álvarez, for example, have been classified both as historical and realist novels (Fernández-Arias Campoamor 1952, 84–85; Read 1939, 293–294) and some works by the Cuban Cirilo Villaverde both as novels of customs and romantic novels (Remos y Rubio 1935, 166–180; Suárez-Murias 1963, 23–24). This makes it difficult for a stylistic analysis of subgenres that aims to select and compare subsets of novels from a corpus. Therefore, the explicit, implicit, and literary-historical labels have been sorted according to a system of discursive categories, as explained above in chapter 3.2.3, resulting in several sets of labels that belong to different discursive levels but whose comparison is more meaningful on each level. It has to be reminded, though, that this system is artificial. In what follows, overviews of the different sets of subgenre labels based on the discursive model of subgenre terms are given for the bibliography and the corpus. First, summaries of how many labels there are on which levels are presented.See figures 144 and 145 in the appendix of figures. In figure 144, the number of different labels related to the realization of the discursive act is smaller than the sum of labels related to its three subgroups (“semantic”, “syntactic”, and “medium”) because the same label can be associated with several levels. For example, the labels “novela filosófica” and “novela psicológica” are categorized both as realization/semantic/theme and as realization/syntactic/mode.representation, because the terms point both to certain themes (e.g., general considerations about the meaning of life in a philosophical novel or the focus on personal, emotional states of characters in a psychological novel) and also to a certain way of representation (e.g., an argumentative style in philosophical novels and an introspective narrative style in psychological novels).

How many different labels on the various discursive levels are there in the whole bibliography? In total, there are 124 different terms. Most of them belong to the thematic group (39 %), followed by the mode the novel is represented in linguistically or narratively (36 %), the cultural-geographical and linguistic identity of the novel (20 %), the medium that the novel uses (11 %), the intention of the author or narrator (10 %), the relationship between the novel and reality (8 %), the literary current of the novel (5 %), and the attitude the author or narrator has towards what is represented in the novel (4 %). The diversity of thematic labels is not surprising, and that there is only a small set of labels related to literary currents is expectable as the number of different currents is limited. The broad range of labels referring to the mode of representation and also to the identity was not expected, though, as these aspects are usually not focused on in studies of subgenres of the Spanish-American novel.

How many different subgenre labels are there in Conha19? In the corpus, there are 90 different subgenre labels, which are distributed similarly over the different levels when compared to Bib-ACMé.The visualization for the corpus is available at https://github.com/cligs/data-nh/blob/master/corpus/overview/subgenres-labels-number-corpus.html. Accessed 15 September 2020. 44 % of the different labels are thematic, 36 % are related to the mode of representation, 14 % to the identity, 13 % to the medium, 10 % to the intention, 9 % to the mode of reality, 7 % to the literary current, and 3 % to the attitude.

The significance of the different discursive levels changes to a certain degree if not the number of different labels in each category is considered, but the overall number of labels belonging to them. How often have such labels been assigned?See figure 145 in the appendix of figures. When the overall number of labels was determined for the various categories, identical labels stemming from different kinds of sources were only counted once for each novel (e.g., if a novel was explicitly labeled as “novela histórica” and also classified as such by literary historians). In total, on the different levels, 3,193 labels were assigned to the novels in the bibliography and 1,317 to the novels in the corpus.Again, on each level, if the same label is assigned to a work by different sources, it is only counted once. Most labels are of the thematic type (38 %), followed by the mode of representation (23 %), literary currents (15 %), the mode of reality (10 %), identity (10 %), the attitude (2 %), medium (2 %), and intention (1 %). So in terms of quantity (instead of diversity), labels related to literary currents and the reality mode are more important, while labels associated with the mode of representation, with identity, the medium, intention, and attitude are less relevant. The picture is similar for the corpus, only that literary currents have a bit more weight than modes of representation.The visualization for the overall number of labels in the different categories in the corpus is available at https://github.com/cligs/data-nh/blob/master/corpus/overview/subgenres-labels-amount-corpus.html. Accessed 15 September 2020. In Conha19, 40 % are thematic labels, 20 % are related to literary currents, 18 % to the mode of representation, 8 % to the identity, 8 % to the mode of reality, 2 % to the attitude, 2 % to the medium, and 1 % to the intention. Clearly, thematic labels are most important both when the number of kinds and the overall number of assignments is considered.

The impression that the overall top lists of labels give is also confirmed when one differentiates between different sources of labels. All in all, there are 108 different explicit subgenre labels in the bibliography and 34 different literary-historical labels. The corpus includes 70 different explicit labels and 32 different literary-historical labels. This means that the diversity of explicit labels doubles the diversity of literary-historical labels in the corpus and is three times higher in the bibliography. Regarding the overall number of labels, 1,669 explicit labels and 1,120 literary-historical labels were assigned to the works in the bibliography, and 564 explicit and 686 literary-historical labels were assigned to the works in the corpus. So on average, a literary-historical label is assigned more often than an explicit label.To determine the overall number of labels, they were counted on each discursive level, so a homonymic label on different levels is counted several times. On the other hand, if several literary-historical sources mentioned the same label for a work, it was only counted once per level. The following table 23 summarizes the importance of the different discursive levels for explicit versus literary-historical labels in the bibliography.See the corresponding visualizations “subgenres-labels-number-explicit-bib.html”, “subgenres-labels-amount-explicit-bib.html”, “subgenres-labels-number-litHist-bib.html”, “subgenres-labels-amount-litHist-bib.html” at https://github.com/cligs/data-nh/tree/master/corpus/overview. Accessed 15 September 2020. Ranks of discursive levels of subgenre labels, explicit versus literary-historical (Bib-ACMé). Rank Different explicit labels Amount explicit labels Different literary-historical labels Amount literary-historical labels 1 mode.representation 43 mode.representation 707 theme 25 theme 558 2 theme 37 theme 363 current 6 current 385 3 identity 25 identity 302 mode.representation 5 mode.reality 121 4 mode.medium 14 mode.reality 203 mode.reality 4 mode.attitude 35 5 mode.intention 13 mode.medium 48 mode.intention 2 mode.representation 10 6 mode.reality 9 mode.intention 28 mode.attitude 2 mode.intention 7 7 mode.attitude 4 current 11 mode.medium, identity 1 mode.medium, identity 2 8 current 2 mode.attitude 7 - - - -

When differentiating by the type of source (explicit versus literary-historical labels), thematic labels are also generally important. Regarding the overall number of labels, the mode of reality also has some importance for both explicit and literary-historical labels. The relevance of the other levels depends more on the provenance of the labels. For explicit labels, the mode of representation and identity are prominent, and for literary-historical labels, in particular, the literary currents. The ranks do not change for the corpus, apart from minor shifts in the last positions.For the corpus, see the charts “subgenres-labels-number-explicit-corp.html”, “subgenres-labels-amount-explicit-corp.html”, “subgenres-labels-number-litHist-corp.html”, “subgenres-labels-amount-litHist-corp.html” at https://github.com/cligs/data-nh/tree/master/corpus/overview. Accessed 15 September 2020. Regarding the importance of the linguistic and cultural-geographical identity of the novels, it has to be said that this aspect is, of course, also a topic in the critical discourse about the novels, but in a different way than in the historical practice of labeling the works. For literary-historical studies, usually, the linguistic and cultural-geographical frame is set from the beginning on. Either it concentrates on one national space, e.g., “the Mexican novel”, or on Spanish America or Latin America as a whole and then differentiating the novels by country. However, usually, no difference is made between one work or the other and the attribution to the cultural-geographical space is made based on general extra-textual parameters. On the other hand, in literary histories, works are also reviewed in terms of how their content reflects local or foreign realities, but usually in more general terms and without explicitly categorizing novels in that way. Furthermore, the question of the own and the foreign is discussed more in aesthetic categories.See, for instance, examples of statements on the relationship to the European literatures and questions of emancipation in Rössner’s literary history of Latin America. About nineteenth-century Caribbean literature, he writes: “Das Streben nach Unabhängigkeit ist nicht nur eine politische Angelegenheit, es bestimmt auch das literarische Leben. Einerseits orientieren sich die karibischen Literaten des 19. Jhs. an europäischen, besonders spanischen und französischen Vorbildern [...], andererseits bemühen sie sich darum, diese Modelle nicht einfach zu kopieren, sondern sich deren theoretisches Gedankengut mit Berücksichtigung all der Spezifika ihres amerikanischen Lebensraums kreativ anzuverwandeln” (Rössner 2007, 153). In this context, there are also discussions of individual authors and works: “José López Portillo y Rojas hingegen distanziert sich von französischen Vorbildern. In dem Vorwort zu seinem Roman La parcela (1898) weist er die ästhetische Wortkunst Flauberts oder der Brüder Goncourt ebenso zurück, wie er auch die Obszönität Zolas meiden möchte. Stattdessen bezieht er sich auf die Spanier Galdós und vor allem Pereda, deren Vorliebe für das naturverbundene Leben in der Provinz auch in Portillos Bauernroman ihren thematischen Niederschlag findet. Die Reserve dem französischen Kulturerbe gegenüber hat sich auch in dem Roman Fuertes y débiles (1919) erhalten. Das hier porträtiert porfiristische Gesellschaftssystem krankt daran, dass sich der französische Positivismus nicht einfach auf die mexikanischen Verhältnisse übertragen lässt und an den espíritus débiles der científicos scheitert” (Rössner 2007, 148).

In the following, the two discursive levels of subgenre labels that were chosen for text analysis are analyzed further on the metadata level: primary thematic subgenres and primary literary currents. In the text analysis part, the setups for thematic labels are further reduced by concentrating on the three most frequent subgenres (novela histórica, novela sentimental, and novela de costumbres), but in the following section, also the less frequent ones are examined in terms of metadata.

Several sets of general features were prepared. They can be distinguished by the unit that is counted, by the number of most frequent items taken into account, and by the kind of normalization applied to the resulting counts. The values that were selected on these three levels are summarized in table 26. Parameters for general feature sets. Level Selected values token unit basic units words, word 2-grams, word 3-grams, word 4-grams, character 3-grams, character 4-grams, character 5-grams character n-gram subtypes “all”, “word” (mid-word, multi-word), “affix-punct” (prefix, end-punct) frequency range (MFW) 100, 200, 300, 400, 500, 1000, 2000, 3000, 4000, 5000 normalization tf, tf-idf, z-scores

Regarding the token units, words are the classical option for a “bag” representation of the texts, which is needed as a basis for most machine learning algorithms (Müller and Guido 2016, 330; Scikit-learn developers 2007–2023d). For a bag-of-words model, the text is tokenized into individual units, which are then counted. As a result, a corpus is represented by the token counts for each document. In such a model, the order of the units does not play a role anymore. By using word n-grams, i.e., sequences of n words, as the basic unit, the relevance of word order can be reintroduced into a bag-of-words model to a certain extent. Word 2- to 4-grams were chosen to test different ranges of word sequences. While word 2-grams are most frequent in the whole corpus, 3- and 4-grams are closer to whole phrases and are more special features that might be relevant for the distinction of subgenres. Diminishing the unit to character sequences, on the other hand, has several other advantages. First, it makes the model less dependent on the exact spelling of the words. In a word-based approach, a difference in spelling results in different features. In a character-based approach, the words are split into several sequences anyhow so that the effect of orthographic differences in specific parts of the words is smaller. For the corpus at hand, this is of interest because some spelling errors persist as a result of the text digitization process.See chapter 3.3.2  on text treatment. Second, character n-grams that include blank space and punctuation marks cover aspects of the text string that are not captured by the classical bag-of-words approach, where words are usually split at blank spaces and punctuation marks are ignored. Third, character n-grams can be modeled so that they cover specific linguistic substructures of a morphological, thematic, or stylistic nature. Here, both general character n-grams (called “all” in the above table) and specifically modeled ones (called “word” and “affix-punct”) are used. How the different types of character n-grams are generated is illustrated with an example sentence given in example 49.The sentence is taken from the novel “Lucía Miranda” (1860, AR) by Eduarda Mansilla de García.

The general approach, which is implemented, for example, in the Python library scikit-learn (Scikit-learn developers 2007–2023f), uses all types of characters in the string and yields the following 3-grams for the example sentence: “«Es”, “Esp”, “spe”, “per”, “ero”, “ro_”, “o_q”, “_qu”, “que”, “ue_”, “e_a”, “_al”, “alg”, “lgú”, “gún”, “ún_”, “n_d”, “_dí”, “día”, “ía,”, “a,_”, “,_v”, “_vo”, “vos”, etc.Underscores are used to represent blank spaces to enhance the readability of the n-grams. As Sapkota et al. state, such general character n-grams are very effective for text classification, for example, by author. However, because of the mixed nature of the n-gram types, it is not clear why they are so useful: “Character n-grams are the single most successful feature in authorship attribution [...], but the reason for their success is not well understood. One hypothesis is that character n-grams carry a little bit of everything: lexical content, syntactic content, and even style by means of punctuation and white spaces [...]. While this argument seems plausible, it falls short of a rigorous explanation” (Sapkota et al. 2015, 93). The authors of the cited paper test this hypothesis by designing different types of character n-grams and by using each type separately in authorship attribution tasks to see which kinds of n-grams are successful. They use three types of super categories: n-grams covering morpho-syntactic features (affix-like n-grams, called “affix”), thematic content (word-like n-grams, called “word”), and style (punctuation-based n-grams, called “punct”). The n-gram types associated with the three super categories are called “prefix”, “suffix”, “space-prefix”, and “space-suffix” for the affix group, “whole-word”, “mid-word”, and “multi-word” for the word group, and “beg-punct”, “mid-punct”, and “end-punct” for the punctuation-based group.For details on how these are implemented, see Sapkota et al. (2015, 94–95). Sapkota et al. (2015, 96–98) found out that the affix-like n-grams are most successful in their single-domain authorship attribution task. For cross-domain authorship attribution, also the punctuation-based features are very strong.With “single-domain”, they mean that the corpus consists of texts about a single topic written by different authors. “Cross-domain” means a corpus with multiple different topics.

The subtypes of n-grams that are modeled here are based on the propositions of Sapkota et al., but because the issue is subgenre classification and not authorship attribution, the selection of n-gram subtypes is led by hypotheses about the features’ relevance for this categorization task. Here, the subtypes “mid-word” and “multi-word” are selected and combined into a word-based character n-gram feature set (called “word”). It aims to cover aspects of the thematic content that is supposed to be different from subgenre to subgenre. Furthermore, the subtypes “prefix” and “end-punct” are used in a set combining morpho-syntactic and stylistic characteristics (called “affix-punct”). Its goal is to test whether the differences between the subgenres may also be captured with features that are not primarily content-based.“Whole-word” was not selected because the general bag-of-words approach already covers this feature type. Of the affix-like features, “space-prefix” and “space-suffix” were not selected because they are congruent with “prefix” and “suffix”, the only difference being that they are one character shorter because they start or end with a blank space. “Suffix” was not selected because, in the Spanish language, the suffixes are highly influenced by how pronouns are used (if they are used freely before or attached to verb forms after them). As has been shown in the chapter on text treatment above (see 3.3.2), the difference in pronoun use is connected to the type of edition and its publication year, so it cannot be reliably associated with the subgenres. Of the three punctuation-based n-gram types, only “end-punct” was used because it is the only one capturing phrase-level structures of the sentences. The main reason for not only using general character n-grams is that they are dominated by blank spaces and the short most frequent words in different variants if only a certain number of top most frequent items is used (see table 28 below illustrating the resulting top most frequent tokens for different types of the general features). The intuition is that these features can hardly be interpreted in terms of the different subgenres. Regarding the chosen range of character numbers, it was decided to use 3-, 4-, and 5-grams because 3-grams are the minimum that is needed to be able to construct the different character n-gram subtypes, and 4-grams have been successfully used in subgenre classification before (Hettinger et al. 2016). The 5-grams are included as an alternative candidate. Examples of the n-gram subtypes are listed in table 27, using the above-mentioned example sentence for their creation. Definition and examples of character n-gram subtypes. Group Subtype Subtype definition Examples (for 3-grams) word mid-word “a character n-gram that covers n characters of a word that is at least n + 2 characters long, and that covers neither the first nor the last character of the word”The source of this and the other subtype definitions in the table is Sapkota et al. 2015, 94f. spe  per  lgú  mar  aré  réi  amb  mbi  bié  osa  lan  anc  nca word multi-word “n-grams that span multiple words, identified by the presence of a space in the middle of the n-gram” o_q  e_a  n_d  s_a  s_t  n_m  s_r  s_b affix-punct prefix “a character n-gram that covers the first n characters of a word that is at least n + 1 characters long” esp  alg  ama  tam  ros  bla affix-punct end-punct “a character n-gram whose last character is punctuation, but middle characters are not” ía,  as»

The examples highlight the differences between the character n-gram subtypes. The most evident one is that the number of resulting n-grams varies. The general approach using all types of character n-grams for the example sentence results in 60 3-gram tokens. In contrast, there are only 13 mid-word, eight multi-word, six prefix, and two end-punct tokens, so depending on the rules set up for the character n-gram subtypes, the feature space is reduced considerably because only some types of words or cross-word sequences meet the conditions.

Besides the different token units, the frequency ranges were chosen so as to cover stylistic features in the narrower sense with only a small number of most frequent words but also a mix of function and content words with several thousand most frequent words. Three types of normalization were chosen for the token counts: term frequency (tf), term frequency – inverse document frequency (tf-idf), and z-scores. The raw, absolute counts are computed as a basis for the normalization steps but are not directly used in the categorization because they depend heavily on the length of the texts, which varies considerably.See the overviews on text length in chapter 4.1.3.2. The term frequency is a relative frequency balancing out the effects that different text lengths have. It is calculated by dividing the absolute number of occurrences of a term t in document d by the frequency of the term t’, which is the term with the maximum frequency in d (see formula 1).

For tf-idf, the tf-scores are weighted by the inverse document frequency, which is defined as the logarithm of the number of all documents divided by the number of documents that contain the term t. This normalization takes into account the structure of the whole corpus: terms that occur in many documents of the corpus are considered less important than terns that only occur in a few documents. This gives the function words, which are likely to occur in all documents, a weaker weight compared to the more specific content words. Here, the tf-idf-scores were calculated with the Python library scikit-learn, which uses a specific implementation of this weighting scheme (see formulas 2 and 3).The 1 that is added to the idf secures that terms that occur in all documents will not be ignored (Scikit-learn developers 2007–2023g).

The third kind of normalization used, the z-scores, also involve the relationship between the different documents in the corpus, but in a different way than the tf-idf-scores. A z-score measures the number of population standard deviations by which the value of a score is above or below the population mean. This results in positive values for scores that are above the mean, zero values for scores that are equal to the mean, and negative values for scores that are below the mean. The population values are calculated by determining the standard deviation and mean for each feature in the corpus (see formula 4).

Like tf-idf-scores, z-scores upgrade the importance of terms that are not so frequent relative to the ones that are very frequent, but in the case of z-scores, it is not decisive in how many of the documents in the corpus they occur. In addition, the range of the resulting values is different than for tf and tf-idf because z-score zero values indicate that a term frequency is equal to the population mean but not that the term in question has a frequency of zero in the document. Instead, terms that do not occur at all in documents get negative z-score values. This different treatment of zero values can have consequences when the texts are categorized.

Furthermore, z-scores themselves do not take into account the effects of different text lengths by computing relative values per document as in tf and tf-idf. By applying the z-score transformation to absolute values, text length effects would only be controlled indirectly and only to a certain extent because the spread of the absolute values in the corpus is factored in through the standard deviation. A term with a higher standard deviation is penalized by the denominator in the formula if it is above the mean and revalued if it is below the mean. However, very long texts would tend to get higher z-score values. In addition, it would not be clear if higher z-score values are due to text length or to above-average uses of a term. It does therefore make sense to apply the z-score-normalization on top of tf-values, which is done here. By considering several normalization strategies in the feature sets, these can be compared to see which effect they have on the kind of features that turn out to be relevant for the categorization. While tf and tf-idf-scores are generally used in document classification, z-scores have especially been used in the context of authorship attribution (Burrows 2002; Evert et al. 2017; Müller and Guido 2016, 338–340). In sum, the combinations of parameters for the general features listed above result in 390 different MFW-based feature sets.The common abbreviation “MFW” is used in the general sense of “most frequent items” here and also includes the most frequent n-grams. The features were generated with a Python script available at https://github.com/cligs/scripts-nh/blob/master/features/general_features.py. The resulting feature sets are available at https://github.com/cligs/data-nh/tree/master/analysis/features/mfw. Both links were accessed on November 4, 2020.

The bag-of-words representations of the novels were created on the basis of full-text files that were extracted from the linguistically annotated TEI files.See chapter 3.3.5 for details about the linguistic annotation. The annotated TEI files were used in their corrected form and are available at https://github.com/cligs/conha19/tree/master/annotated_corr and the corresponding full-text files at https://github.com/cligs/conha19/tree/master/txt_annotated_corr. Accessed December 14, 2020. This was done to be able to remove tokens that were annotated as proper nouns, such as person and place names, from the files. This preprocessing step is considered important because especially the names of the protagonists can have distorting effects on the lists of MFW, as they occur very often. By taking out these named entities, of course, conclusions of the form “the female name ‘Blanca’ is overrepresented in sentimental novels” or “the place name ‘Madrid’ is mentioned above-average in Cuban novels” cannot be drawn anymore. Instead, the features are intended to lead to insights into more general linguistic-stylistic characteristics of the texts. In addition to the replacement of proper nouns, a list of stop words was created, consisting of the lists of proper names and place names that were compiled for the spell check of the text files.See chapter 3.3.2 about the text treatment. The resulting stop word list is available at https://github.com/cligs/data-nh/blob/master/analysis/features/stopwords/mfw_stopwords.txt. Accessed November 4, 2020. The stop word list is used by scikit-learn’s CountVectorizer (Scikit-learn developers 2007–2023f) in the process of creating the bag-of-words models that are based on word units.Per default, the CountVectorizer uses the token pattern (?u)\b\w\w+\b, which looks for sequences of word characters separated by word boundaries. For the Spanish texts in the corpus, the pattern was slightly modified to also cover words consisting of just one character, such as “y” or “a”: (?u)\b\w+\b. The stop word list was applied to the full texts in a customized manner before the character n-gram features were created.The special subtypes of character n-gram units were created with a Python approach designed specifically for this purpose (see the link to the script in footnote 518) because the CountVectorizer only supports the general character n-grams that include all types of characters. The combination of entity removal and stop word list has two advantages. A stop word list alone would not work to detect proper names that also have a general meaning in the language, and there are a lot of such names in Spanish, for example, “Blanca” (white), “Clemencia” (mercy), “Rosa” (rose), “Gloria” (fame), “Hidalgo” (nobleman), “Salvador” (saviour), “Gil” (silly), “Cortés” (polite). On the other hand, the stop word list can compensate for some errors of the named entity recognition.

Several visualizations were created to characterize the resulting feature sets.For this purpose, also the script https://github.com/cligs/scripts-nh/blob/master/features/general_features.py was used. The resulting plots are available at https://github.com/cligs/data-nh/tree/master/analysis/features/mfw/overviews. Accessed November 4, 2020. It is, for example, useful to know if a feature set is sparse, i.e., if many of the features have mostly zero values or if almost all the features are present in all instances of the corpus. It is also of interest to know how much the feature values vary. Both these characteristics have an influence on how well specific categorization methods work and if further steps of preprocessing are needed before the feature sets can be used for categorization.Random forest classifiers, for example, tend not to work very well with high-dimensional sparse data. Support Vector Machines (SVM), on the other hand, usually handle these well but work better when features have similar scales (Müller and Guido 2016, 90, 103–106). Figure 36 shows how the overall proportion of zero values increases from 100 to 5,000 MFW from less than 1 % to one-third of all the values.This chart only shows the feature sets based on single words. Charts displaying the proportions of zero values in the feature sets based on other token units (word ngrams and character ngrams) are available on GitHub. See footnote 523 above.

Looking in more detail into how many features have how many zero values in figures 37 and 38, it can be seen that the zero values in MFW100 stem from a few features only, of which one has over 40 zero values. 94 % of the features are never zero. In contrast, in MFW5000, only 7 % of the features are never zero. The form of the histogram shows that there are many features that are zero in about half of the texts. For the frequencies between 100 and 5,000 MFW, the overall proportion of zero values lies between these extremes, and the bulk of features has zero values in an increasing part of the corpus.

Character n-grams are much more stable regarding the proportion of zero values than words. For example, there are no zero values for the 100 most frequent character 5-grams. For the 5000 most frequent ones, the amount increases only to 1 %. For word n-grams, on the other hand, the proportions of zero values increase abruptly from the 3-grams on. The 100 most frequent word 2-grams, for instance, have less than 1 % of zero values. In contrast, the 100 most frequent word 3-grams have 11 %, and the word 4-grams 43 %. The 5000 most frequent word 2-grams have 32 % of zero values, the word 3-grams 57 %, and the word 4-grams 82%.For details, see the overviews at https://github.com/cligs/data-nh/tree/master/analysis/features/mfw/overviews. Accessed November 4, 2020.

In the following figures 39 to 42, it is illustrated for MFW1000 how the different normalization strategies reduce the variance of the feature values. For absolute values, the variances range between 6 and 2,976, for tf-scores between 0.001 and 0.150, for tf-idf-scores between 0.0006 and 0.0581, and for z-scores between 0.999999999999998 and 1.000000000000002, so the absolute values of variance are smallest for the tf- and tf-idf-scores, but the range of variance is smallest for the z-scores.The last number or decimal of the values was rounded. In addition, the distribution of the variances changes from a right-skewed distribution for the tf- and tf-idf-scores to a left-skewed distribution for the z-score values, for which the variance is expressed in relative terms and, therefore, more balanced for the different frequency ranges of the features. With z-scores, most features vary to a similar degree, and only a few features are more consistent throughout the texts in the corpus. A hypothesis of the effect that these differently normalized features have on classification tasks is that less frequent features will have a higher probability of being selected as important features if the variances of the features are more balanced.See, for instance, the remarks made on the scikit-learn website on how different feature variances influence the selection of coefficients in linear models (Scikit-learn developers 2007–2023c).

The top 10 features for different frequency sectors and token units are listed in table 28 to get a sense of the kind of features contained in the matrices of the most frequent items. Most frequent tokens. Frequency sector (MFW) Token unit Tokens 1–10 word de  el  la  que  y  a  en  se  los  no 101–110 word esto  habían  allí  hay  desde  todas  amor  vez  toda  mucho 501–510 word alto  última  voluntad  pensar  dirigió  delante_de  alegría  siquiera  además  pasos 1001–1010 word daban  no_obstante  basta  tranquilo  hablando  vivo  mozo  batalla  cólera  poco_a_poco 4091–5000 word injusticia  apoyada  sensaciones  bendito  cuadra  insignificante  personalmente  distinción  significa  fijo 1–10 word 2-gram de_el  a_el  de_la  a_la  en_el  en_la  de_los  de_su  que_se  lo_que 101–110 word 2-gram como_si  si_no  todas_las  por_su  pero_no  sobre_el  y_le  por_los  a_una  la_joven 501–510 word 2-gram calle_de  dijo_que  como_los  una_palabra  era_de  sobre_las  mi_madre  sobre_los  el_primero  que_tiene 1001–1010 word 2-gram de_otro  tenía_que  en_toda  el_instante  sus_hijos  con_lo  ver_lo  hombres_de  preguntó_el  dueño_de 4091–5000 word 2-gram esos_hombres  eso_y  mañana_de  sin_querer  en_ti  qué_había  dando_se  sus_soldados  paz_y  amigo_el 1–10 char 4-gram _de_  _el_  que_  _que  _la_  _en_  _con  _se_  o_de  a_de 101–110 char 4-gram ada_  _cua  más_  as_d  a_ca  a_y_  o_co  ue_l  _más  e_qu 501–510 char 4-gram es._  as._  ar_e  _cam  adre  ía_d  ue_c  _ni_  no_p  nte, 1001–1010 char 4-gram o,_q  s,_a  _tem  an_c  s,_l  olvi  s,_q  uego  aber  ber_ 4091–5000 char 4-gram neci  pitá  siti  rmas  ría,  lia_  rici  no_f  _obr  vend 1–10 char 4-gram (word) o_de  a_de  s_de  e_la  de_l  e_el  de_e  a_el  ente  ando 101–110 char 4-gram (word) el_a  ones  enta  n_qu  os_e  ue_n  ntra  o_nha  ar_s  mbre 501–510 char 4-gram (word) nera  ar_d  otra  ra_p  o_re  s_ve  es_l  s_el  ende  echa 1001–1010 char 4-gram (word) nqui  dich  ía_q  casi  sibl  feli  l_mi  alid  uant  dici 4091–5000 char 4-gram (word) rcan  masi  idam  clas  ifes  mbia  pica  e_be  mó_e  uili 1–10 char 4-gram (affix-punct) ente  ando  para  ment  esta  ento  ante  como  habí  cont 101–110 char 4-gram (affix-punct) ismo  hast  pera  ensa  algu  acio  ales  hora  prim  vida 501–510 char 4-gram (affix-punct) grad  inad  cura  cult  rinc  ingu  vers  engo  camp  oras 1001–1010 char 4-gram (affix-punct) sueñ  dorm  fere  ritu  alud  denc  creí  pend  saca  lara 4091–5000 char 4-gram (affix-punct) dudo  line  aza.  nel,  mutu  sinu  genu  oso;  iro,  decr

For the word unit, the frequency range 1–10 contains exclusively function words (propositions,  conjunctions, articles, and pronouns) and the particle “no”. In the range 101–110, the nouns “amor” and “vez” enter the list, and there are several verbs and adverbs with general meanings. Range 501–510 contains more specific nouns (“voluntad”, “alegría”, “pasos”), semantically more specific verb forms (“pensar”, “dirigió”), and adjectives (“alto”, “última”). On the positions 1001–1010 and 4091–5000, there are nouns with quite specific meanings (“mozo”, “batalla”, “cólera”, “injusticia”, “sensaciones”, “cuadra”, “distinción”), verbs with more different tenses (gerund: “hablando”, imperfect: “daban”, participle: “apoyada”), and also adjectives and adverbs with increasingly specific meanings (“tranquilo”, “bendito”, “insignificante” and “no_obstante”, “poco_a_poco”, “personalmente”). The word lists from the different frequency sectors show how the feature space gets more complex in terms of the grammatical forms of words that are included and also regarding the semantic specificity of the words as more most frequent items are included.

In the word 2-grams, function words dominate the most frequent tokens in all frequency ranges. On positions 1–10, the 2-grams consist entirely of function words. In the range 101–110, the particle “no” (“si_no”, “pero_no”), a noun (“la_joven”), and an adjective (“todas_las”)  appear as one part of the word combinations. The first verbs are visible in the range 501–510 (“dijo_que”, “era_de”, “que_tiene”). A 2-gram without function words appears in the 4091–5000 most frequent items (“sin_querer”). So due to the combination of two words, the semantic specificity of individual words increases later than for the one-word unit. As for the character n-grams, it is directly visible that the most frequent items of the classic n-grams contain many combinations of white spaces and short function words (e.g., “_de_”, “o_de”, “a_de”). Nouns and verbs can be recognized from the 501–510 MFW group onwards (“adre”, “olvi”, “aber”, “rmas”). The special types of n-grams – the ones only including mid- and multi-word n-grams (“word”) and the ones consisting only of prefixes and word endings with punctuation marks (“affix-punct”) – are more homogeneous throughout the different frequency ranges.

The general features presented here are open with regards to the grammatical and semantic characteristics and have the advantage that they cover a broad spectrum of the linguistic material in the full-text basis, the extent of which depends on the chosen frequency range and token unit. Therefore, they lend themselves well to exploratory analyses. A disadvantage coming along is that these features can be difficult to interpret. In the high frequency ranges, this is because semantic elements are scarce, and in the broader frequency spans, it is because of the heterogeneity of the features. The preparation of the semantic features that are used as a counterweight and an alternative to these general features is outlined in the next chapter.

The method used to determine thematic elements in the texts is topic modeling, a text-mining method that is unsupervised and not deterministic. Unsupervised means that the set of topics is not predefined but emerges from the text collection that is analyzed. A method that is not deterministic does not produce the same output every time it is run, even if the start conditions are fixed, which means that an element of randomness is involved in the process. The goal of topic modeling is to uncover hidden semantic relationships between the words used in the documents of a large text collection and thereby determine what the texts in the collection are about.The method was initially developed in the context of Information Retrieval as a general approach to model text corpora: “The goal is to find short descriptions of the members of a collection that enable efficient processing of large collections while preserving the essential statistical relationships that are useful for basic tasks such as classification, novelty detection, summarization, and similarity and relevance judgements” (Blei, Ng, and Jordan 2003, 993). The semantic relationships between the words are described as hidden because a topic model involves the basic units corpus, document, word, and topic, of which the first three – in their input version – are predefined structures and processable parts of the text collection’s initial state. The topics, on the other hand, constitute a medium layer between the words and the document, or the words and the corpus. The topic layer is inferred from the distributional characteristics of the words in the documents with a probabilistic approach but is not inherently manifest in the text strings and their structural organization. The basic idea comes from the field of distributional semantics and was pointedly put by J. R. Firth already in 1957: “You shall know a word by the company it keeps!” (Firth 1968, 179).The cited version of the paper is a reprint of Firth 1957. The wider context of this quote is: “The placing of a text as a constituent in a context of situation contributes to the statement of meaning since situations are set up to recognize use. As Wittgenstein says, ‘the meaning of words lies in their use.’ [...] The day-to-day practice of playing language games recognizes customs and rules. It follows that a text in such established usage may contain sentences such as ‘Don’t be such an ass!’, ‘You silly ass!’, ‘What an ass he is!’ In these examples, the word ass is in familiar and habitual company, commonly collocated with you silly—, he is a silly—, don’t be such an—. You shall know a word by the company it keeps! One of the meanings of ass is its habitual collocation with such other words as those above quoted” (Firth 1968, 179). The essence of the quote is that words that repeatedly precede and follow a particular word of interest in a defined textual context contribute to its meaning so that words that occur in the same contexts tend to have similar meanings. This distributional hypothesis is also the theoretical basis for topic modeling.For a general discussion of the foundations of the distributional hypothesis, see Sahlgren 2008. Words that occur in the same context in many documents of a text collection are grouped together into topics because it is assumed that their meanings are related.What exactly the “context” is and which scope it has depends on the implementation of the distributional analysis. The two main approaches are to either use surrounding words or to use text regions in which the words occur together as a basis. Current implementations of topic modeling, for instance, Latent Dirichlet Allocation (LDA), usually follow the latter strategy (Sahlgren 2008, 33). Depending on the kind and size of the context, Sahlgren differentiates between two types of semantic similarity. According to him, a wider, document-oriented context leads to models that capture “semantic relatedness (e.g. ‘boat’ - ‘water’)”, while a narrower, word-oriented context models “semantic similarity (e.g. ‘boat’ - ‘ship’)” (Sahlgren 2015). As topic modeling is document-oriented, topics are characterized by semantic relatedness.

In the topic modeling approach, the terms “word”, “document”, “corpus”, and “topic” have specific meanings. Words correspond to tokens that the text is split into and the tokens that are selected for the topic modeling analysis. They can but need not concur with words in a linguistic sense, depending on how the tokens are defined. Often a list of stop words is applied before the topic model is built so that not all the words that are part of the initial text string stay in the resulting model. The goal is to remove words that carry little semantic value, such as function words, in order to get topics that consist mainly of content words whose distributional relationships are easy to interpret as being of a semantic kind. The term “document” has different meanings, depending on the level on which it is used in a topic modeling workflow. In the narrow sense, “document” is defined as a structure resulting from the topic modeling process. In a wider sense, though, “document” also has a special meaning as an input structure for the topic modeling algorithm. To be able to differentiate between these different uses of the term “document”, they are distinguished here by indexing the term. For the purpose of this dissertation, documentS means the original source document in the form of a continuous text string of characters, documentIn-1 is the input document that the topic modeling algorithm gets, documentIn-2 is the input document that the topic modeling algorithm creates internally and uses as a basis for the modeling process, and documentOut is the output document resulting in the final topic model.

DocumentsIn-1 are the text units that constitute the contextual frame in which the co-occurrence of the words is analyzed and counted. They can but need not correspond to entire documentsS of a text collection such as books or articles. Alternatively, they can be combinations of documentsS or subparts of them. In practice, the size of the documentsIn-1 for topic modeling is often chosen with the aim of balancing the length of the text units and optimizing them for the algorithm to produce good results. Usually, very long texts and texts with a significant variance in length, such as novels, are segmented into smaller units.So far, the decisions for a certain length of the documentsIn-1 follow empirical experience or rules of thumb, because there are no theoretical foundations for this parameter of text preprocessing for topic modeling yet. In a survey on LDA-based topic modeling in Digital Humanities, Du notes: “The common preprocessing procedures include lemmatization, part-of-speech (POS) tagging and document chunking. [...] Chunking allows us to capture topics which only appear at certain points. My survey pays particular attention to the reasons of applying (or not) a preprocessing procedure in practice. [...] Document chunking is very diverse: the chunk-size could be several hundred or several thousand words, or a page of a book, or to split a book into ten equal segments. But almost no approach explained the reason of their chunking choices” (Du 2019). For a topic model, the documentsIn-1 are not kept as sequences of words, punctuation marks, and blank spaces. Instead, internally, they are converted to a collection of word counts following the bag-of-words approach, the documentsIn-2, as demonstrated with the sentence in example 50 and the resulting word count matrix in table 29.

In the example, the proper names “Belgrano” and “Buenos Aires” are treated as stop words and are not included in the word count matrix. Furthermore, all the tokens are converted to lowercase before counting them. Punctuation marks and blank spaces are removed, and the order of the words is suspended. Already in this single sentence, the counts give an impression of the kind of words that dominate from a quantitative point of view: “de” (7), “la” (3), and “el” (2), which are all function words. Table 30 shows how the matrix is further reduced if the sentence is preprocessed by only selecting the lemmas of nouns. Word count matrix with word lemmas. calle casa ciudad hombre mayo noche patio 1 1 1 1 1 1 1

It becomes clear that the decisions on how to preprocess the texts significantly influence what a word and a documentS become in the topic modeling. In the case of lemmatization, also morphological information is lost, and by selecting only one type of word category, the initial sentence is reduced to content buzzwords. Furthermore, it is not only the structure inside of sentences that is reduced. The sentences themselves and higher-order structures inside of the documentsS are also not preserved.

In the context of topic modeling, also the term “corpus” has several meanings if both the input and output states of the model are considered. As for the documents, I differentiate between corpusS as the collection of documentsS, i.e., the original full-text files, corpusIn-1 as the collection of documentsIn-1, that is, the set of full-text snippets that the topic algorithm receives as an input and that already can have been preprocessed by stop word removal, linguistic annotation and selection, or chunking. The corpusIn-2 is the collection of documentsIn-2, which is a matrix of terms and their counts per document.This is the so-called “term-document matrix”, “a data structure, a computationally tractable (to use McCarty’s term) representation of the texts able to be modeled by a computational process” (Burton 2013). Finally, there is also a corpusOut, which is part of the final topic model.

The term “topic” is only defined on the output level, as this structure does not exist on the input levels and is found by the algorithm. Technically, a topic is a probability distribution over words. For each word that is part of the corpusIn-2’s vocabularyHere, “vocabulary” means the set of different words contained in the corpus., a probability value indicates how important it is for the topic at hand (Steyvers and Griffiths 2007, 2). Table 31 illustrates the structure of topics by showing the 15 top words of two different topics that are part of a topic model created for the corpus Conha19.The example topic model is available at https://github.com/hennyu/papers/tree/master/family_resemblance_dsrom19/features/topicmodel. Accessed November 12, 2020. It was created with the following parameters: 100 topics, 5,000 iterations, and a hyperparameter optimization interval of 100. The texts were preprocessed by lemmatizing them and only using nouns. The original input full-text files were chunked into segments with a length of 1,000 tokens. A list of stop words was applied, which contained the 50 most frequent nouns plus some nouns that were added manually. The structure of the table corresponds to the examples shown in Steyvers and Griffiths (2007, 2). Top 15 words of two example topics. Topic 77 Topic 10 Word Probability Count Word Probability Count amor 0.09260 8839 bandido 0.05592 763 corazón 0.06372 6082 jefe 0.02844 388 alma 0.03663 3496 ladrón 0.02375 324 pasión 0.02052 1959 robo 0.01752 239 felicidad 0.01882 1796 policía 0.01708 233 palabra 0.01761 1681 crimen 0.01700 232 ser 0.01235 1179 compañero 0.01634 223 sentimiento 0.01187 1133 bandolero 0.01158 158 flor 0.01173 1120 lugar 0.01099 150 mirada 0.01113 1062 justicia 0.00880 120 pensamiento 0.01009 963 sociedad 0.00777 106 placer 0.00950 907 camarada 0.00770 105 esperanza 0.00819 782 silencio 0.00770 105 cielo 0.00777 742 comandante 0.00696 95 ángel 0.00776 741 camino 0.00660 90

An important insight is to realize that each topic consists of all the words in the corpus vocabulary and that the differences between the topics are the result of the different weights that the words have in each topic. When humans interpret topics, usually, they only examine a certain number of words – those with the highest probability. The semantics of the topics emerges from the combination of high-probability words. A human inspector can evaluate the semantic relationships between the terms and label the topics with titles. A topic title can consist of individual words but also descriptions. In the above example, topic 77 could be entitled “love” or “love and feelings” and topic 10 “crime” or “crime and society”. Computers can also evaluate the resulting topics by calculating the similarity and semantic relationships of the most important terms in each topic by using external semantic resources, for example, word embeddings,Word embeddings are another method from the area of distributional semantics. In word embeddings, words from a vocabulary are converted to vectors of real numbers, and semantic relationships can be inferred from the proximity or distance of the word vectors to each other (Sahlgren 2015). word nets,Word nets are lexical databases of semantic relations between words, for example, synonymy or hyponymy (Miller 1995). or dictionaries. Computers could also label topics, for example, by finding superordinate terms. Nevertheless, for creating the topics, no explicit semantic knowledge is necessary because the algorithm bases the assumptions on the relationships between the terms in a topic only on statistical, distributional patterns. Another advantage of topic modeling is that polysemy is not problematic. As every word occurs in every topic, the same word can be important in several topics where its meaning is determined by the surrounding words. In the table, the “probability” column indicates how probable a word is in a specific topic, and the “count” column shows how often a word has been assigned to a topic across all documents. When the counts are compared, the love topic is more important in the corpus than the crime topic because the number of tokens assigned to it is higher. In total, the love topic has 95,452 token assignments, and the crime topic has only 13,644.This information about token counts per topic can be found in the diagnostics file of the tool MALLET, which was used to create the model (McCallum 2018a). Inside each topic, the probabilities indicate the relative importance of each word. In the love topic, the first words have higher probabilities than the first words of the crime topic. However, the differences between the probabilities decrease for the lower word ranks, which means that the love topic is more dominated by a few very important words than the crime topic.

In the topic model output, the documentOut is a probability distribution over topics. For each topic that was defined for the corpus, a probability value is given for every document, and topics with higher probabilities are considered especially relevant for the document in question. It follows from this that the corpusOut is a matrix of probability distributions over topics. The structure of three example documentsOut is given in table 32, showing the probabilities of five selected topics of the model.See footnote 536 above for information about the underlying topic model. In the table, the topic probability values are rounded. Example documentsOut from a topic model. Document number Document name Topic 0 Topic 1 Topic 2 Topic 10 Topic 77 0 nh0217§0030.txt 0.04178 0.00009 0.00003 0.00362 0.00029 1 nh0107§0052.txt 0.00002 0.00384 0.00004 0.00004 0.00406 2 nh0103§0012.txt 0.00002 0.00010 0.00004 0.00005 0.14777

Here, the names of the documents are a combination of the novels’ identifiers in Conha19 (“nh0217”) and the numbers of the text segments that served as documentsIn-1 (“§0030”). To be able to evaluate the probabilities of the topics in the entire novels, the values for the individual segments have to be aggregated again, for example, by using average probabilities. The different probability values show how important the topics are in the documentsOut. Regarding the love topic (topic 10) and the crime topic (topic 77), the table shows that the segment of the first document in the list (“nh0217§0030”) has a higher probability for the love topic than the other two segments from other documents and the third segment (“nh0103§0012”) has a higher probability for the crime topic than the preceding ones. The identifier “nh0217” belongs to the novel “El espejo de Amarilis” (1902, MX) by Laura Méndez de Cuenca and “nh0103” to the novel “El mendigo de San Ángel” (1865, MX) by Niceto de Zamacois. The first one is a novel of customs involving romantic plot elements, and the latter is a historical novel. The results of the topic model seem reasonable, since a prominent love topic can be expected in a novel with a romantic plot, and it is also plausible that the crime topic would carry greater weight in a historical novel.

The topic model was created with the tool MALLET (McCallum 2002), in which the topic modeling algorithm Latent Dirichlet Allocation (LDA) is implemented (Blei, Ng, and Jordan 2003; Blei 2012). Other mathematical and technical approaches to topic models exist, but LDA is the most prominent current technique and is widely used.A forerunner of LDA is, for example, Latent Semantic Analysis (LSA), which is not probabilistic (Landauer and Dumais 1997; Landauer, Foltz, and Laham 1998). Newer approaches are Nonnegative Matrix Factorization (NMF) and a network-based approach using a stochastic block model (SBM) (Arora, Ge, and Moitra 2012; Gerlach, Peixoto, and Altmann 2018). In the context of deep learning, the method lda2vec has been developed, which combines the learning of word, document, and topic vectors (Moody 2016). LDA is based on a generative model, assuming the documents are generated based on the topics. The method works by choosing a distribution over topics for each document and subsequently choosing a topic from this distribution for each word in each document. Finally, a word is chosen from the topic’s distribution over words. The process starts with an initial distributionThis is the Dirichlet distribution which gives the algorithm its name (Blei, Ng, and Jordan 2003, 996–997). and approximates it to the data by iterating over the words in the corpus.This process is called Gibbs sampling (Steyvers and Griffiths 2007, 7–9). The most important parameter that needs to be set for the algorithm is the number of topics to model.

For humanists who want to evaluate results obtained from topic modeling or who are interested in using the method to produce their own results, it is fundamental to recognize the different meanings of the basic terms word, document, corpus, and topic in the context of topic modeling compared to the context of thematic analyses based on linguistic and literary theories. In the “Reallexikon of German Literary Studies”, for instance, the term “Thema” is defined as “Die einem Text zugrundeliegende Problem- oder Gedankenkonstellation” (Schulz 2007, 634), which means the central topic or subject of a text and corresponds roughly to the English term “theme”. This sense of topic is different from the topicsTM resulting from topic modelingThe index TM is used here to differentiate the general term topic from the term as it is defined in the context of the topic modeling method. in that it is defined on a more abstract level. The underlying theme of a text can be interpreted from its linguistic material but does not necessarily have to be directly present in terms of formulations using the word or words that describe the theme. For example, one could think of a short story describing a romantic dinner and containing dialogues of a couple, but in which the word love is never used. Nevertheless, a reader could conclude that the central topic of the story is just love.

The topicsTM, on the other hand, are closer to the textual surface. A topic model captures how topics are realized in text segments. However, to find one or several central themes of a text, the topicsTM need to be interpreted on a more general level.For a useful discussion of the similarities and differences of topicsTM to literary theoretical terms on thematic aspects, see Horstmann (2018). Besides the term Thema, Horstmann also compares the German terms Motiv, Topos, and Sujet to topicsTM and comes to the conclusion that they all cover different thematic and content-related aspects of literary texts. A term that is better suited to be related to the topicsTM  is the text-linguistic term thematische Entfaltung. It focuses on how the theme or central topic of a text is unfolded in the overall content of the text. Brinker, Cölfen, and Pappert (2014, 57–80) define the thematic unfolding as a combination and linkage of relational and logical-semantically defined categories that express the relationship of partial topics present in individual parts and substructures of the text to its central topic. They mention justification and specification as examples of such relational categories. Obviously, the same central topic can be unfolded in different ways. Brinker, Cölfen, and Pappert also define a set of basic forms of thematic unfolding: descriptive, narrative, explicative, and argumentative thematic unfolding.In German, the term Themenentfaltung is defined as “die gedankliche Ausführung des Themas” and, more specifically, “Die Entfaltung des Themas zum Gesamtinhalt des Textes kann als Verknüpfung bzw. Kombination relationaler, logisch-semantische definierter Kategorien beschrieben werden, welche die internen Beziehungen der in den einzelnen Textteilen (Überschrift, Abschnitten, Sätzen usw.) ausgedrückten Teilinhalte bzw. Teilthemen zum thematischen Kern des Textes (dem Textthema) angeben (z. B. Spezifizierung, Begründung usw.)” (Brinker, Cölfen, and Pappert 2014, 57). These concepts go back to Brinker (1992). Brinker, Cölfen, and Pappert stress that the thematic unfolding is influenced significantly by communicative and situational factors, such as the intention or purpose of the communication, which leads to several different possibilities of unfolding the same central topic. However, little is known about the factors which have an effect on the exact unfolding. From the point of view of digital literary studies and especially stylistics, this means that also stylistic intentions can have an influence on the thematic unfolding, or, looked at from another perspective, that the exact thematic unfolding can be interpreted as an intended or unintended stylistic effect. The distribution of topicsTM in the documentsOut of a topic model could be interpreted as the result of thematic unfolding. On the level of the corpusOut, the specific set of topicsTM can be understood as resulting from the thematic unfolding across all documents in the text collection. The thematic unfolding itself, as defined by the combination of relational and logical-semantical categories, though, is located on a more abstract level, which is intermediate between the topicsTM and the theme of the text. Brinker, Cölfen, and Pappert point out that the theme or central topic of the text is the shortest possible summary of the textual content and that only a reader’s interpretation can achieve this reduction: Man muss sich überhaupt darüber im Klaren sein, dass die textanalytische Bestimmung des Themas primär auf interpretativen Verfahren beruht; es kann hier keine ‘mechanische’ Prozedur geben, die nach endlich vielen Schritten automatisch zur ‘richtigen’ Themenformulierung führt. [...] Die Bestimmung des Themas ist vielmehr abhängig von dem Gesamtverständnis, das der jeweilige Leser von dem Text gewinnt. Dieses Gesamtverständnis ist entscheidend durch die beim Emittenten vermutete Intention bestimmt, d. h. durch die kommunikative Absicht, die der Sprecher / Schreiber mit seinem  Text nach der Meinung des Rezipienten verfolgt. (Brinker, Cölfen, and Pappert 2014, 53–54)

Regarding the possibilities of a mechanical (or computational) method to determine the theme of a text, I take a more moderate position. Even if a reader is indispensable as the last instance of interpretation, formal methods can be used to evaluate how a theme is thematically unfolded in a text. Furthermore, the basic types of thematic unfolding have been taken up by Schöch and Rißler-Pipka in a topic modeling analysis of literary texts. In a contribution to a conference panel on drama analysis, Schöch and Rißler-Pipka analyzed the distribution of argumentative, narrative, descriptive, and discursive topics in 1,100 novels, 800 dramatic texts, and 1.8 million Wikipedia articles in French to see if the proportions of topic types vary by genre.For the abstract of the conference panel, see Willand et al. (2017). The results of the contribution by Schöch and Rißler-Pipka about topic types can be found in the presentation, which can be downloaded at https://github.com/christofs/dramenanalyse-dhd/. Accessed November 14, 2020. Moreover, different forms of thematic unfolding are also studied in computational studies in general.See, for example, the study on Argument Mining using topic modeling presented by Lawrence and Reed (2017). These approaches perform much of the text-analytical determination of topics and theme and push the role of a human reader or a human interpreter of analytical results to a higher level.

Whatever the relationship between literary-theoretical and linguistic terms that cover content-related aspects and topicsTM, so far, the systematic analysis of content has not been a key concern in literary scholarship: ’Inhaltsanalyse’ (engl. content analysis) ist eigentlich kein literaturwissenschaftlicher Begriff. Diese [in den Sozialwissenschaften verbreitete] hier stark vereinfacht beschriebene, mit diversen Kontrollverfahren zur Einhaltung von Intersubjektivitätsmaßstäben begleitete, insgesamt sehr aufwändige und inzwischen bei der Texterfassung, -bearbeitung und -auswertung weitgehend computergestützte Vorgehensweise ist lieraturwissenschaftlichen Umgangsformen mit Texten fast völlig fremd. Allenfalls in der Computerphilologie [...] zeigt man sich mit ihnen vertraut. (Anz 2007, 55)

As Anz points out, systematic content analysis has a longer tradition in the social sciences than in literary studies. According to him, the core interest of literary studies is much more the specific literary function of the texts, which manifests itself in particular linguistic and structural forms (Anz 2007, 57). In this regard, the method of topic modeling, with its dependence on the textual surface, is closer to the concern of literary scholarship than content analysis because it allows us to examine the relationships between content and form more directly. In computational literary studies, topic modeling has indeed been taken up with interest and applied to a variety of literary texts with success (see, among others, Jockers 2013; Rhody 2012; Schöch 2017c). Literary scholars that used topic modeling soon discovered that the relationships between the terms in the topics are not necessarily content-related and do not even need to be of a semantic nature. When applied to literary texts, topic modeling can also discover rhetoric structures and elements of discourse (Jannidis 2016, 27).The method has even been applied intentionally to discover non-thematic structures, for example, by Rhody: “By locating ‘figurative language’ as an aspect of address for topic modeling, I choose to constrain my consideration of poetic texts and agree to a caricature of poetry that hyper-focuses on its figurative aspects so that we can better understand how topic modeling, a methodology that deals with language at the level of word and document, can be leveraged to identify latent patterns in poetic discourse” (Rhody 2012). That such structures were especially noticed when topic modeling was applied to literary texts shows that the method, which originally aimed at enabling content analysis, has been developed and optimized on the basis of non-literary text types. It also indicates that the distributional hypothesis could be understood in a wider sense as not only applying to semantic relationships but also to discourse relationships. However, to what extent non-thematic topics can be produced also depends on the kind of preprocessing applied to the texts. If only nouns are processed, it is more probable that the words in the topics are held together by semantic, thematic, and content-related connections.

For the analysis of subgenres of the novel pursued here, this traditional way of using topic models is intended. Therefore, the linguistically annotated version of the novels is used, and only noun lemmas are selected, aiming at semantic and content-related topics. This way of preprocessing the texts has also been applied in other digital literary studies that used topic modeling for genre distinction and categorization.For instance, only noun lemmas are used in Hettinger et al. (2016) and Schöch, Henny et al. (2016). Noun, verb, adjective, and adverb lemmas are used by Schöch (2017c). Primarily stylistic aspects, on the other hand, are intended to be covered by the general, most frequent words-based features, so that the noun-based topics function as a counterpart and as semantic features. As for the general features, also for the topics, several feature sets were prepared using different parameters, as summarized in the following table 33. Parameters for topic feature sets. Parameter Selected values Number of topics 50, 60, 70, 80, 90, 100 Optimization  interval 50, 100, 250, 500, 1000, 2500, 5000, None

The first parameter that was varied is the number of topics. Here, a range of 50 to 100 topics is considered reasonable for the corpus with 256 novels. In his study of French classical drama, Schöch (2017c) used 60 topics for a collection of 391 novels. Schöch, Henny et al. (2016) used 70 topics for a corpus of 150 Spanish and Spanish-American nineteenth-century novels, and Hettinger et al. (2016) found the best performance for classifying 628 German novels with 100 topics. In general, a lower number of topics produces more general results, and a higher number covers more specific thematic aspects of a text collection.

The second parameter that was selected for variation here is the optimization interval, that is, the interval at which the Dirichlet hyperparameters for the LDA model are optimized during the iterations of the topic modeling process. The two hyperparameters, alpha (α) and beta (β), influence the form of the topic probability distribution over words and the document probability distribution over topics, respectively. The lower the parameters are, the more the resulting distribution is concentrated on single values. Otherwise, it is more even. It follows that a lower alpha value leads to topics with a few dominant and many unimportant words and a lower beta parameter to documents that are dominated by a few topics rather than by a larger number of roughly equivalent topics. In the MALLET implementation of LDA, which was used here,MALLET was used in version 2.0.8RC3. the parameter optimize-interval indirectly influences the form of the two types of distribution. A lower rate of optimization results in more even distributions and a higher value in more skewed ones. By not setting this parameter at all, no hyperparameter optimization is performed, which leads to even probability distributions (McCallum 2018b; Wallach, Mimno, and McCallum 2009). Schöch (2016, 2017c) has analyzed how different optimization intervals influence the results for subgenre classification in French novels and drama.He recommends setting the optimization interval depending on the goal of the analysis: “If your goal is to identify small numbers of texts about specific themes in a large collection, then a lot of optimization may be good. However, if your goal is to identify topics typical of certain authors, periods, genres or some other reasonably large subset of your collection, then it may be better to optimize a bit less” (Schöch 2016). The range of intervals that has been selected here varies from 50 to 5,000 and also includes None. With a fixed number of 5,000 iterations, this results in 50, 20, 10, 5, 2, 1, and no optimizations.

As to the preprocessing of the texts, the following procedure was followed: first, the linguistically annotated versions of the TEI corpus files were used to extract the noun lemmas for each novel and to create full-text files only containing the nouns. As for the general features, also here proper names were excluded.The script for creating the noun lemma files is available at https://github.com/cligs/scripts-nh/blob/master/corpus/derivative_formats/get-plaintext-annotated-nouns.xsl. The resulting plain-text files are available at https://github.com/cligs/conha19/tree/master/txt_annotated_nouns. Accessed November 15, 2020. Then, the novels are segmented into chunks of 1,000 tokens, which are fed into the topic modeling workflow. A stopword list consisting of the 50 most frequent nouns and of proper names and place names was created and used by MALLET.The proper names and place names in the stopword list can cover cases that were missed by the named entity recognition. The stopword list for the topic models is available at https://github.com/cligs/data-nh/blob/master/analysis/features/stopwords/topics_stopwords.txt. Accessed November 14, 2020. The list was refined by inspecting selected topic models and adding words with very general meanings that dominated some topics and reduced their interpretability. For text segmentation, modeling, and post-processing of the results, the tool tmw (Topic Modeling Workflow) was used. Tmw was developed by Christof Schöch and Daniel Schlör in the context of the CLiGS project. It is a set of python scripts that especially supports the pre- and post-processing of the texts that are used for the topic modeling. It also includes a set of functions to visualize the topic modeling results. Its development is organized in a GitHub repository (see Schöch and Schlör 2017).The tool was presented at a workshop at the DH2017 conference in Montréal by the CLiGS team (Betz et al. 2017). A more recent presentation about topic modeling that includes a description of tmw is Schöch (2019). The tmw scripts have been slightly adapted to create the topic models used here. Mainly the routines for calling the core functions were changed.The version of tmw which was used here is available at https://github.com/cligs/scripts-nh/tree/master/features/tmw. Accessed November 14, 2020. An important feature of tmw is the text segmenting procedure. Before the novels are processed with MALLET, they are split into segments of a defined length,Tmw is even able to apply a smooth segmenting that respects paragraph boundaries, but this feature was not used here. Instead, a fixed segment length was defined. so that the topic models are created for the segments, not the entire texts. In the post-process step, tmw recombines the segments of the novels and calculates average probability values. The segmenting step is considered important to counter the effects of different text lengths, and the recombination of the segments is necessary to be able to interpret the results per novel. Finally, because topic modeling is not deterministic, five different models were produced for each parameter constellation.The script that applies the different parameter settings, controls the topic modeling workflow, and calls the functions of tmw is available at https://github.com/cligs/scripts-nh/blob/master/features/topics.py. The resulting topic models are available at https://github.com/cligs/data-nh/tree/master/analysis/features/topics/.  Accessed November 14, 2020. Through the different combinations of topic modeling parameters and the repetitions of the modeling process, 240 different topic models were created for the whole corpus.

As for the MFW-based features, also for the topics, some general characteristics of the feature sets are evaluated here. Unlike for MFW, the number of zero values is not relevant for the topics, though, as no zero values were found in the whole set of topic models created for the corpus of Spanish-American novels. This means that each topic has at least some probability in each document, even if it is very small.The charts summarizing this finding can be viewed at https://github.com/cligs/data-nh/tree/main/analysis/features/topics/overviews. Accessed December 22, 2020. Another aspect of the models that was evaluated is the coherence of the topics. The tool MALLET includes a set of diagnostic measures which can be saved as an additional output of the topic modeling process and can be used to assess the formal quality of the topic models. One of the measures is the coherence metric, which evaluates whether the words in a topic actually occur together in the texts of the corpus. This is examined by taking each pair of top words in a topic and calculating “the log of the probability that a document containing at least one instance of the higher-ranked word also contains at least one instance of the lower-ranked word” (McCallum 2018a). The resulting scores are negative, and values closer to zero mean that words co-occur more often. For the models at hand, the number of top words was set to 50. As the coherence measure depends on the top words, it is expected that lower numbers of top words lead to coherence scores that a closer to zero than the scores for models with a higher number of top words. However, as the number of top words is the same for all the models here, the resulting coherence values can be compared. A summary is given in figure 43.

For the figure, the mean coherence of all the topics in a topic model was calculated. This was done for each combination of the number of topics and optimization intervals. As there are five different models for each setting, their mean coherences were averaged. The results show two trends. First, the coherence of the topics tends to decrease with an increasing number of topics, and second, the coherence also decreases when the optimization intervals are bigger (so that the hyperparameters are optimized less often). What can also be seen is that the values drop more for bigger optimization intervals than for higher numbers of topics. It makes sense that less optimization leads to topics that are “further away” from the actual distribution of words in the corpus so that the topics tend to be more abstract. It is more difficult to explain that higher topic numbers produce less coherent topics. A hypothesis for this is that with more topics, many of them are more special, so it is less probable that texts containing some top words of a topic also contain all the others. Interestingly, though, the coherence values go up again from 90 to 100 topics for the lower optimization intervals. Regarding the usage of the topic models for subgenre categorization, the coherence values are of interest in two ways. One is the question of whether models with more coherent topics are better feature sets for classification, and the other is whether a better formal coherence of the topics also facilitates their interpretation as features related to the subgenres. Four example topics are shown in figure xxx to get an impression of the kind of topics that are part of the different models.

Each word cloud depicts one topic and includes its 40 most important words. The top left topic stems from a model with 50 topics and an optimization interval of 50, and the top right from a model also with 50 topics but without hyperparameter optimization. The bottom left topic is part of a model that has 100 topics and an optimization interval of 50, and the bottom right topic was produced in a model also with 100 topics but without hyperparameter optimization. With optimization, the topics can be ranked because they can have different overall probabilities in the corpus. The tenth most important topic was selected from the two models with optimization (as indicated by the first number in parentheses after the topic number). Although the overall number of topics is different in the two models, the tenth topic is similar in both. Without optimization, all the topics in the model have the same weight in the whole corpus, so of the two models without optimization, two topics with similar words were selected. The two topics to the left can be interpreted as representing a ball. In the top left topic, the three most important words are “baile”, “salón”, and “amor”, and in the bottom left one, “baile”, “salón”, and “brazo”. The other words in the topics are also related to a ball situation and cover music, dance, social life and conversation, and dresses. The main difference between the topic from the 50-topics model and the one from the 100-topics model is that in the former, words about love and family relationships are more prominent (“amor”, “mamá”, “novio”, “corazón”, and “papá”), suggesting that here the ball situation is connected with central plot elements, whereas the latter has a stronger focus on the ball as a social event and seems to be more descriptive, and hence more specialized (“brazo”, “traje”, “movimiento”, “efecto”, “concurrencia”, “reunión”). The two topics to the right can be interpreted as covering travel to or in a city. In the topic from the 50-topics model, the three most important words are “coche”, “carruaje”, “ciudad”, and the three top words in the topic from the 100-topics model are “coche”, “carruaje”, and “cochero”. As in the case of the ball topics, also here the first topic includes some top words which add elements from another type of situation. The word “hotel” is more important in that topic than in the second one, and there are the words “mesa”, “juego”, “jugador”, “reloj”, and “oro”, which seem to describe a situation at a gaming table. Such a subtopic is not noticeable in the second topic. Overall, from the point of view of a human interpreter, all four topics stemming from models that were created with different topic modeling parameters have good quality and are semantically coherent, with differences only on the level of detail. Interestingly, the two topics from the 100-topics model give a more coherent impression than the two topics from the 50-topics model, which is contrary to the outcome of the MALLET diagnostics and shows that the degree of formal coherence does not necessarily coincide with the semantic coherence that can be observed in the topics by a human interpreter. In any case, topics as semantic features are mostly easy to interpret, so it can be expected that the feature sets resulting from topic modeling are useful for gaining new insights into the subgenres of the novel.

The method which is used here to group the novels into discrete classes of texts is statistical classification, as it is defined and implemented in the context of machine learning. According to Alpaydin, machine learning is programming computers to optimize a performance criterion using example data or past experience. We have a model defined up to some parameters, and learning is the execution of a computer program to optimize the parameters of the model using the training data or past experience. The model may be predictive to make predictions in the future, or descriptive to gain knowledge from data, or both. (Alpaydin 2016, 3)

Machine learning is applied in cases in which it is very difficult or not possible to know the rules that connect a certain type of data to a characteristic that is attributed to this data. In the case of genres, this means that there are texts about which we have some knowledge: we know, for example, which words are used in the texts, how they are organized syntagmatically, to which grammatical categories they belong, and so on. On the other hand, we know that certain texts belong to specific genres because it is indicated in the paratexts of their publication, because the books they are published in are offered in a genre-specific section of a bookstore, because some literary scholar has said that they belong to a genre, etc. We may have some ideas about how the input characteristics of the texts are related to the output labels, but we do not know enough to be able to design corresponding algorithms for the computer that could convert the input (the features) to the output (the category labels) in a direct way.

In machine learning, the computer uses the data that has been labeled with some output value to learn the rules from it. This is done by using a model of a predefined type (for example, a linear model aiming to describe connections in the data with linear relationships), and adapting and optimizing this model by setting its parameters in a way that fits the data and their labels best. The goal of training such a model is to be able to automatically label new cases of the data by applying the trained model to them, i.e., to make predictions. In the case of genre, this would mean being able to assign a genre label to a new, unknown text by handing it to the computer. As Alpaydin states, another goal can be to inspect which part of the data the computer used to make the predictions to learn something about the rules that connect the input data to the output labels. For the genres, this would mean reproducing which features of the texts, for example, the frequency of a certain word or the distribution of word categories, really are relevant for the decision to assign the texts to certain genres.This description of machine learning is based on Alpaydin (2016, 1–3).

Approaches in machine learning are broadly classified into unsupervised and supervised learning methods. In an unsupervised case, knowledge is extracted from the data without information about some output value or category. The data may, for example, be grouped based only on how the features are distributed. In a supervised approach, on the other hand, a specific outcome is known for a given input. Classification belongs to the latter group because the goal is to find a model for the relationships between the input data and the output class labels (Müller and Guido 2016, 27, 133).

Usually, the aim of a trained machine learning model, or more specifically, a classifier, is not just to be able to treat the data from which it has learned. Instead, it should be a general model, in the sense that it can differentiate between instances of the classes independently of the specific data set. For the corpus of nineteenth-century Spanish-American novels at hand, for instance, this implies that a classifier trained to recognize the difference between historical and non-historical novels should be able to make correct predictions for another corpus. A different corpus may be built using the same population of nineteenth-century Spanish-American novels as represented in the bibliography but selecting different authors and works. This would mean that the classifier can tell us something about the general difference between historical and non-historical novels of that period and cultural-geographical context, and not just about how the genre novela histórica is realized in the specific works in the selected corpus. By training a classifier not just on nineteenth-century but also on twentieth-century novels, one could create a model that is independent of the time period and learn a more abstract data-based concept of the historical novel. Similarly, also the cultural-geographical and linguistic context could be broadened or narrowed down, for example, to only learn about the historical novel in Mexico. As in literary studies in general, these examples show that also in digital literary studies applying machine learning and classification, the design of the corpus is decisive for the kind of conclusions that can be drawn from the text analysis results.For a formal definition of the generalization problem, see Alpaydin (2016, 23–27, 37–41).

To be able to build a model that is general for the selected context, a supervised machine learning task is usually performed in several successive steps. First, part of the data is used as a training set to build the model and find the best parameters for it. Another part of the data is kept as a validation set to be able to check how well the model performs if it has to classify data on which it has not been initially trained. Subsequently, the model is refined by adjusting its parameters so that it performs best on the validation set. That way, the validation set becomes part of the training process of the model. By repeating the process of splitting the data into different training and validation sets (cross-validation), the effect of random chance in selecting a specific training set can be reduced. The parameters of the model should be fit in a way that makes the model neither too specific for the data (overfitting) nor too general (underfitting) (Alpaydin 2016, 39–41). If, for example, a model is built with the goal of differentiating between historical and sentimental nineteenth-century Spanish-American novels, but the training set is dominated by a certain type of historical novel, then it would be disadvantageous if the model learns too many details about the special types of historical novels when compared to sentimental novels. Such special types of historical novels are, for example, novels that were part of large series of historical novels published by a few Mexican authors, such as the “Episodios nacionales mexicanos” (1902–1903, MX) by Victoriano Salado Álvarez, or the “Leyendas históricas de la independencia” (1886–1913, MX) by Ireneo Paz. If other types of historical novels are presented to the model as a test case, it could happen that the specialized model is not able to classify them correctly. This could happen, for instance, if an individual Argentine or Cuban novel dealing with the conquest or colonial times is presented to a model trained with Mexican historical novels. Such a model would be overfitting. Conversely, if a model does not predict the classes of the data it was trained on well, it has probably learned too little about the data structures representing the classes and is underfitting. Finally, a fully trained and validated model can be used with a third part of the data, the test set, which has not been part of the training cycle at all.

Obviously, not just the type of model and the selection of the best parameters for it are decisive for the quality of a classification task, but also how – based on what data – the instances of the classes are represented in the machine learning process. The selection of the text features for genre classification should be based on good hypotheses about their relevance for the problem. Here, it was decided to use the most frequent words and topics as two comparatively generic types of features to classify the nineteenth-century Spanish-American novels by subgenre. Both types of features have already been used successfully in the classification of literary genres and other text types. Hettinger et al. (2016), for instance, used the 3,000 most frequent words, the 1,000 most frequent character 4-grams, and topic models to classify a corpus of German nineteenth-century novels by subgenre. They achieved accuracy scores between 70 and more than 90 %, depending on the feature set used and on the constellation of subgenres. Schöch (2017c) classified French dramas of the Classical Age and Enlightenment by subgenre, using a set of topic models with varying parameters. In Schöch’s study, the accuracy reached between 70 and 87 % for different numbers of topics, optimization intervals, and classifiers. Schöch also tested to classify the dramatic subgenres based on the most frequent words. He found the best results with 3,500 MFW and a z-score transformation of the word frequencies.

Although the most frequent words and topics are quite different types of features, both cover much of the underlying textual material. The hypotheses that the distributions of most frequent words and topics in the texts can be related to subgenre distinctions are fairly general.See chapter 4.2.1 on the chosen feature sets for details. By testing different sets of the two types of features, the results of the classification can help to refine the hypotheses about the features’ relevance to capture the differences between the subgenres.It would also be possible to use more specific types of features, for example, temporal expressions (including dates or words related to duration or repetitive events). By using such special features, the textual material is reduced much more, making it more difficult for classifiers to optimize a model for the differences between subgenres. As a consequence, the hypotheses about the relevance of these features for genre classification would have to be much stronger, for example, based on assumptions about how temporal expressions are used in historical novels versus other subgenres, and the classification could serve to confirm or reject the hypotheses.