4.1.5.2.1 Theme

540So far, the discursive model has served to give a general overview of what kind of subgenre labels are relevant in Bib-ACMé and Conha19. Proceeding to the individual levels and the actual labels, it becomes clearer what these subgenres are and which ones dominate from a quantitative point of view. The different levels are presented here in the order of their relevance for analyses of the novels in the corpus from a quantitative point of view. Starting with the thematic labels, of all the works in the bibliography, 695 (84 %) have such a label. In the corpus, all the works have been assigned at least one thematic label. There are 48 different thematic labels in the bibliography, of which 18 are assigned to at least 10 works.⁴³⁶ In the bibliography and in the corpus, as well, four thematic subgenres are predominant: the sentimental novel, the historical novel, the social novel, and the novel of customs. Especially in the corpus, also the political novel is significant. Comparing bibliography and corpus, the order and ratio of the top thematic subgenres are different. In Conha19, the social novel and the novel of customs are relatively more important, and the sentimental and historical novels are less relevant. However, the absolute amount of these four subgenres is more balanced in the corpus than in the bibliography.

541An analysis of the sources for the top 18 thematic labels shows the different statuses they have as explicit historical subgenres, implicitly signaled subgenres, and literary-historically discussed subgenres.⁴³⁷ As can be seen, the historical novel and the novel of customs are the most important historically explicit subgenres, followed by the “leyenda” and the “novela policial”. Especially the sentimental novel is implicitly signaled. All four top subgenres play an important role as literary-historical subgenres. Although this list of thematic labels is already the result of a process of systematization, two kinds of one-to-many relationships persist: the same label can also play a role on other discursive levels (the “novela histórica”, “leyenda”, and “novela contemporánea”, for example, are also among the labels referring to the relationship between the novel and reality) and each novel cannot just carry one, but several thematic labels. The latter is very often the case.⁴³⁸ In the bibliography, 16 % of the novels do not have any thematic label, 45 % have just one, 23 % have two, 11 % have three, and 5 % have more than three thematic labels. In the corpus, all the novels have thematic labels: 38 % have just one, 33 % have two, 18 % have three, and 11 % have more than three different thematic labels. The difference between the bibliography and the corpus shows that the more is known about the works, the more differentiated (and less clear) their categorization is in terms of subgenres. Besides the one-to-many relationships, i.e., one work to many thematic subgenre labels, there are also relationships of overlap or inclusion among the different thematic labels. For example, the “novela abolicionista” can be considered a special type of social novel, and the “leyenda” is associated with historical content just as the “novela histórica”.⁴³⁹ Relationships of this kind also become visible in the most frequent combinations of thematic subgenre labels, which are listed in table 24 below.⁴⁴⁰

542The most frequent combination of thematic subgenre labels both in the bibliography and the corpus is “novela de costumbres” and “novela social”. Most top combinations are very frequent in both Bib-ACMé and Conha19, even if the ranks are not exactly the same. Combinations that are only in the top ten of the bibliography are highlighted in blue, and the ones that are only in the top list of the corpus are highlighted in orange. As can be seen, combinations including the “leyenda” and “novela indigenista” have more weight in the bibliography and combinations including the “novela política” and the “novela abolicionista” have more weight in the corpus. Clearly, combinations of the “novela sentimental” with other types are the most frequent ones that are not characterized by semantic overlap.

543Quantity distributions as the ones described for thematic subgenre labels here (a few frequent groups and many infrequent ones) and complexities (multiple assignments and interrelationships) have consequences for a digital quantitative analysis that aims to analyze novels in terms of subgenre categories. First, for very small groups, it is hardly possible to achieve general results as they will be influenced very much by the particular characteristics of the few individual works that form the group. Furthermore, groups of very different sizes cause problems when evaluating the performance of a categorization task. In the case of a large majority, good results could simply be achieved by always choosing that group. To avoid that, the groups to compare would have to be balanced down to the smallest size by undersampling. Another strategy is oversampling, where instances of the smallest group are duplicated, which again can cause problems of overfitting and a lack of generalizability.⁴⁴¹ This study will concentrate on the largest groups of novels that are represented in the corpus to avoid these problems as far as possible. In the case of thematic subgenres, these are sentimental, historical, and social novels, novels of customs, and political novels.

544One way to solve the problem of multiple assignments is to decide on primary labels. For thematic labels and those relating to literary currents, this was done by choosing the label

• that was most prominent in the signals,
• that was most often assigned by literary historians or mentioned by them as the primary label,
• that corresponded to the same level of generality as the major groups,
• or that was most informative.

545The process of choosing a primary label changes the order of the top thematic subgenres both in the bibliography and in the corpus.⁴⁴² In Bib-ACMé, the historical novel becomes more important than the sentimental novel, and in Conha19, also the historical novel rises from the fourth to the top position and changes place with the social novel. This shows that both the sentimental and the social are often secondary thematic aspects in the novels, but at the same time, they can also be primary concerns. The primary labels can be used as target values in a standard categorization task, keeping in mind that part of the works has secondary labels when evaluating the results. Another possibility to handle multiple assignments of subgenre labels on the same discursive level is to include them all in the analysis. For a categorization task, this would mean allowing multiple target values, as, for example, in a multi-label classification.⁴⁴³ In both cases, one has to be aware of using subgenre labels that stem from multiple sources and that have been assigned to the works by people with different perspectives on them and different interests in them. The goal of such a categorization can only be to examine to what extent the collective labelings actually reflect characteristics of the texts, and to find out to which traits of the texts they correspond. If literary-historical sources are included, these already introduce an element of systematization, so the digital text analysis then aims to analyze in what way this system matches the one found based on a computational treatment of the texts. The collectivity of labelings also applies to explicit historical labels when the whole set of novels is concerned, and to a lesser degree also per work, for example, when there is a label mentioned on a title page and the same or another one in a foreword or introduction written by someone who is not the author or editor of the work. Usually, though, historical labels are not multiple because they stand for a will to mark a novel as a representative of a certain subgenre and do not aim to describe the text systematically in all its aspects. Still, cases of multiple assignments also exist for historical labels, especially on different discursive levels, but also on the same ones. An extreme case is the novel “Los bandidos de Río Frío” (1892, MX) by Manuel Payno which has the subtitle “Novela naturalista, humorística, de costumbres, de crímenes y de horrores”, including two thematic labels (“novela de costumbres”, “novela de crímenes”), two labels of intention (“novela humorística”, “novela de horrores”) and one label indicating the literary current (“novela naturalista”). Another combination that occurs three times in the bibliography is that of “novela histórica” and “novela de costumbres”.⁴⁴⁴

4.1.5.2.2 Literary Currents

546Going on to another level of subgenre labels, the ones related to literary currents, 405 works (49 %) in the bibliography and 201 works (79 %) in the corpus have such labels, so compared to thematic labels, the coverage of works is lower in both the bibliography and the corpus. This is mainly because the assignment of a work to a literary current is, above all, a critical task, so there are fewer sources than for thematic labels, as the latter draw from explicit and implicit signals and literary-historical sources alike.

547To how many works were the labels related to literary currents assigned in Bib-ACMé and Conha19?⁴⁴⁵ In the whole bibliography, there are only six different labels of literary currents. The ranks of the subgenre labels related to literary currents are the same for the bibliography and the corpus. Of the works that are associated with a literary current, most were labeled as “novela romántica”, followed by “novela realista” and “novela naturalista”. Three other subgenre labels related to literary currents occur but do only play a minor role from a quantitative point of view: “novela modernista”, “novela verista”, and “novela clasicista”. Comparing Bib-ACMé and Conha19, realist and naturalistic novels are relatively more important in the corpus than in the bibliography. Regarding the sources for the labels related to literary currents, literary-historical sources and implicit signals play the most important role.⁴⁴⁶

548As with thematic subgenres, multiple assignments also occur with literary currents. All the novels that were labeled as naturalistic novels were also marked as realist novels here because Naturalism is considered a current that evolved from and is closely related to Realism, and there are literary-historical sources for Spanish-American novels that do not differentiate clearly between the two. On the other hand, realist novels are not automatically also considered naturalistic novels, so the label “novela naturalista” is a more specific marker. Furthermore, there are eight novels in the bibliography, seven of which are also in the corpus, with the labels “novela romántica” and “novela realista” in combination.⁴⁴⁷ Three novels in the bibliography and two in the corpus are labeled both as “novela romántica” and “novela naturalista”.⁴⁴⁸ Concerning the multiple labels for literary currents, it must be remarked that it was not always easy to evaluate the literary-historical sources because the influences of several different literary currents in the works are often mentioned. Very differentiated descriptions are not fully represented in the subgenre labels. Instead, the main labels that were mentioned in the literary-historical sources were collected, for instance, the heading of a chapter in which a novel is included or the dominant literary current to which it is attributed in the source.⁴⁴⁹ Multiple labels were only assigned where different literary historians classified a novel differently or where works are clearly and repeatedly described as representatives of several currents.⁴⁵⁰

549As the literary currents are related to literary periods, looking at the publication dates of the novels that were labeled as belonging to a certain current can contribute to clarifying the question of when the different currents dominated and replaced each other. To this end, the distribution of works in the bibliography over the years per current is analyzed here, considering the publication date of the first known edition.⁴⁵¹ The analysis confirms several aspects that have been highlighted in literary-historical literature regarding the periodization of Spanish-American novels belonging to specific literary currents. First, it can be confirmed that different currents that followed each other temporarily in Europe were in vogue simultaneously in Spanish America. For instance, the majority of the naturalistic novels were published at the same time as the realist novels, approximately between 1880 and 1900.⁴⁵² Furthermore, the novela modernista came up approximately at the same time as the other currents of the late nineteenth century, around 1890. That the Romantic novel was a phenomenon that persisted during the whole century is also confirmed by evaluating the publication dates of the novels that have been labeled as romantic.⁴⁵³ The earliest works classified as novela romántica were published in 1836, and the latest one in 1905.⁴⁵⁴ The majority of the romantic novels was published between 1860 and 1886, though. The simultaneous publication of romantic and realist novels can also be attested.⁴⁵⁵ Moreover, the 1880s, which have been described as a turning point regarding the prevalence of literary currents (Rössner 2007, 200Rössner, Michael. 2007. Lateinamerikanische Literaturgeschichte. 3rd ed. Stuttgart, Weimar: J.B. Metzler.; Varela Jácome [1982] 2000, sec. 3Varela Jácome, Benito. (1982) 2000. Evolución de la novela hispanoamericana en el siglo XIX (en formato HTML). Alicante: Biblioteca Virtual Miguel de Cervantes. https://www.cervantesvirtual.com/nd/ark:/59851/bmct14z8.), can be confirmed as the decade in which the proportions of romantic novels versus novels of other literary currents shifted. However, based on Bib-ACMé, this shift can be dated more precisely here. Three-thirds of the romantic novels were published up to 1886, and three-thirds of the realist and naturalistic novels from 1888 and 1887 onwards, respectively. So instead of the year 1880 itself, 1887 is the year with a quantitative move from romantic to post-romantic currents.

550Concerning categorization tasks, the literary-historical labels are suitable for standard classification because the number of different classes is manageable, and apart from the realist-naturalistic relationship, multiple labels are the exception. At least the groups romantic versus realist-naturalistic are both represented with many members. One useful application would be to train a model with the novels in the corpus that have labels related to literary currents and use the model to also label the 55 works for which no such label was found.

4.1.5.2.3 Mode of Representation

551The next level of subgenre labels analyzed is related to the mode of representation. In Bib-ACMé, 556 works (67 %), and in Conha 19, 188 works (73 %) have a label related to the representational mode. Compared to the previously examined levels, in Bib-ACMé, these are fewer works than those with thematic labels but more than with labels related to literary currents. For Conha19, the proportion of works with a label related to the mode of representation is comparable to those with a label related to literary currents but less than the proportion of works with thematic labels. In the bibliography, 45 different labels belong to this level. Of these, 17 are assigned to at least five works.

552The distribution of the top labels on this discursive level is quite different from the one of literary currents and themes, though, because the general label “novela” is by far dominating.⁴⁵⁶ It is assigned to 49 % of the works in the bibliography and to 52 % of the works in the corpus. In the bibliography, also the labels “memorias” (assigned to 7 % of the works), “episodios” (7 %), “drama” (3 %), and “historia” (3 %) have a certain, although minor importance. In the corpus, labels that are assigned to at least 3 % of the works are “episodios”, “drama”, “memorias”, “cuento”, “historia”, and “cuadros”, but compared to the bibliography, “memorias” and “episodios” are underrepresented.

553The sources for the labels referring to the representational mode are almost exclusively explicit historical signals.⁴⁵⁷ Only the terms “memorias” and “crónica” were also part of the literary-historical discussion of the novels.

554On this level of subgenre terms, multiple assignments occur as well. In the bibliography, 123 works – i.e., 15 % of all the works in the bibliography and 22 % of the works with labels on this level – have combinations of several labels related to the mode of representation. In the corpus, there are 42 works with more than one label of this kind, which corresponds to 16 % of all the works in the corpus and 22 % of the corpus works with labels on this level. The top 5 combinations in the bibliography and the corpus are summarized in table 25.⁴⁵⁸

555In the bibliography, even the triple combinations “novela-episodios-memorias” and “novela-historia-episodios” are found repeatedly because they occur in series of historical novels, in particular the “Episodios nacionales mexicanos” by Enrique Olavarría y Ferrari. In the corpus, on the other hand, the combinations “novela-cuadros” and “novela-crónica” are more important.

556All in all, only the presence or absence of the label “novela” is meaningfully analyzable with a quantitative approach. Beyond that, the wide range of different labels related to the mode of representation lends itself more to a qualitative analysis of the respective novels.

4.1.5.2.4 Mode of Reality

557In total, there are ten different subgenre labels in the bibliography that involve the relationship of the text to reality. In Bib-ACMé, 279 works (34 %) have such a label, and in Conha19, 95 works (37 %), so this discursive level is quantitatively less relevant than thematic labels, labels related to literary currents, and labels related to the mode of representation.

558As with subgenre labels related to the mode of representation, also the labels related to the mode of reality are quantitatively dominated by a single label, the “novela histórica”.⁴⁵⁹ In Bib-ACMé, 29 % of all the works have that label, and in Conha19, 32 %. In the bibliography, also the “leyenda” plays a role, as there are 44 works with that label, which corresponds to 5 % of all the works. On the other hand, in the corpus, the label “leyenda” is underrepresented. There are only a few instances of the other labels associated with the mode of reality. In addition, the variety of labels in this group is much smaller than for the mode of representation.

559The sources for the different labels related to the mode of reality are mainly explicit signals, but for the historical novel also implicit signals and literary-historical sources are important. Other labels that are a subject of discussion in literary histories are the “novela científica” and the “novela de misterio”.⁴⁶⁰

560Clearly, in the novels, the historical perspective is the most important aspect of the relationship between the text and reality, at least when subgenre labels are evaluated. On this discursive level, only the opposition of historical versus non-historical novels is suited for a comparative, quantitative analysis of subgenres.

4.1.5.2.5 Identity

561In the bibliography, there are 25 different subgenre labels related to linguistic, geographical, and socio-cultural identity. These were assigned to 273 works (33 %) in Bib-ACMé and 101 works (39 %) in Conha19 so that the share of works having such a label is comparable to the ones with a label related to the mode of reality. Nevertheless, the variety of identity labels is larger, and this level is less dominated by a single label.⁴⁶¹

562The most important identity label is the general term “novela original”, carried by 113 (14 %) of the works in the bibliography and 36 (also 14 %) of the works in the corpus. It is followed by the labels related to the three selected countries (“novela mexicana”, “novela cubana”, “novela argentina”), by other general labels (“novela nacional”, “novela regional”), and by labels referring to the American continent (“novela americana”, “novela criolla”, “novela india”). Among the various identity labels of minor importance, there are several related to the countries’ capitals (“novela bonaerense”, “novela porteña”, “novela habanera”)⁴⁶², to specific regions or cities in Mexico or Cuba (“novela yucateca”, “novela suriana”,⁴⁶³ “novela tapatía”,⁴⁶⁴ “novela de Tabasco”, “novela camagüeyana”), and to Mexican indigenous people (“novela mixteca”, “novela azteca”). Furthermore, there are references to European regions and culture (“novela romana”, “novela franco-argentina”, “novela siciliana”, “novela kantabro-americana”, “novela andaluza”). Comparing the bibliography and the corpus, the label “nacional” is underrepresented in the corpus, while the labels referring to the three countries of Argentina, Mexico, and Cuba are overrepresented. Apart from the “novela regional”, the sources of all the identity labels are explicit signals.⁴⁶⁵

563Combinations of identity labels are not very frequent. Only the label “episodios nacionales mexicanos”, containing both the identity label “nacional” and “mexicano”, occurs 24 times in the bibliography and two times in the corpus. The label “novela original” is also combined with other identity markers, but only a few times. There are two “novelas cubanas originales”, one “novela americana original”, and one “novela mexicana original”, both in the bibliography and the corpus.⁴⁶⁶

564In a quantitative approach, the subgenres related to the linguistic, geographic, and socio-cultural identity could be analyzed in the following setups:

• novela original versus novels not carrying that label
• labels referring to the American context versus novels not carrying such a label⁴⁶⁷
• labels referring to the three countries of interest versus novels not carrying such a label⁴⁶⁸

565It is suggested here that one could concentrate on the most frequent identity label, “novela original”, and analyze the other labels as groups combining several individual labels that refer to a similar spatial context. In the corpus, a group of 36 “novelas originales”, 67 novels with a label referring to the American context, and 30, 22, and 9 novels with a label related to the Mexican, Cuban, and Argentine context, respectively, can be compared to the other novels in the corpus not carrying such labels.⁴⁶⁹

4.1.5.2.6 Medium

566Fourteen different subgenre labels in Bib-ACMé refer to a medial aspect of the novel. In the bibliography, only 47 novels (6 %) carry a label of this group, and in the corpus, only 23 novels (6 %), which is marginal from a quantitative point of view.

567In the bibliography, the three most frequent medial labels are “escenas” (12 novels), “cuadros” (11 novels), and “páginas” (8 novels).⁴⁷⁰ In the corpus, they are “cuadros” (8 novels), “páginas” (3 novels), and “esbozos” (also 3 novels). Some of the subgenre labels related to medial aspects are connected to novels of customs but also to social, realist, and naturalistic novels, as in the following titles:

• “El guajiro. Cuadro de costumbres cubanas” (1842, CU) by Cirilo Villaverde
• The series of short novels “Del natural. Esbozos contemporáneos” (1889, MX) by Federico Gamboa
• “Escenas populares. Cuadros vivos de la clase ínfima del pueblo mexicano. Novela original de costumbres” (1901, MX) by Pablo Zayas Guarneros
• “La sociedad y sus víctimas. Escenas bonaerenses” (1902, AR) by Matías Calandrelli

568Others are associated with sentimental novels, for example, “Amalia. Páginas del primer amor” (1891, MX) by José Rafael Guadalajara or “Páginas íntimas” (1895, MX) by Manuel Blanco. However, there are also historical or political novels with labels referring to a medial aspect, for instance, “Campaña y Guarnición. Escenas de la vida militar” (1892, AR) by E. Mayer, “El Señor Gobernador. Breves apuntamientos sobre cosas nacionales del siglo pasado” (1901, MX) by Manuel H. San Juan, or “Vía Crucis. Páginas de ayer” (1910, CU) by Emilio Bacardí Moreau. The sources of the labels associated with medial aspects are almost exclusively explicit signals.⁴⁷¹

569Because the frequency of these labels is low, especially in the corpus, they are not examined further in a quantitative setup but only considered for interpreting the results of other analyses.

4.1.5.2.7 Attitude

570Subgenre labels related to the attitude the author or narrator has towards what is presented in the novel are not frequent. In the bibliography, five different labels of this level were found. In Bib-ACMé, they are assigned to 57 (7 %), and in Conha19, to 32 (13 %) of the works.⁴⁷²

571The main subgenre in this group is the political novel, which is at the same time a thematic subgenre label. In the bibliography, there are 51 political novels, and in the corpus, 28. Besides the political novel, also the “novela satírica” is a label that has been assigned to novels by literary historians, but only four times in the bibliography and three times in the corpus. The other three labels, “reseña”, “novela festiva”, and “elegía”, all go back to explicit signals and occur only once each. This group of labels is not analyzed further because of its minor importance.

4.1.5.2.8 Intention

572The last group of subgenre labels considered here are the ones related to the intention the author or narrator pursues with the novel. In the bibliography, there are 13 different labels of this kind. In Bib-ACMé, 34 works (4 %), and in Conha19, 15 novels (6 %) carry an intention label, making this discursive level the least important in quantitative terms.⁴⁷³

573In the bibliography, there are four labels related to the intention that are assigned to at least five works, each: the “novela moralista” (13 works), “estudio” (7 works), “novela humorística” (6 works), and “novela didáctica” (5 works). In the corpus, “estudio” is the most frequent with five works, and the other labels only occur three times or less. The “novela moralista” and the “novela didáctica” are subgenres that are a topic in literary histories, but the other labels all have explicit or implicit signals as their source. Roughly, the different labels of this group can be classified into labels of entertainment (“novela humorística”, “novela cómica”, “entretenimientos”, “novela curiosa”, “juguete”, “comedia de carácter”, “novela de horrores”) and instruction (“novela moralista”, “estudio”, “novela didáctica”, “novela de propaganda”, “lecturas”, “novela enciclopédica”). Because this last group of subgenre labels is small, it is not examined further in the following analyses.

574Reflecting upon the empirically driven discursive model of subgenre labels that was used here to organize the subgenres contained in the bibliography and the corpus and which served to get an overview of their quantitative distributions, it can be concluded that the model is helpful in enhancing the comparability of subgenre labels. Generic labels usually refer to certain semiotic or discursive levels of a literary work, which can be quite different, for example, the spatial context versus the syntactical realization of a work, and which cannot be compared directly in a useful way.

575The specific model was created based on the metadata collected for Bib-ACMé and Conha19 and helped determine which discursive levels are relevant from a quantitative point of view. Applying it to the metadata and evaluating the resulting quantities clarifies which levels have a certain weight, either regarding the variety of labels or the number of labels assigned to the novels. The relevance of some levels was expected because they are traditionally a focus of critical concern: thematic labels and labels related to literary currents. Other levels – the mode of representation, the mode of reality, and identity – resulted in having a certain quantitative significance, as well. This highlights perspectives on subgeneric terms that have not been discussed widely so far and that are mainly derived from explicit historical subgenre signals. It has to be analyzed if and to what extent the historical practices of assigning such labels to the novels actually correspond to textual patterns. In addition, some levels are dominated by individual subgenre labels, which trigger the overall quantitative importance of the respective level, for example, the general term “novela” on the level of the representational mode or the “novela histórica” on the level of the reality mode. All in all, it became clear that compared to the overall variety of subgenre labels, relatively few of them are very frequent.⁴⁷⁴

576In Bib-ACMé, 89 labels, which corresponds to 72 % of all the different labels in the bibliography, are only assigned to up to 9 works. In Conha19, this is true for 73 labels, which is 81 % of all the different labels in the corpus. In the bibliography, only 13 labels are assigned to 50 works or more.⁴⁷⁵ Regarding the sources of labels, there are 108 different explicitly signaled labels and 34 different literary-historical labels. All these numbers are based on normalized terms, so variances in spelling, formulation, and syntactic constructions are not causing the broad range of subgenre labels. Following these numbers, it can be assumed that in the historical practice of explicitly labeling novels, in particular, creativity and the emphasis on individuality are important factors besides the wish to mark a work as belonging to some established or widely practiced subgenre. In literary-historical approaches, on the other hand, subgenres are studied independently of their quantitative weight in the whole production of novels. In comparison, a digital quantitative approach to the subgenres sets different focuses of analysis, or rather, it is only usefully applicable to a subset of the whole range of generic signals and classifications.

577Some aspects of the nature of generic terms that complicate classificatory and comparative studies of subgenres are not solved by applying the discursive model. Actually, the semiotic and discursive models of generic terms are above all descriptive models and do not claim to tackle these issues:

• generic terms can be semantically complex and loaded with meaning on several discursive levels (the term “novela histórica”, for instance, refers to certain themes but also to the relationship between the text and reality)
• even if generic terms refer to the same discursive level, they can have a different, narrower or broader, semantic scope, which can lead to relationships of inclusion or overlap (the “novela abolicionista”, for instance, can be understood as a special kind of “novela social”, while both are terms on the thematic level); in addition, no clear line is drawn between generic and subgeneric terms (the term “novela”, for example, is more general than “novela epistolar”, but both refer to the mode of representation)
• the degree of relationships between a work and the generic terms is not controlled (none, just one, or several generic terms of different levels or of the same level can be associated with one work) so that the generic information is not necessarily complete on all the levels or unique on individual levels (for instance, the literary current of a part of the novels was not determined; on the other hand many novels had multiple labels, for example, on the thematic level)
• nothing is said about the sources of the generic terms and how they are defined. As for general language terms, also the semantics of generic terms is determined by their use, synchronously and diachronically, by numerous different speakers, writers, and critics, at least if the terms are not restricted to selected scholarly defined ones; on the other side, assigning terms to the different discursive levels involves presupposing that they cover certain semantic aspects so that another level of uncertainty is introduced

578When classificatory tasks are designed based on the subgenre labels in the bibliography and the corpus, which have a collective background, are potentially semantically multifaceted, historically bound, and neither exhaustive nor unique, and when these tasks are based on how the labels are ordered in the discursive model of subgenre terms, this constitutes a simplifying choice of perspective on a more complex system of generic relationships. To conclude this section of metadata analysis of the subgenres, only the constellations of labels selected for further analysis are presented in more detail, i.e., differentiating the overviews also by country, time period, and related to corpus-specific metadata and characteristics of the texts. This is done in the following subchapters.

4.1.5.3.1 Primary Thematic Labels

580This chapter analyzes the proportions of novels with a particular primary thematic subgenre label in Bib-ACMé and Conha19. Here, only the top primary thematic subgenres are analyzed in detail, summarizing the remaining ones as “other”. Primary thematic subgenres are considered “top” if they cover at least 5 % of the works in the corpus. Although the proportion may be higher in the bibliography, the coverage in the corpus is decisive because, in the end, only the full text of the novels in the corpus is analyzed. First, the general proportions of novels having a certain primary thematic subgenre label in the bibliography and the corpus are assessed.⁴⁷⁶ This general overview serves as a reference point for the overviews differentiating by further parameters such as country, time period, etc. In Bib-ACMé, 16 % of the works do not have any thematic subgenre label, whereas in Conha19, all the works were labeled thematically. In both resources, historical novels are most frequent, followed by sentimental novels. In the corpus, novels of customs, social, and political novels have more weight than in the bibliography.

581Next, the primary thematic subgenres are analyzed by the three countries Mexico, Argentina, and Cuba.⁴⁷⁷ As for historical novels, in the bibliography, they are overrepresented in the Mexican works (38 %). In the Argentine novels, and even more in the Cuban ones, historical novels are underrepresented (19 % and 12 %, respectively). In the corpus, the distribution of historical novels is more balanced by country. However, even there, the Mexican works have a greater proportion of historical novels (32 %) than the Argentine (23 %) and Cuban ones (20 %). Sentimental novels are slightly underrepresented in the Mexican and Argentine works contained in Bib-ACMé (19 %, respectively), and clearly overrepresented in the Cuban works (33 %). Also for this subgenre, the distribution is a bit more balanced in the corpus. The novels of customs are proportionally overrepresented in the Cuban works in the bibliography and also in the corpus. Social novels are above average in Argentine novels and below average in novels of other countries, both in Bib-ACMé and Conha19. Finally, in the bibliography, political novels are a bit overrepresented in the Argentine works and underrepresented in the Mexican and Cuban works, slightly in the first and more in the latter case. In the corpus, in contrast, there are relatively more political novels from Mexico, while the Argentine novels correspond to the average, and for Cuba, there are no political novels at all. So regarding the distribution of primary thematic subgenres by country, the differences between Argentine, Cuban, and Mexican works that are visible in the bibliography are, for the most part, also reflected in the corpus, where they are balanced out a bit. In summary, there are more historical novels from Mexico, more sentimental novels and novelas de costumbres from Cuba, and more social novels from Argentina. In the following, the distribution of primary thematic subgenres is given per decade, for Bib-ACMé and Conha19, respectively.⁴⁷⁸

582The most important point to conclude for the bibliography is that all of the top primary thematic subgenres occur in almost all of the decades. The political novel is missing in the 1830s and 1840s, and the social novel in the 1840s. In the corpus, there are some more decades without works of individual subgenres, especially the first and last decades. There is only one novel of customs from the 1830s and one novel of a different subgenre. There are only three novels from the 1910: one political novel, one social novel, and one historical novel. In the more central decades, no political novel is included in the 1840s, 1860s, and 1870s. No social novel is present in the 1840s and 1870s. In contrast, historical novels, sentimental novels, and novels of customs are represented in all the central decades of the corpus.

583Regarding the relative amount of different subgenres over the decades, in the bibliography, the proportion of historical and sentimental novels is higher up to the 1860s. After that, especially the social novel gets more significant. When compared to the bibliography, in the corpus, especially the 1860s stand out as a decade with an over-proportional number of historical novels and the 1890s with an above-average number of social novels. The change of proportions of primary thematic subgenres over time becomes clearer if the two periods before 1880 and in or after the year 1880 are compared.⁴⁷⁹

584In Bib-ACMé, in particular, the number of sentimental novels is lower after 1880 (17 % instead of 28 % before), and the number of social novels higher (14 % instead of 6 %). The political novel, which is in general not very frequent, also raises from 1 % of the works before 1880 to 4 % after that year. Furthermore, the share of works without any thematic subgenre label is considerably higher in the later period (19 % instead of 10 %). On the other hand, the differences between the proportion of historical novels and novels of customs are not very big between the two periods (4 % in the case of the historical novels and 3 % for the novels of customs). In general, and regardless of the shifts of proportions, the absolute number of novels of all the primary thematic subgenres is higher in the second period than in the first one.

585In Conha19, in contrast, the absolute number of sentimental novels and novels with other subgenres drops for the period after 1880. On the other hand, the relative increase of social and political novels is even higher than in the bibliography. There is almost no change in the proportion of novels of customs between the two periods, but the relative importance of the historical novel drops more in the corpus than in the bibliography. It can be assumed that the change of proportions of the various thematic subgenres over time reflects the preferences of the different literary currents, which overlap and succeed each other during the nineteenth century. The romantic current is apparently more closely related to semantic and historical themes, and the realist and naturalistic novels to social and political topics. However, the historical novel persists as an important subgenre throughout the whole century, and also the novel of customs remains relevant. It is striking that the only two thematic subgenres that are frequently explicitly named in the novels’ subtitles, the novela histórica and the novela de costumbres, are the ones that are less subject to change in terms of relative quantities over time, which means that they are less bound to the dominant literary currents than the other thematic subgenres.

586In what follows, the distribution of primary thematic subgenres is analyzed in relationship to corpus-specific metadata and text characteristics, namely the prestige of the texts, the narrative perspective, the continent and time period of the setting, and text length. The first of the corpus-specific aspects to be analyzed is the prestige of the texts.⁴⁸⁰ Historical novels are equally present both in high- and low-prestige novels. The sentimental novels, in contrast, are underrepresented among the high-restige novels and overrepresented in the low-prestige ones. Interestingly, of the novels of customs, relatively more are classified as high- than as low-prestige. The social novels are slightly overrepresented in the high-prestige group and the political novels in the low-prestige one. Especially in the latter case, the difference is not considered significant because of the comparatively low number of political novels in the corpus. It can be concluded that one of the primary thematic subgenres, the sentimental novel, is clearly marked as having a low-prestige branch, but none of the other subgenres sticks out as a particular high-prestige subgenre.

587Considering the narrative perspective, the great majority of historical novels are written in the third person. Social novels also primarily have a third-person narrator. The first-person narrator is most frequent in sentimental novels but also overrepresented in the group of political novels. In the novels of customs, both narrative perspectives are almost equally in use. It is noticeable that 20 % of the novels with a first person narrator have a primary thematic subgenre that is part of the “other” group, so apparently, this perspective is favored in some of the subgenres that are less frequent.⁴⁸¹

588For the continent of the setting, there are also correlations with some of the primary thematic subgenres.⁴⁸² Almost half of the novels set in Europe (46 %) are sentimental novels. Novelas de costumbres rarely have a European setting – they make only up 4 %. The political novels are exclusively set in America. There are social novels set on both continents, but also for that subgenre, the American setting is more prominent. Only the proportion of historical novels is almost the same for the American and the European setting. A hypothesis that follows from this is that European models of sentimental novels were more often just copied and that models of other subgenres were more often adapted and appropriated to reflect the local circumstances, or even that they developed more into independent varieties of the subgenres.

589Correlations between the different primary thematic subgenres and the time period of the setting become visible, as well.⁴⁸³ Obviously, most of the novels set in the past are historical novels (85 %), and there are no primarily sentimental or political novels in this group at all. Most of the novels set in the recent past are also historical novels (68 %), but here there are also some representatives of the other subgenres, in particular sentimental novels and novels of customs. Novels with a contemporary setting are led by the sentimental subgenre, novelas de costumbres, and social novels. However, also 7 % of the novels with a contemporary setting are primarily classified as historical novels, which shows how widely the concept of historicity was interpreted in the novels.

590Analyzing the lengths of the works of different primary thematic subgenres reveals that all the groups overlap.⁴⁸⁴ There are short novels among all the subgenres, and differences become only visible regarding the median and the variance of length. The longest novels are historical novels, followed by novels of customs and sentimental novels. Apart from the outliers, the subgenres with the highest median length are historical novels (89,000 tokens) and social novels (57,000 tokens), and the ones with the lowest sentimental novels (40,000 tokens) and political novels (43,000 tokens). The novels of customs lie in between with a median of 48,000 tokens, so the median sentimental and political novels are less than have as long as the median historical novel.⁴⁸⁵ A test for significance shows that the difference in length is significant for the historical versus all other types of thematic subgenres, but not for the other pairs of subgenres.⁴⁸⁶ Analyzing the two central quartiles, the variety of length is greatest for historical novels (50,000 to 139,000 tokens) and novels of customs (34,000 to 94,000 tokens) and lowest for social novels (37,000 to 76,000 tokens). The ratio of variances and, thereby, the difference in statistical variance between selected pairs of subgenres is greatest for the historical versus the political novel, the historical versus the social, and the historical versus the sentimental novel and lowest for the sentimental novels when compared to the novels of customs.⁴⁸⁷  As with the development of the number of works associated with the primary thematic subgenres over time, also in terms of text length, the two explicit subgenres, novela histórica and novela de costumbres, stick out as the subgenres with the longest novels on the one hand (considering the outliers and upper fences) and the greatest variety of length on the other hand (in terms of the ranges of the two central quartiles). It can be hypothesized that these are signs of long-lived subgenres for which there is historical variability and for which experimentation to the extremes has taken place. When statistical differences in the distributions of lengths are considered, the historical novel is the one that is significantly longer than the other types of thematic subgenres and, at the same time, the one with a comparatively bigger variance in text length.

4.1.5.3.2 Primary Literary Currents

591In this chapter, the primary literary currents to which the novels in the bibliography and the corpus were assigned are analyzed.⁴⁸⁸ In the bibliography, the literary current is only known for half of the works (49 % or 405 novels). Normally, this information is available from literary-historical accounts of the novels and not from explicit historical subgenre labels. As many of the novels in the bibliography have not been the focus of critical literary-historical work, this information is missing. In the corpus, the number of novels for which the literary current is known is much higher (79 % or 201 novels), which again shows that the corpus represents works that are better known and more canonized than the works in the whole bibliography.⁴⁸⁹ Here, only the primary literary currents assigned to the novels are analyzed. Secondary literary currents are not taken into account here. Examples are novels that are borderline cases or mixtures of romantic and realist novels but also all naturalistic novels that can as well be understood as realist novels in a general sense.

592In the bibliography, the largest group of novels by literary current are the romantic novels, with 32 % or 263 novels. This is also the case in the corpus, where the romantic novels have a share of 45 % or 116 novels. In the bibliography, the amount of realist and naturalistic novels is equal (8 % or 66 novels each), while in the corpus, there are a bit more naturalistic works (18 % or 45 novels) than realist ones (14 % or 35 novels). Other literary currents only have a minor quantitative importance in both the bibliography and the corpus. Compared to the bibliography, the relationship between romantic novels on the one side and realist and naturalistic novels on the other is more balanced in the corpus. However, it is difficult to say if this turns the corpus further away from the population of novels or not because the literary current is unknown for so many novels in the bibliography.

593When the proportions of literary currents are viewed by country, some differences become visible.⁴⁹⁰ In Bib-ACMé, the share of novels for which the literary current is unknown is similar in the three countries, but in the corpus, it is highest for the Cuban novels and lowest for the Mexican ones, with the Argentine novels in-between. The higher number of unknown cases for Cuba reflects that apart from a few very well-known works, novels from that country appear to be less studied. Regarding the distribution of the different literary currents by country, in the bibliography, approximately one-third of the novels are romantic ones, also by country. In Conha19, however, the proportion of romantic novels is bigger for Mexico (55 %) than for the other two countries. Most realist novels come from Mexico, in the bibliography, and also in the corpus. The naturalistic novels have the most weight in Argentina, both in Bib-ACMé and Conha19.

594Next, the distribution of the novels’ primary literary currents over time is analyzed, considering the number of works per decade in the bibliography and the corpus, respectively.⁴⁹¹ What becomes clearly visible is that the romantic novel is a phenomenon that persisted throughout the whole nineteenth century. It was the dominant current up to the 1870s and only gradually gave way to the other currents in the following decades. The first realist and naturalistic works are found in the 1860s and 1870s, but it was from the 1880s onwards that these two currents gained more weight. If one compares the bibliography and the corpus, the proportions are similar in both, only that there are relatively more realist and naturalistic novels in the corpus. If only the relationship between realist and naturalistic novels is examined, the latter are a bit overrepresented in the corpus in the 1890s and 1900s. Considering the period before and after 1880,⁴⁹² the dominance of the romantic novel in the early period again stands out, both in Bib-ACMé and Conha19. Moreover, even after 1880, the number of romantic novels is higher (in the bibliography) or comparable (in the corpus) to the amount of either realist or naturalistic works. Concerning the number of works for which the literary current is unknown, it is very high in the bibliography for the period after 1880, amounting to 59 % of all the novels. In the corpus, the proportion of works with unknown currents is more balanced before and after 1880. As was seen before in chapter 4.1.3.1, where the general number of works was assessed independently of the subgenres, the number of works doubled from the 1870s to the 1880s. For the 1880s and 1890s, the corpus contains approximately 30 % of the novels that are registered in the whole bibliography, but for the 1900s, only 20 %. So the high proportion of novels after 1880 in Bib-ACMé for which the literary current is unknown shows that there is a mass of works that is still mostly unexplored.

595We now turn to the analysis of the corpus-specific metadata, starting with the proportions of high- and low-prestige novels in Conha19 for the different literary currents.⁴⁹³ The most striking difference is the higher proportion of novels without a literary current label in the low-prestige group, which is at 34 %, compared to 16 % in the high-prestige group. This result is mainly due to the way that the prestige of the novels was determined. Only works of which at least one new edition was published between 1960 and 2020 were considered high-prestige novels. The high proportion of novels without known literary current in the low-prestige group is just another perspective on them as a group of texts which has been largely forgotten or has not been investigated in the last 60 years.⁴⁹⁴

596The next aspect that is analyzed is the kind of narrator that the novels related to the different literary currents have.⁴⁹⁵ In general, the third-person narrator is much more frequent in the corpus than the first-person narrator: there are 44 novels (17 %) with a first-person narrator and 212 novels (83 %) with third-person narrator. Interestingly, the proportion of realist novels in the first-person narrator group is considerably higher than in the third-person group. On the other hand, naturalistic novels are overrepresented in the third-person group, and also the romantic novels have a higher proportion in the latter one. In his overview of the history of Latin-American literature, Dill introduces the realist novel as follows: “Die Realität wurde nicht costumbristisch-dokumentarisch kopiert, vielmehr mit literarischer Inszenierung ein Ähnlichkeits- oder Realitätseffekt (effet du réel) durch einen heterodiegetischen Erzähler erzeugt, der dem impliziten Leser die neue Gesellschaft und adäquate Verhaltensweisen modellierte” (Dill 1999, 159Dill, Hans-Otto. 1999. Geschichte der lateinamerikanischen Literatur im Überblick. Stuttgart: Reclam.). So the third-person narrator is usually seen as typical for the realist novel, but the overview of the novels in Conha19 suggests that there is a subgroup of realist novels narrated in first person and that the choice of this narrative perspective is a factor that differentiates the realist from the naturalistic novel, where the third person is proportionally more important. Checking which ones are the realist novels with a first-person narrator reveals that they belong to several different thematic subgenres: the sentimental novel, the historical novel, the novel of customs, and the political novel, so one specific type of thematic subgenre is not responsible for the number of realist novels with first-person narrator. In addition, they were written by different authors from the three countries.⁴⁹⁶

597The next property to be analyzed is the continent of the novels’ setting.⁴⁹⁷ Comparing the proportions of novels with an American and European setting reveals that the romantic and also the realist novels are preponderant for the American setting. Naturalistic novels have a similar proportion in both groups, which means that the European setting is relatively more important in them when compared to romantic and realist novels. Apparently, the French origin of the naturalistic current had an influence on the choice of the setting in some cases. Nevertheless, the numbers of the continent of the setting must be interpreted with caution because the great majority of the novels are set in America (90 %, 231 novels) and only a few in Europe (9 %, 24 novels).⁴⁹⁸ Furthermore, the numbers for Europe are biased because the proportion of novels for which the literary current is not known is considerably higher than in the case of an American setting. Here there is a correlation with the countries of origin of the novels because knowledge about the literary current is missing for many Cuban novels, and novels from Cuba are also the ones that have a European setting more often than novels from Argentina or Mexico.

598How does the time period of the setting relate to the literary currents? Clear tendencies are visible in this respect:⁴⁹⁹ the romantic novel has high proportions in all three categories but is overrepresented in the groups of novels with a setting in the past or recent past. The realist novel primarily has a contemporary setting or one in the recent past, and the naturalistic novel is, above all, set in the present. That the romantic novel is inclined towards the past is in line with the fact that the historical novel was a very popular thematic subgenre in that current. However, the results also show that the romantic novel is a multi-faceted phenomenon because all the different time periods of the setting are covered by it to a significant extent. That the naturalistic novel primarily has a contemporary setting confirms its role as the type of novel that served to depict the process of social, economic, and technological modernization going on in the Spanish-American countries in the last decades of the nineteenth century. In terms of text style, the preferences for certain time periods of the setting that the novels associated with the different literary currents have might influence, for example, the usage of temporal expressions or verb forms, which is an aspect that needs to be investigated further.

599Finally, the differences in text length between the novels of the three main literary currents are analyzed.⁵⁰⁰ The romantic novels have a median length of 66,000 tokens, the realist novels of 56,000, the naturalistic novels of 51,000, novels with other literary currents of 62,000, and novels with an unknown literary current of 42,0000 tokens.⁵⁰¹ Of most interest are the differences between the romantic, realist, and naturalistic novels because the group of novels with another literary current only consists of 5 novels, and the “none” group is probably mixed. A test for significance shows that the differences in length between the romantic, realist, and naturalistic novels are not statistically significant, though.⁵⁰² Even if the differences in median text length are not significant, the group of romantic novels has a bigger variance in length than the other two currents, which are more similar regarding the variation of the novels’ length.⁵⁰³ That the length of romantic novels varies more is plausible because the group of romantic novels in the corpus is bigger (116) than that of the realist (35) or naturalistic (45) novels, but it can also be a sign of the variability of romantic novels in themselves. Romantic novels can be part of several distinct thematic subgenres, of which, for example, historical novels can be very long and sentimental novels are usually shorter. In addition, romantic novels spread over the whole nineteenth century, whereas the realist and naturalistic novels were concentrated in the decades 1880 to 1910. The general comparison of text length by decade that was made in chapter 4.1.3.2 above showed that the variability of text length was lowest from the 1870s to the 1890s, and very long novels were the exception in these decades. This suggests that the form of the novel in terms of length was more stabilized in the last decades of the nineteenth century. In Spanish America, romantic novels participated in all the phases that the novel as a genre underwent in the nineteenth century, so they were variable, whereas the realist and naturalistic novels were anchored in a more precise literary-historical moment. That the variance of the length of the three major currents is smallest for the naturalistic novels can be interpreted as a sign for the relatively uniform and defined form of the novels participating in that subgenre.

600Finally, the results of the metadata analysis on the novels in the bibliography and the corpus will be briefly summarized and evaluated here in overview. All in all, the proportions of works in the corpus and the bibliography are very similar in most aspects. In the corpus, some balancing can be noted, for example, concerning the distribution of subgenres in general or by country. This means that the corpus deviates from the distributions in the bibliography to mitigate certain tendencies of imbalance and to achieve a dataset that is more suited for an analysis of subgroups, even if it doesn’t represent the proportions of the sampling frame closely. In addition, there are effects that are due to the fact that the sources for the corpus are limited because specific works are difficult to access in digital format. As a whole, the corpus is not entirely balanced in all its aspects. Its closeness to the distributions of subgroups in the bibliography is bigger than their balance in terms of similar numbers in the corpus itself, for example, by subgenre. The overviews of the subgenres selected for analysis highlighted which metadata and also textual factors are connected to the subgenres. They pointed out some interdependencies (e.g., regarding narrative perspective or text length). Some influencing factors only became visible from a single perspective, whereas others manifested themselves repeatedly, such as the very long Mexican historical novels, for instance.

601Such an analysis of metadata in connection with the subgenres is of interest in itself, but it also provides useful background knowledge for further textual analysis to be able to evaluate possible biases in the results. Especially, differences between the works of the three countries and also some chronological trends became visible. When conducting analysis on the texts, it must be kept in mind that the sampling frame and also the corpus are not entirely balanced datasets – instead, there are many different factors influencing each other and some asymmetries that are already characteristic of the underlying set of novels. In the next chapter, the textual features used for text analysis and the methods used to categorize the novels by subgenre are presented.

4.2.2.1.1 Thematic Subgenres

680This chapter presents the results of the classification of novels by thematic subgenres. First, the distribution of the primary thematic subgenres in the corpus is presented in figure 51.

681The four most frequent types of primary thematic subgenres are the novela histórica, with 67 novels; the novela sentimental, with 55 novels; the novela de costumbres, with 50 novels; and the novela social, with 45 novels. Smaller groups are the political novels (13), the science fiction and crime novels (5 each), and the anti-slavery novels (4). The latter can also be subsumed under social novels but are marked separately because historically and by literary critics, they were categorized with this more specific label. Twelve novels participate in other primary thematic subgenres, which all occur only one to three times. The classification of primary thematic subgenres is conducted for the following selected subgenre constellations:

• novela histórica versus other novels
• novela sentimental versus other novels
• novela de costumbres versus other novels
• novela histórica versus novela sentimental
• novela histórica versus novela de costumbres
• novela sentimental versus novela de costumbres

682With these six constellations, the three most frequent primary thematic subgenres are analyzed. They correspond to the subgenres that developed in the romantic period and to the ones that literary historians have often distinguished as relevant thematic subgenres in nineteenth-century Spanish America.⁵⁷⁸ The presentation of the results starts with the topic features and then goes on to the MFW-based ones. The results for the topic features are discussed in more detail, inspecting feature importances, numbers of correct and wrong classifications, and topic profiles of selected novels. For the MFW, the overall classification results are presented to see how well these features work, but a deeper analysis of the features and novels is left as a future task.

683First, it is analyzed which classifiers worked best to classify the thematic subgenres in all six constellations.⁵⁷⁹ For this, the accuracy and F1 score values for all classification runs were evaluated, and top values, mean values, and the standard deviation (SD) were calculated. In total, 144,000 classification runs were considered for the topic features, including all topic model parameter constellations, topic modeling repetitions, repeated data selections, and 10-fold cross-validation.⁵⁸⁰ The results for the topic features are shown in table 36.

684In the case of the topics, SVM and RF work best with a mean accuracy of 0.80, but the results for KNN are also quite good, with a mean accuracy of 0.77. The relatively high standard deviation indicates that either the results vary for the different topic model parameters or for the different subgenre constellations, or both. For the following, more detailed analyses of the results, SVM is chosen because the standard deviation is lower than for RF, even if there is only a difference of one percentage point. Once the best classifier is chosen, the next step consists in evaluating the feature sets in more detail. The goal is to find out which influence the different parameter values have on the results and to find the ones that are best suited for the classification of primary thematic subgenres. The classification results for the different topic modeling parameters are presented in figure 52.

685The top mean accuracy of 0.81 is reached with several different combinations of topic numbers and optimization intervals: with 60 topics and an interval of 500, 80 topics and an interval of 250, and 90 topics and an interval of 50, 250, or 2,500.⁵⁸¹ If one evaluates the general development of the results for different topic numbers and optimization intervals, for some intervals, there are trends, and for others, oscillations. With an interval of 100, for example, the score is highest at 60 topics and decreases for higher topic numbers. The score for no optimization at all rises up to 70 topics and then falls again. For intervals of 50 and 250, the scores get higher with more topics. Except for the interval of 2,500, which means very little optimization, all the scores drop with more than 90 topics. This suggests that the models get too specific to detect subgenres if they have more topics. Otherwise, it appears that the number of topics interacts with the optimization interval so that specific combinations which produce topics and topic distributions that are not too general and not too specific work well to model thematic subgenres.⁵⁸²

686Overall, the differences between the mean accuracies using different kinds of topic features are very low, ranging from 0.78 for 80 topics and an interval of 500 to the highest value of 0.81. This means that the higher standard deviation observed in the general accuracy mean for topic features is not due to the different topic modeling parameters but must be connected to the different kinds of thematic subgenres or to the influence of individual novels. With such similar results, the decision to choose a certain combination of topic numbers and optimization intervals does not only need to be based on the top mean accuracies but can also take into account the kind of topics that are likely to result from a model with a specific combination of parameters. Here, 90 topics are chosen because they lead to the highest mean score several times and also because the topics are more specific than they would be in a topic model with a lower number of topics so that the resulting text types can be described in more detail. As for the optimization interval, a number of 250 is chosen so that the topics are still relatively balanced regarding the different weights they have in the corpus and individual texts.

687Having decided on the kind of classifier to use and on the specific kinds of features, the classification results for the different constellations of thematic subgenres can now be inspected. The results for the topic models with 90 topics and an optimization interval of 250, using SVM as a classifier, are shown in table 37. These average results are based on 500 classification runs for different data samples and the 10-fold cross-validation.

688The constellation novela histórica versus novela sentimental has the highest mean accuracy with 0.89. The second best result is reached for novela histórica versus novela de costumbres, and the third for novela histórica versus other. This shows clearly that the historical novel is the subgenre that can best be distinguished from the other two thematic subgenres that were chosen here but also from the big group of all other kinds of novels, which include not only sentimental novels and novels of customs, but also crime novels, science fiction novels, social and political novels, and more. This result confirms the expectation that the novela histórica as a subgenre, which is by convention firmly established in Argentina, Cuba, and Mexico in the nineteenth century, also is united textually. Although there are different subtypes of historical novels dealing with different past (or not-so-past) epochs and different specific topics, the subgenre can be classified well by using topic features. Furthermore, it is astonishing that it can be separated best from the sentimental novel because there are many historical novels with a sentimental plot and also because most historical and sentimental novels in the corpus are romantic novels that have much in common. This result confirms what most literary histories of nineteenth-century Spanish-American novels describe (see chapter 3.1.3 on the novela sentimental): that the sentimental novel is an important and recognizable thematic subgenre as well. After the historical novel, the sentimental novel has the second-best results. The mean accuracy for novela sentimental versus other novels is at 0.81, which is only slightly worse than for the historical versus the other novels. Apparently, the sentimental novel can be separated well from other subgenres, even if there are many novels that also have a sentimental plot. However, the results also show that this is not so easy if sentimental novels are compared to novels of customs. Here, the mean accuracy is at 0.77, which is the worst result for the direct comparisons of subgenres. This confirms what has has been remarked by Janik (2008, 67–68Janik, Dieter. 2008. Hispanoamerikanische Literaturen. Von der Unabhängigkeit bis zu den Avantgarden (1810–1930). Tübingen: Narr Francke Attempto.): that the description of local customs in novels and the expansion of the cuadros de costumbres to longer fictional narrative works in the form of novels of customs needed the sentimental plot as a basic structure, at least in the romantic variant of the subgenre. The results are even worse when the novels of customs are contrasted with all other novels (with an accuracy of 0.71), which can be interpreted as a sign that costumbrista elements are not only found in sentimental novels but also in other subgenres. As the group of other novels also contains all the social and political novels that are associated with the realist (or naturalistic) current, it might well be the case that this causes lower accuracy. The two results for the novels of customs can be interpreted as a sign of mixtures of sentimental novels and novels of customs, but also similarities of the novels of customs with other social and political novels in terms of the topics that they treat. Thus, for example, Kohut has noted that there are not only romantic, but also realistic and naturalistic novelas de costumbres (Kohut 2016, 196Kohut, Karl. 2016. Kurze Einführung in Theorie und Geschichte der lateinamerikanischen Literatur (1492–1920). Berlin: Lit Verlag.). However, mean accuracies of over 0.70 do mean that even for the novels of customs, a model can be learned that classifies the texts better than by chance, only that this thematic subgenre is textually less coherent than the others. Overall, the standard deviations for the accuracies (and F1-scores) are still quite high, which means that the results also depend on the individual novels that are selected for the training and test sets.

689The next step consists in having a look at the coefficients that have the highest values in each subgenre constellation to inspect the characteristics of the text types that were learned by the classifiers. Starting with the historical novels, figure 53 shows the top 25 coefficients for the classification of novela histórica versus novela sentimental. As these weights are based on a specific topic model, they represent the average weights of 100 classification runs (ten different selections of novels and ten cross-validation runs, as explained in chapter 4.2.2.1 above on the classification workflow). In the plot, the topics are identified by the number they got in the topic modeling process (0 to 89 for the 90 topics) and the three words that are most important in each topic.⁵⁸³

690The bars with negative values are the topics characteristic of historical novels and those with positive values of sentimental novels.⁵⁸⁴ The most important topic for distinguishing historical from sentimental novels is “baile-salón-fiesta”, which is a topic concerned with a ball situation that is typical for sentimental novels. The second most important distinguishing topic is “sangre-voz-arma”, this time for historical novels. The topic covers a situation in a fight or battle. The word clouds of the two topics are given in figure 54. It is interesting that both top topics describe situations that are typical for the two subgenres in question and that can be considered characteristic scenes and elements of the plot. As the topics consist only of nouns and verbs are not included, the dynamic aspect of the topics does not stand out directly, but even the nouns describe actions if they are analyzed together. In the case of the ball situation, for example, the words “mirada”, “brazo”, “sonrisa”, “conversación”, and “palabra” point to acts of the characters. In the fight situation, also the word “brazo” occurs, as well as “cabeza”, “golpe”, “grito”, “espada”, “pistola”, “puñal”, “combate”, “lucha”, “herida”, and “muerte”, so in both cases parts of the body and physical actions are involved. The ball topic contains also words that describe the setting of the situation, for example, “salón”, “sala”, “pieza”, “teatro”, “flor”, “música”, “piano”, “gente”, “reunión”, “concurrencia”, “sociedad”. That the word “teatro” is part of the ball topic could be a sign that similar thematic elements are relevant if the situation is a theater performance as a social event. The fight topic is less descriptive and contains more words related to emotions and sounds: “miedo”, “peligro”, “terror”, “venganza”, “ruido”, and “silencio”.

691All in all, the top feature weights for historical versus sentimental novels are balanced in that there are as many features that are distinctive for the historical as for the sentimental novel. They also alternate by importance in the range of top 0 to 25. Besides the fight topic, other highly weighted topic features for the historical novels are concerned with crime and prison (“crimen-cárcel-preso”), politics and administration (“gobierno-pueblo-ministro”, “policía-gobierno-pueblo”, “virrey-oidor-señoría”), monarchy and nobility (“príncipe-rey-palacio”, “conde-condesa-señorito”), the countryside (“caballo-camino-árbol”, “caballo-hacienda-compadre”), church (“cura-iglesia-sacerdote”), and another topic about battles (“soldado-enemigo-fuego”). These topics point to different subtypes of historical novels dealing with colonial and contemporary history and different urban and rural surroundings.

692Among the top topics for historical novels, there is also one which is a bit more difficult to interpret: “v_d-instante-corazón”. It is dominated by the historical form of address “vuestra merced”, which is actually not a noun but was misinterpreted by the linguistic tagger because of the abbreviation “vd”. The topic, which is shown in figure 55, is a bit more abstract and mixed than the other top topics for historical novels. It seems to be about physical (“pecho”, “pena”, “doctor”) and mental states (“alma”), feelings (“temor”, “duda”, “amor”, “placer”, “satisfacción”, “esperanza”, “corazón”, “desgracia”, “felicidad”), and thoughts (“pensamiento”, “idea”), but it also contains other words that are not related to these aspects. The words “instante”, “puerta”, and “vista” together with “fisonomía”, “cabeza”, “rostro”, “gracia”, and “virtud” seem to describe a moment in which someone is observed and the appearance of a person is described. The impression that this topic gives is that it joins together several thematic aspects that are not prototypical for historical novels but are accompanying material related to the plots and the representation of the characters’ feelings.

693Besides the ball topics, the other top distinctive topics for sentimental novels cover private conversation (“voz-cabeza-mirada”), love (“amor-corazón-alma”, “idea-pasión-alma”), family relationships (“mamá-muchacho-papá”), material aspects (“v_d-país-negocio”), interiors and meals (“mesa-puerta-sala”, “mesa-copa-vino), art (“poeta-maestro-artista”), and travel or movement (“coche-carruaje-cochero”). The three topics “tía-do-aire”, “ojos-años-manos”, and “ojos-hija-años” need a closer look. In the first one, which is visualized in figure 55, the word “do” sticks out because it is not a regular Spanish word. The novels in which this topic has the highest probability were checked, and the “do” is a spelling error (instead of “de”) that remained unnoticed in some of these novels.⁵⁸⁵ Moreover, the other top words of this topic suggest that it can be interpreted as being about outward appearances and physical encounters because it contains words referring to people (“muchacho”, “hijas”, “mujeres”, “viejita”, “virgen”), body parts (“cabeza”, “ojos”, “labios”), clothes (“camisa”, “pana”, “abrigo”, “alfiler”), meeting places (“sitio”, “calle”, “playa”, “pieza”, “sillas”, “cama”, “flores”, “vela”), and also the word “besos” and “belleza” occur in it. The word “aire” can then be read in its sense of “look” and “airs and graces”. The other two topics that were not directly clear from the three top words (“ojos-años-manos” and “ojos-hija-años”) can be interpreted better when more of the top words are considered. The first one has mixed meanings and appears more abstract than the other top topics for sentimental novels. It can be described as treating views, nature, body, and time. The second one is about looking, also talking, mainly female persons, and emotions. In the first case, the word “años” probably refers to time, and in the second case, it may refer to age.⁵⁸⁶ Summing up the findings about the topics that are most important for the classifier to distinguish historical from sentimental novels, it can be stated that most topics can be interpreted easily from the top words and some by examining them more closely. The topics of the historical novels are more easily recognizable as themes, whereas some topics of the sentimental novels are more subtle or mixed but also make sense for the concept of the subgenre. For the aspects covered by the topics that are at first sight less clear, it may have been favorable to also have verbs and adjectives and not only nouns in the topics. In what follows, not all the topics can be analyzed in the same depth as the ones that are most distinctive for historical versus sentimental novels, and the presentation of the results concentrates on the three top topic words unless there are aspects that need to be clarified or are of special interest. Nevertheless, a look into more top words of the topics and, if needed, also into the texts of the novels shows that the topics can be interpreted in more depth to characterize the subgenres that are classified. The next subgenre constellation for which the feature weights are analyzed is novela histórica versus novela de costumbres (see figure 56).

694The positive bars to the right are the topics that are distinctive for the historical novels and the negative bars to the left are the ones that are distinctive for the novels of customs. Interestingly, the top topic for novels of customs is the same as for sentimental novels: “baile-salón-fiesta”, which shows that the text types of these two kinds of conventional subgenres have common stylistic traits. The top topic for the historical novels is, in this case, the topic of battles: “soldado-enemigo-fuego”. The number of top topics for each subgenre is less balanced in this case than for the constellation novela histórica versus novela sentimental because, among the top 25, there are 10 topics for the historical novels and 15 for the novels of customs. This is a sign that the novels of customs are more diverse in terms of characteristic topics. Besides the first one, the other topics with high weights that characterize the novels of customs are concerned with lower class work (“muchacho-dinero-año”), with tobacco, sugar, and coffee plantations and work on them (“gallo-finca-negro”, “amo-estancia-muchacho”), which clearly points to the Cuban novels, with work on a ranch (“cutter-indio-tierra”), which is a topic that seems to be related to cattle and sheep breeding in Argentina, with the countryside (“pueblo-molino-sol”, “caballo-camino-árbol”, “muchacho-año-estancia”), money, business, and gambling (“dinero-fortuna-oro”), meals (“mesa-copa-vino”), the description of rooms (“mesa-puerta-sala”), the description of characters (“boca-cabeza-perro”), including pets (dogs and cats), youth and student lifestyle (“ventana-año-piano”), and reflections about society (“año-sociedad-idea”). Of these topics, the one about meals was also among the top topics for sentimental novels. Furthermore, the mixed topic “ojos-años-manos” about views, nature, body, and time appears again. Among the top topics for the novels of customs, there are several that cover rural life and surroundings, which confirms that the novelas de costumbres were oriented towards that sphere. In addition, the aspect that the life and working conditions of people are realistically described in the novels of customs becomes visible in the top topics. Finally, there are several descriptive topics as well as a reflective one, which is in line with the aim of the novel of customs to represent different areas of society closely and also to provide a social critique or vision. For inspection, the two descriptive topics “mesa-puerta-sala” and “boca-cabeza-perro” are given in figure 57.

695The other top topics for the historical novels are very similar to the ones that also appeared in contrasting the historical novels with the sentimental ones, and they are semantically much more homogeneous than the various top topics of the novels of customs. Again, elements of colonial history and also contemporary politics are present. A topic that is new here is, for example, “unitario-patria-cabeza”, which is related to the struggle between Unitarians and Federalists that took place in Argentina in the first half of the nineteenth century. The last constellation of thematic subgenres for which the coefficients of the linear classification model are analyzed is novela de costumbres versus novela sentimental (see figure 58).

696Among the top 25 topics, there are 14 that are typical for the novels of customs and 11 that are typical for the sentimental novels. Many topics that were already relevant in the other subgenre constellations appear again, for example, the lower class work topic “muchacho-dinero-año”, the youth and student life topic “ventana-año-piano”, or the countryside topic “muchacho-año-estancia” as topics that are distinctive for the novels of customs, and the love topics “amor-corazón-alma” and “idea-pasión-alma”, as well as the private conversation topic “voz-cabeza-mirada” as topics that are typical for the sentimental novels. In addition, some new topics appear, for instance, “carta-papel-duda” for the sentimental novel, which addresses the writing and reading of letters, or “tío-sobrino-primo” for the novel of customs, which is a topic about relatives and cliques. Furthermore, some topics change the side, that is, they become distinctive for another subgenre than in the other constellations that were already examined. This is the case for the airs and looks topic “tía-do-aire”, which is now typical for the novel of customs but was important for the recognition of sentimental novels when they were contrasted with historical ones, and also for the abstract views, nature, body, and time topic “ojos-años-manos”, which was distinctive for novels of customs in contrast with historical novels and is now typical for sentimental novels when compared to novels of customs. This fluctuation shows that even though some characteristic traits of the subgenres remain constant independently of the subgenre constellation, others are relative and depend more on the kind of subgenres that are compared. The fact that topics are, in one case, typical for the sentimental novel and, in the other, for the novels of customs is again a sign of how these two subgenres are intertwined thematically. That topics are in one case typical for the sentimental novel and in the other for the novels of customs is again a sign of how these two subgenres are intertwined thematically. The plots for the constellations of the individual thematic subgenres against all other novels are not shown in detail here. What is interesting about them is that in all three cases, there are more top topics for the positive class than for the other group, which shows that the classifier focuses on the aspects that are specific for the individual subgenre that is compared to the big group of novels with many different subgenres, which makes complete sense.⁵⁸⁷

697Besides the feature importances, also the cases of correct and false classifications of individual novels were analyzed for each subgenre constellation, with the aim to find out how many and which of the texts that are part of the conventional subgenres are typical or untypical for the text types. The results of this analysis are summarized in the form of histograms, which show the distributions of true positives, false positives, and false negatives. The true positives are interpreted as instances of the conventional genre, as well as of the text type and hence of the textual genre. The false positives are instances of the text type but not of the conventional and textual genre, and the false negatives are part of the conventional genre but not of the text type and, therefore, also not of the textual genre. For each novel in the corpus, it was counted how often it fell into one of the three groups in absolute and relative terms. As the data selection process was random and repeated several times (ten times for the undersampling and ten times for the cross-validation), not every novel of the corpus is present in every classification, and some novels were classified more often than others. The relative numbers of correct classifications and misclassifications can still be used to examine how prototypical the novels are for the subgenres. The chart for novela histórica versus other novels is given in figure 59.⁵⁸⁸

698Of special interest are the bars on the right side because they mean that the novels covered by them were classified correctly or wrong in many cases. In the histogram, the bars are grouped in steps of 10 %, starting with 0–9 % and ending with 90–99 %, and then 100 %. So the rightmost group of bars stands for the novels that were classified correctly or wrong in 100 % of the cases. If the novels were classified correctly, they can be interpreted as prototypical instances of the subgenre. If not, there is a discrepancy between the conventional and the textual genre for these novels. Of course, if a novel was only classified once and then correctly, it would appear as prototypical. That characterization would be less sure than for a novel that was classified 100 times and each time correctly, so these details have to be taken into account when this kind of chart is interpreted. For the constellation novela histórica versus other novels, 252 of the 256 novels in the corpus were classified at least once, among them all the 67 historical novels. Of these, 46 (69 %) were classified correctly in all 100 classification runs and can thus be considered the prototypical core of the text type novela histórica. Among them are, for example, the romantic historical novel “La cruz y la espada” (1866, MX) by Eligio Ancona and the realist historical novel “Puebla” (1903, MX) by Victoriano Salado Álvarez, which both carry the explicit historical subgenre label “novela histórica” and for which there is also agreement among literary historians to classify them as historical novels. As an example of such prototypical historical novels, the probabilities of the topics that are most distinctive for historical versus other novels are visualized for the novel “La cruz y la espada” in figure 60.⁵⁸⁹

699In the plot, the topics are ordered by the weight that they have for the distinction of historical and sentimental novels in general, so that “baile-salón-fiesta” is the topic with the most weight and “ejército-insurgente-sangre” with the least. The colors indicate for which subgenre the topics are typical. The plot for “La cruz y la espada” shows that several topics that are distinctive for historical novels also have higher probabilities in this novel. For example, this is the case for “sangre-voz-arma”, “príncipe-rey-palacio”, “conde-condesa-señorito”, “cura-iglesia-sacerdote”, and “convento-fraile-sor”. Especially the last four topics characterize this novel as one that is set in a more remote past (it is about the Spanish conquest of Mexico) because they refer to monarchy, nobility, and church. So the novel is a specific subtype of the historical novel, but still always classified correctly. On the one side, it has a high probability for the topic “sangre-voz-arma”, which is a general topic for historical novels, but there are also top topics of historical novels which are not so important in “La cruz y la espada”, as, for example, “soldado-enemigo-fuego”, “virrey-oidor-señorita”, “capitán-coronel-gobernador”, or “gobierno-pueblo-ministro”. The first one of these is a general topic about battles. The second would be typical for a novel that is set in the colonial era because it mentions the viceroy and “oidor”, which was a judge in the colonial judicial system. The last two would be typical for a novel that treats contemporary history. This shows that the top topics of historical novels, in general, combine elements of different subtypes of the subgenre. As the case of “La cruz y la espada” shows, it is not necessary for each individual novel to have high values for all of these topics in order to be always classified correctly as a historical novel. This means that not only the idea of prototypically organized categories is present in the statistical classes that are determined in the machine learning process, but also the idea of family resemblance: The novels can have high probabilities of specific subsets of the most distinctive topics, but they do not all need to have the same distributional profile. However, the aspect of family resemblance that is covered by the statistical classification is also limited because the boundaries between classes are still strictly drawn, and there are no loose networks of novels with overlapping similarities. What can be done is to interpret the novels inside of a class as members of a family that share different characteristics, or to analyze novels at the edges between two classes, that is, novels that are often misclassified for each class, to see to what extent they share properties.

700In the case of “La cruz y la espada”, not only the topics that are distinctive for historical novels are of interest, but also the ones that are typical for other types of novels because this novel also has high probabilities for some of these topics. For example, a top topic that has much weight in the novel is “voz-cabeza-mirada”, and also “año-sociedad-idea” and “ojos-años-manos” have higher weights. However, in sum, these topics are less dominant than the ones that are typical for the historical novel. What this shows is that elements that are typical for other subgenres can be present and have a certain weight, but as long as they are not dominating, the novel is still classified correctly. So the class of a novel is really determined in quantitative terms: which topics “win” in terms of numbers? This also means that a prototypical historical novel can have elements of other subgenres, as well. This aspect is particularly useful in the case of the nineteenth-century Spanish-American historical novels, of which many were romantic and included a sentimental plot or were more realistic and contained passages with descriptions of customs. In this view, instances of prototypes are not necessarily pure but have certain feature values and distributions that are quantitatively dominating and can be interpreted as salient.

701Returning to the group of prototypical historical novels that were always classified correctly, there are also novels that have been discussed not only as historical novels but also in other terms by literary historians. For example, the novel “Amalia” (1855, AR) by José Mármol has been described as a political novel, a historical novel, and also a sentimental novel, but is here always classified as a historical novel. Regarding the historical genre conventions, an interesting aspect of this novel is that the main title, “Amalia”, is typical for sentimental novels because it refers to a female first name, but the novel also has the subtitle “novela histórica” in most editions that are published from 1874 onwards. It was therefore labeled with the primary thematic label “novela histórica” here. The results of the SVM classifier make clear that the sentimental plot elements do not prevent this novel from being classified as a historical one. The four novels of the series “Dramas militares”, which were published between 1884 and 1886 by Eduardo Gutiérrez and in which the protagonist is an Argentine gaucho, are also always classified as historical novels.

702There are three novels that resulted as false positives in every case and two novels for which this happened in more than 90 % of their classifications so that they can be interpreted as belonging to the text type of the historical novel, although they did not have the primary thematic label “novela histórica”. One of them is “Las gentes que son así (Perfiles de hoy)” (1872, MX) by José Tomás de Cuéllar, which is a novel of customs that is part of the series “La linterna mágica”. It was classified as a historical novel 60 times out of 60. Another one, which was classified 39 times as a historical novel and one time as “other”, is “Los bandidos de Río Frío. Novela naturalista, humorística, de costumbres, de crímenes y de horrores” (1892, MX) by Manuel Payno, which has been assigned the primary thematic label “novela de costumbres” in the corpus. Already the subtitle shows that it refers to several genre conventions at once. It is about banditry and has also been classified as a novel of customs by Brushwood:

Manuel Payno, for example, published another serial novel Los bandidos de Río Frío, from 1889 to 1891. The subtitle even states that it is ‘naturalistic’, but it has much Romantic overstatement that identifies it with an earlier time. Taking the theme of banditry that had by that time become very popular, Payno wrote another essentially costumbrista novel. (Brushwood 1966, 116Brushwood, John S. 1966. Mexico in its Novel. A Nation’s Search for Identity. Austin: University of Texas Press.)⁵⁹⁰

703In these two example cases, it is not entirely clear why they are repeatedly classified as historical novels, so the top topic profiles of them are checked to see if they explain these results (see figures 61 and 62).

704As to the novel “Las gentes que son así”, a look into the probabilities that it has for the topics that are most distinctive for historical novels reveals that especially the topics “crimen-cárcel-preso” and “sangre-voz-arma” have high values. The first one is about crime, justice, and imprisonment, and the second one is a general topic about fights. These topics are not exclusively historical but can as well occur in novels of other subgenres. However, because they are typical for historical novels, the novel of customs “Las gentes que son así” in which they are so important, is mistaken as a historical novel, even if it does not have high probabilities for topics that are more specific for historical novels, such as “soldado-enemigo-fuego” or “sargente-jefe-soldados”. What this demonstrates is that the classifier has no idea about the conventional genres and does not make a difference between necessary, typical, or sufficient features. It just compares the topic values and distributions, and if it finds that they are similar to the ones typically found in a certain subgenre, the novel is classified as such. The example of “Las gentes que son así” is one where the novel has textual similarities to historical novels and is recognized as part of the corresponding text type, but it should not be treated as part of the textual genre, because by convention and also in terms of necessary features it is no historical novel.

705The second case, “Los bandidos de Río Frío”, has another quality. The probabilities for the top topics that are typical for historical novels show that some of them also have high values in this novel, above all “capitán-coronel-gobernador”, “sangre-voz-arma”, “crimen-cárcel-preso”, “conde-condesa-señorito”, and “gobierno-pueblo-ministro”. They are more than in the novel “Las gentes que son así”, and even though the general fight and crime topics appear again, there are also topics concerned with the military, politics, and nobility. The novel “Los bandidos de Río Frío” was published in 1892 but is set in the 1830s. The male protagonist has a military position, and also social and political problems of banditry, insecurity, and corruption are treated in the novel. That the occupation of the protagonist has an influence on the topics of the novel and that this, in turn, affects how the novel is classified in terms of subgenre again makes clear that a machine learning classifier cannot distinguish between different reasons for which topics are present in the texts. However, the combination of themes which are treated in “Los bandidos de Río Frío”, together with the fact that it treats a period that is several decades away from its publication time, does not make it far-fetched to compare it to other novels of contemporary history, so here, the result is not considered so misleading as in for “Las gentes que son así”.

706The three other novels that frequently resulted as false positives are “Misterios del corazón” (1875/1897, AR) by Rafael Barreda, “La hija de Tutul Xiu. Novela yucateca” (1884, MX) by Eulogio Palma y Palma, and “La Chapanay” (1884, AR) by Pedro Echagüe. The first one was classified as a historical novel 50 out of 50 times, the second one 30 out of 30 times, and the third one 28 times as a historical novel and 2 times as “other”. Barreda’s novel has the primary label “novela sentimental” and the secondary thematic label “novela histórica” because it is set in the Rosas’ era. Therefore, it is not surprising that it is mistaken as a historical novel, but what is surprising is that this happens every time. It seems that, in this case, the setting has a bigger influence on the text type than the intended primary theme of the novel. Or, from another perspective, aspects of the sentimental novel are not untypical for nineteenth-century Spanish-American historical novels and are part of the historical novel’s text type. The case of “La hija de Tutul Xiu. Novela yucateca” can also be easily explained because it is an indianist novel set in the pre-Spanish period, and it also has the secondary thematic label “novela histórica” in the corpus. So here, the problem was that only the more specific thematic label was considered in the classification and not the more general secondary one. As to “La Chapanay”, it has the primary thematic label “novela de costumbres”. This novel is an account of the life of Martina Chapanay, a historical Argentine personality of the nineteenth century which is only of regional importance. Lichtblau summarizes the novel’s contents as follows: “Regional types, customs, and the daily life of the San Juaneses are vividly portrayed in the novel as we see Martina first as a rebellious young girl, then as a member of a band of highwaymen, and finally as a contrite woman aiding the forces of law and order” (Lichtblau 1997, 299Lichtblau, Myron. 1997. The Argentine novel: an annotated bibliography. Lanham, Maryland: Scarecrow.). This description identifies it as a novel of customs, but “La Chapanay” also contains passages in which the narrator expresses his political opinion and criticizes the Rosas regime (Lichtblau 1997, 299Lichtblau, Myron. 1997. The Argentine novel: an annotated bibliography. Lanham, Maryland: Scarecrow.). So similarly to “Misterios del corazón”, where it was the historical background that approximated the novel stylistically to other conventional historical novels, here there are argumentative passages that influence the text style in terms of topics. In addition, this is another case of a hybrid between the novel of customs and the historical novel.

707As to the novels that were labeled as historical novels but often classified as other types (the false negatives), there is none that was misclassified in every run. There are three significant cases, though. The first one is “Los esposos” (1893, AR) by Lola Larrosa de Ansaldo, which was classified as “other” 77 times and as a historical novel 23 times. In the corpus, the novel has the primary thematic label “novela histórica” because it had this subtitle in the first edition. The second edition, in contrast, does not have that subtitle anymore. Moreover, Lichtblau describes the novel as follows: “A cloyingly romantic novel, Los esposos interweaves several plots that demonstrate the author’s self-righteous morality and her traditional beliefs about women’s place in the home” (Lichtblau 1997, 530Lichtblau, Myron. 1997. The Argentine novel: an annotated bibliography. Lanham, Maryland: Scarecrow.). Finally, it also has the main title “Los esposos”, which seems rather typical for a sentimental novel. In the corpus, it has therefore been assigned a secondary subgenre label: “novela sentimental”. As it seems, the explicit historical label that this novel carries is not entirely detached from the textual material and fits in 23 % of the cases, but the novel is textually closer to another subgenre, most probably the sentimental novel. A look into the weight that the top distinctive topics for novela histórica versus “other” have in this novel confirms that it does not correspond to the historical novels textually (see figure 63). The topic with the highest value is “voz-cabeza-mirada”, which is indeed one of the top topics for sentimental novels when compared to others, but the other two topics, “año-sociedad-idea” and “cuerpo-amor-ocasión”, are neither distinctive for historical nor for sentimental novels, and also not for the novels of customs,⁵⁹¹ but for other subgenres. As the novel was published in 1893, it is possible that it includes elements that are typical for the realist or naturalist novel. The topic “año-sociedad-idea” points in the first direction and “cuerpo-amor-ocasión” in the second one.

708The second false negative is “Vía Crucis” (CU) by Emilio Bacardí Moreau, of which the first part, “Páginas de ayer”, published in 1910, is included in the corpus. It was classified as “other” 73 times and as a historical novel 27 times. This novel is an account of the Cuban struggle for independence in the second half of the nineteenth century, which means that it treats contemporary historical events. It has nonetheless been characterized as a historical novel by several literary historians, for example, by Remos y Rubio, who compares Bacardí’s style to the one of Pérez Galdós, a Spanish writer who published the famous series of historical novels entitled “Episodios Nacionales”. He also relates Bacardí Moreau’s style to the ones of Alexandre Dumas and Walter Scott (Remos y Rubio 1935, 42–43Remos y Rubio, Juan J. 1935. Tendencias de la narración imaginativa en Cuba. La Habana: Casa Montalvo-Cárdenas. https://dloc.com/UF00078289/00001/images.).⁵⁹² However, Remos y Rubio also remarks that the novel “Vía Crucis” contains accomplished costumbrista passages: “Por el alto valor costumbrista que hay en ella, por lo admirable de las descripciones locales, como cafetales, campiñas, amenas, hogares criollos, etc., ha sido estimada Vía Crucis por algunos críticos, superior a Cecilia Valdés” (Remos y Rubio 1935, 44Remos y Rubio, Juan J. 1935. Tendencias de la narración imaginativa en Cuba. La Habana: Casa Montalvo-Cárdenas. https://dloc.com/UF00078289/00001/images.). The similarity with other novels of customs, especially the Cuban ones that also contain descriptions of coffee plantations, might be an aspect that leads to the frequent misclassifications in the case of “Vía Crucis”. Its top topic profile is visualized in figure 64. It becomes visible that this novel has high weights for topics that are typical for the sentimental novel (“voz-cabeza-mirada”, “baile-salón-fiesta”, and “ojos-años-manos”) and indeed a topic about slavery (“esclavo-negro-amo”), but none of the topics that are distinctive for novels of customs versus other novels are strong in this novel. Instead, the two topics “año-sociedad-idea” and “cuerpo-amor-ocasión” are important, which are the same ones as in “Los esposos”. As “Vía Crucis” was published in 1910, it is a novel that is clearly not a romantic historical novel and also Remos y Rubio remarks on the new kind of historical novel that develops in the early twentieth century and of which “Vía Crucis” is an example: “La novela histórica se oscurece con el advenimiento de las nuevas tendencias, hasta que a partir de 1910 la reaniman con nuevos bríos y moderna fisonomía, el selecto Rodríguez Embil y el ameno narrador santiaguero, EMILIO BACARDI MOREAU” (Remos y Rubio 1935, 42Remos y Rubio, Juan J. 1935. Tendencias de la narración imaginativa en Cuba. La Habana: Casa Montalvo-Cárdenas. https://dloc.com/UF00078289/00001/images.). So the misclassifications of “Los esposos” and “Vía Crucis” can also be interpreted as signs of a new type of historical novel related to Realism, Naturalism, and Modernism, which is less frequent in general and in the corpus than the romantic type of historical novel and is therefore not recognized well, because it is not learned as part of the model of the quantitatively dominant historical novel. These results underline how important the construction of the corpus is for the digital text analysis, and the attention to balance subparts of it. However, if a certain type of novel simply was not frequent enough to be balanced against another type (as the modernist historical novel, for instance, against the romantic historical novel), then qualitative analysis or an approach that is different from statistical classification is more meaningful to describe its characteristics.

709The third novel that carries the label “novela histórica” but is often misclassified is the novel “Las ranas pidiendo rey. Confesiones de una afrancesada (1861–1862)” (1903, MX) y Victoriano Salado Álvarez. It is part of the series of historical novels “Episodios Nacionales Mexicanos”. The novel was classified as historical 30 times and as “other” 70 times. Like “Vía Crucis”, this novel treats events of the recent past. Furthermore, it is written in the form of a diary and thus in the first person, which is quite unusual for a historical novel. Here it is assumed that especially the latter aspect turns it into a novel that is stylistically closer to other subgenres than the historical novel, but the narrative perspective does not need to have much influence on the topics. The plot for the novels’ top topic probabilities is given in figure 65 to check that. What it shows is that the most important topic which is not distinctive for historical novels is “año-sociedad-idea”, followed by “baile-salón-fiesta”. The second one is typical for sentimental novels, and it might have more weight here because of the personal account that is given of the events, but not necessarily. It can also be an element that is inherited from romantic historical novels with a sentimental plot. The other topic about intellectual work and society is the same that had much weight in “Los esposos” and “Vía Crucis”. Now, “Las ranas pidiendo rey” was published in 1903, and the historical novels of Victoriano Salado Álvarez are commonly attributed to the realist current. So this is yet another example of a later type of historical novel that challenges the classifier.

710The findings for the textual coherence of the historical novel can be summarized as follows: The majority of novels are prototypical historical novels and are both part of the conventional and the textual genre. There are only a few novels that are consistently misclassified or classified wrongly in more than 70 % of the cases. Five novels are regular false positives and interpreted as members of the text type but were, for different reasons, not labeled as members of the conventional genre. Three novels are persistently false negatives or members of the conventional genre but not of the text type, for individual reasons but also because they were published in the late nineteenth or early twentieth century. In all of these cases, the results can be explained by comparing and relating the textual and the conventional generic levels to each other, so their misclassifications are not considered real errors, except in the case of “Las gentes que son así”. Howver, also in that case, the decision of the classifier can be explained. Besides, there are groups of novels that are only sometimes misclassified. For instance, 13 of the “novelas históricas” are classified as other types only in up to 20 % of the cases. They are still considered members of the text type but less prototypical ones. On the other hand, there are 38 novels of other subgenres that are classified as historical novels in up to 20 % of the cases. These are members of other text types that have similarities with the historical text type and that are stylistically located at the edge of the historical prototype category. For the sentimental novels and the novels of customs, no individual cases are discussed, but their generic profiles are evaluated on a general level. The histograms for the constellations novela sentimental versus “other” and novela de costumbres versus “other” are given in figures 66 and 67.

711All of the 55 sentimental novels in the corpus were included in the classifications for the constellation novela sentimental versus “other”. In total, 243 of the 256 novels were part of this contrast at least once. Like the historical novel, also the sentimental novel has a strong prototypical core, but it is smaller than in the case of the historical novel. Of the sentimental novels, 33 (60 %) were classified correctly in 100 of 100 cases, compared to 67 % of the historical novels. For the sentimental novel, there are twelve instances that are regular false positives (seven with 100 %, four with 90 or more, and one with 80 %) and three instances that are very frequent false negatives (one of 100 % and the other two with 97 and 89 %). So the number of novels that were not labeled as sentimental novels but are recognized as such on the textual level is higher than for the historical novels, while the number of novels that carry the label but are not textually congruent with the subgenre is the same. Otherwise, the distribution is similar to the one of the historical novel. Less prototypical are 15 novels that have the label but are not recognized as sentimental novels in up to 10 % of the classifications. At the edge of the category are 34 novels that are classified as sentimental novels in up to 20 % of the cases but were not labeled as such.

712The novel of customs has 50 instances in the corpus, all of which were included in the classifications of novela de costumbres versus “other”. 241 novels out of the 256 in the whole corpus participated in these classifications. The histogram for the novel of customs looks a bit different than the ones for the historical and the sentimental novel. The group of novels that are always or almost always classified correctly is smaller, and there are more false positives which are relatively frequent. The amount of frequent false negatives, in contrast, is not especially high. This means that the prototypical core of this textual genre is smaller: 20 novels (40 %) are classified correctly in 100 of 100 cases. In addition, more novels that are not labeled as novels of customs still belong to the text type (according to these results): 19 novels are classified as novelas de costumbres in 70 % of the cases or more. On the other hand, there is only one novel which is labeled as novel of customs and always classified as “other”: the Cuban anti-slavery novel “El negro Francisco” (1873, CU) by Antonio Zambrana y Vázquez. An interpretation of these results is that the description of customs is a textual element in many of the nineteenth-century Spanish-American novels. However, it is not always marked on the level of historical genre convention and also not always considered as the primary thematic element by literary historians even if it dominates a text quantitatively. As was seen before, there are many different topics that are typical for the novels of customs, which might contribute to this lesser degree of overlap between the larger text type and the smaller conventional genre. The results for all three thematic subgenres demonstrate that the primary thematic labels that the novels have are not totally disconnected from the stylistic characteristics of the texts; quite the contrary. Nevertheless, there are also novels in which the conventional subgenre identity does not correspond to the textual one, either completely, as in the extreme cases, or in part. All constellations are possible on the way from a prototypical core to the edge of a textual genre, but some are more frequent than others, depending on the kind of subgenre that is analyzed.⁵⁹³

713In what follows, the classification results for the MFW-based features are presented. These features consist of three main groups: words, word n-grams, and character n-grams. The results for each group are reported separately. In total, 18,000 classification runs were considered for basic MFW, 54,000 for word n-grams, and 162,000 for character n-grams.⁵⁹⁴ The results for the MFW-based features are summarized in table 38.⁵⁹⁵

714With the KNN classifier, a mean accuracy of 0.68 is reached with MFW features and also with character n-gram features. The mean accuracy for word n-grams is only 0.64. With SVM, MFW yield the best mean accuracy value of 0.77, character n-grams 0.76, and the word n-grams 0.75, so here, the results for the three subtypes of MFW-based features are very close. The RF works best with MFW features, which lead to a mean accuracy of 0.78, followed by character n-grams with 0.76 and word n-grams with 0.74. So MFW are the best token unit across all classifiers, and character n-grams work equally well for KNN, but word n-grams work less well for all three classifiers. SVM and RF are almost equally successful and have the same average accuracies across the different types of feature sets (MFW, word n-grams, and character n-grams). Because RF is slightly better for the basic MFW, it is chosen here for evaluating the classification results for primary thematic subgenres with MFW-based features further. Figure 68 displays the classification results for different numbers of MFW and tf-idf versus z-scores.⁵⁹⁶

715It is directly visible that the lower numbers of MFW are not suited for classification by thematic subgenre. On the other hand, the results stay at a high level, approximately from the 1,000 MFW onwards. As was shown in the chapter on the MFW-based features, in the range of 1,000 MFW, there are already verbs, adjectives, adverbs, and nouns, so it can be concluded that these content words are essential for the classification by thematic subgenre.⁵⁹⁷ As to the difference between tf-idf and z-scores, the former are better up to 3,000 MFW and the latter from 4,000 MFW onwards. When one rounds to two decimal places, the top mean accuracies of 0.81 are reached with tf-idf and 3,000 or 2,000 MFW.⁵⁹⁸ Because of the above results, tf-idf and 3,000 MFW are chosen to analyze the classification results further. Next, word n-gram features are examined. The classification results relying on these feature sets are summarized in figure 69.

716In general, word 2-grams give higher mean accuracies than 3-grams or 4-grams, which was expected because the more words are involved, the less frequent the combinations are in the corpus and potentially less helpful to classify texts by genre if they only occur in a few texts. Here, too, the results are generally better from 1,000 MFW onwards than with fewer MFW, but they also drop again for the 2-grams and 4-grams, even if slowly. The differences between tf-idf values and z-scores are very small. The best mean accuracy is achieved with tf-idf, word 2-grams, and 2,000 MFW, which is at 0.79. If the values are rounded to two decimal points, the same result is achieved with 2-grams from 1,000 to 4,000 MFW with both tf-idf and z-scores, and also with 5,000 MFW and z-scores, so the differences between these different constellations are minimal.

717The third set of MFW-based feature sets that is examined are the character n-grams. Here, different parameter constellations for three n-gram subtypes are evaluated. The first subtype are the  “classic” character n-grams containing all characters, punctuation marks, and blank spaces. The classification results for these are summarized in figure 70.⁵⁹⁹

718With character n-grams, the best mean accuracy is achieved with the combination of tf-idf, 3-gram characters, and 4,000 MFW, reaching 0.80. With values rounded to two decimal points, the same result is achieved with tf-idf, 4-gram characters, and 5,000 MFW.⁶⁰⁰ Besides the classic n-gram features, also two special types of n-grams were created. The first of them only uses mid-word and multi-word n-grams and is called “word n-grams” and the second type only uses prefix n-grams as well as n-grams ending with punctuation marks and is called “affix-punct”.⁶⁰¹ In the following, it is checked if they have advantages over the classic n-grams. Figure 71 visualizes the results for the “word” character n-grams.

719With the “word” character n-grams, the best result is an average accuracy of 0.80, which is reached with tf-idf, 3-grams, and 1,000 MFW and also with tf-idf, 4,000 MFW, and 4-grams or 5-grams, as well as with tf-idf, 5-grams, and 3,000 MFW (rounded values). The top mean accuracy is the same as for the classic n-gram features, so it was no advantage to only use n-grams that are derived from words or word boundaries. The classification results for the second special type of n-gram features, the “affix-punct” character n-grams, are given in figure 72.

720In this group of feature sets, the highest mean accuracies of 0.80 are achieved with character 5-grams and tf-idf values for several numbers of MFW (1,000 to 5,000, with accuracies rounded to two decimal places). The 4-grams with tf-idf reach this value several times, too. Again, there is no difference in the mean accuracy compared to the other types of n-grams, so for the classification of the nineteenth-century Spanish America novels by thematic subgenres, it is not decisive which kind of n-gram features are used.

721When the findings for different token units are summarized, it can be observed that the top mean accuracies are all very similar, as they range from 0.79 in the case of word 2-grams over 0.80 for the character n-grams to 0.81 for normal MFW. All of these token units seem equally suited to classify the novels by thematic subgenres. As to the numbers of MFW, higher numbers of 1,000 or more tokens work better in all cases. Regarding the normalization techniques, tf-idf was, in general, the one that reached the top values, but the differences between the different maximum mean accuracies were small in many cases. Only the standard MFW features are chosen here to inspect the classification results for the different subgenre constellations, with 3,000 MFW and tf-idf. The results for this combination, using the RF classifier, are given in table 39 below.⁶⁰²

722What is striking about the results for MFW is that the ranking of subgenre constellations is the same as for the topic features. The constellation that is easiest to distinguish is novela histórica versus novela sentimental, which reaches a mean accuracy of 0.90, the second best novela histórica versus novela de costumbres with 0.88, and the third best novela histórica versus “other” with 0.86. Again, the historical novel is the subgenre that can be separated best from the other thematic subgenres, followed by the sentimental novel. For the latter, a score of 0.78 is achieved if the subgenre is compared to all other novels and a score of 0.74 if it is contrasted with the novel of customs. As with the topics, also with MFW, the novelas de costumbres have the lowest mean accuracy of all constellations when they are classified against all other novels. In part, it is understandable that 3,000 MFW leads to similar results as topics because content words and, thereby, semantic features which can capture thematic elements or other semantic surface structures are also part of this feature set. Nevertheless, in the MFW set, not only nouns are included, but all kinds of word categories, and there is no layer of hidden semantic distributions that is analyzed, but the (normalized) word frequencies are directly used as features. So obtaining such similar classification results with different feature sets means that the results are meaningful regarding the characteristics of the different subgenres.

723Another noticeable aspect of all the results for the classification of thematic subgenres is that the standard deviations were not reduced considerably from the first average results based on all feature sets and parameter constellations to the individual subgenre constellations. They were about 0.15 in the first overviews and are still at about 0.10 in the direct comparison of subgenres with a certain feature set. This means that approximately two-thirds of the variation are to due to different selections of novels for the various classification runs, which shows how important it is to have a corpus and subcorpora for the different subgenres as large as possible and how important it is to prevent the classifiers from focusing too much on specific novels. Furthermore, it means that in the case of the nineteenth-century Spanish America novels analyzed here, the thematic subgenres are not categories of very homogeneous texts but of texts that share textual features so that they are recognizable as instances of the text types, but they do so to different degrees. Furthermore, that the results for different classifiers, feature sets, and feature parameters are so similar indicates that the approximately 20 % of accuracy that is missing to reach the perfect classification is probably not due to a wrong choice of features or settings in the classification procedure but due to the discrepancies between the conventional genres and their text types. This could also be shown in the discussion of examples of historical novels that were frequently misclassified.

4.2.2.1.2 Literary Currents

724The classification of the novels by their primary literary currents has several purposes. As for the thematic subgenres, the results of the classification are analyzed to see which classifiers and feature sets work best to capture the differences between the various literary currents. The best constellations are then chosen to see how well the different constellations of literary currents can be classified. A short analysis of the features that are decisive in the classification as well as overviews of how often individual novels were classified correctly or wrong is given for the MFW-based features in this case but not for the topics. Word-based features are considered more interesting for the literary currents because they are not primarily thematically defined, even if ranges of different topics are also characteristic for them. For the literary currents, style in a narrower sense is assumed to play a more important role.⁶⁰³ Figure 73 shows the distribution of primary literary currents in the corpus.⁶⁰⁴

725Of the 256 novels, there are 55 for which the primary literary current is unknown.⁶⁰⁵ Romantic novels are most frequent with 116 instances, followed by 45 naturalistic, 35 realist, and 5 modernist novels. The following subgenre constellations are selected for the classification of primary literary currents:

• novela romántica versus other novels
• novela realista versus other novels
• novela naturalista versus other novels
• novela romántica versus novela realista
• novela romántica versus novela naturalista
• novela realista versus novela naturalista

726The novels for which the literary current is unknown were not included in classifications because they could be instances of the positive class. As for the thematic subgenres, also here the results for topic features are presented first, and then the ones for MFW-based features. First, it is determined which of the three classifiers, KNN, SVM, and RF, worked best. The overall results for the classification by literary current with topic features are shown in table 40. The results are based on 144,000 classification runs for each classifier, including all parameter variations, repetitions of data selection, and cross-validation steps.⁶⁰⁶

727SVM works best with a mean accuracy of 0.85, but KNN and also RF follow closely with 0.83 each. These mean accuracies are higher than the ones for the thematic subgenres, which were 0.77 and 0.80 for topic features. This was not expected for the topic features because especially the novels of the romantic current are subdivided into specific thematic subgenres. Apparently, despite the different conventional and textual thematic subgenres, the romantic novels all have topics in common when contrasted with novels of the two later realist and naturalistic currents. The possibility that the topics are not thematic in a narrow sense also has to be kept in mind because they can also represent other structures (descriptions, argumentation, motifs, etc.). That so high results are already reached on average for all kinds of features and subgenre constellations demonstrates that the literary currents were not merely conventional movements but that the novels associated with them had textual and stylistic traits in common. The next step consists in evaluating which feature parameters worked best with the SVM classifier. An overview of the results for different numbers of topics and optimization intervals is given in figure 74.

728The range of the mean accuracies goes from 0.83 to 0.86, so the differences are small between the different parameter constellations. The top mean accuracies are reached with 70 topics and an optimization interval of 1,000 and with 90 topics and an interval of 2,500.⁶⁰⁷ Regarding the developments of different optimization intervals with an increasing number of topics, the curves for intervals of 50, 100, and 250 first fluctuate and then rise from 80 topics onwards. Intervals of 500 and 5,000 have downward trends, the curves of 1,000 and 2,500 mainly fluctuate but also fall with more topics, and the curve for no optimization at all is the most stable one. Here, the best combination of 90 topics and an optimization interval of 2,500 is chosen to analyze the classification results for literary currents further. That a constellation with less hyperparameter optimization works better for the literary currents than for the thematic subgenres is a sign that it is helpful if the topics are more evenly distributed across the corpus and in the individual novels, maybe because the literary currents are more general phenomena than the thematic subgenres.

729With the decision for a certain combination of parameters for the topic feature set, the classification results for the different subgenre constellations can be inspected. The results for the literary currents with SVM, 70 topics, and an optimization interval of 1,000 are listed in table 41. These results are based on 500 classification runs for each constellation because of five topic modeling repetitions, ten random data selections, and 10-fold cross-validation.

730The mean accuracies range from 0.77 to 0.92, so there are clear differences as to how well the different literary currents can be distinguished from each other. The best mean accuracy of 0.92 is achieved for the classification of the novela romántica versus other novels and the novela romántica versus the novela naturalista. That the latter constellation yields the best results could be expected because the naturalistic novel can be considered a further development and specialization of the realist novel, carrying the realist aesthetic to the extreme so that it is also poetically more distant from the romantic novel. The best result for the novela romántica versus other novels with 0.92 and also the third best one for the novela romántica versus the novela realista with a mean accuracy of 0.91 was not expected for several reasons: because of the smooth transition of Romanticism into Realism in Spanish America, the existence of several novels that combine elements of both currents and the novels that included realistic elements before Realism (in particular the novels of customs). Furthermore, the romantic novel included several thematic subgenres, each of which is recognizable by its own topics, as was seen in the previous chapter. As it seems, the difference between romantic and realist novels is still big enough to lead to such good classification results. The classification of naturalistic novels against all others achieves a mean accuracy of 0.85, the realist novel versus the others 0.80, and the realist versus the naturalistic novels 0.77. That the score is lowest for the novela realista versus the novela naturalista could be expected because of the temporal overlap of both literary currents in nineteenth-century Spanish America but also because of the similar aesthetic concept of these two currents. The realist novel can be interpreted as the current that has a middle position between the romantic and the naturalistic novel. Therefore, it is not surprising that it has the lowest mean accuracy when it is contrasted with all the other romantic and naturalistic works. Next, the results achieved with MFW-based features and the different classifiers are analyzed for the literary currents. An overview of the results for the three main types of MFW features (words, word n-grams, and character n-grams) is given in table 42.

731As with topics, so with MFW-based features, the SVM classifier achieved the best results with a mean accuracy of 0.86 for the basic MFW. The second-best mean accuracy is 0.84 for RF, and the worst is 0.78 for KNN. For all three classifiers, word n-gram features produced worse results than single words. For KNN, the best results were achieved with the character n-grams, and for SVM and RF, with words. As the SVM classifier worked best, it was chosen to further evaluate the results. The mean accuracies for different values of MFW and the three normalization techniques are shown in figure 75.

732For the basic MFW features, the highest mean accuracy of 0.88 is reached with 3,000 MFW and tf-idf values. The differences between tf, tf-idf, and z-scores are minimal. The mean accuracies for the other numbers of MFW from 1,000 up to 5,000 are also at 0.87 or 0.88 if the values are rounded to two decimal points. So the most important aspect is to use 1,000 MFW as a minimum, which was also the case for the thematic subgenres. As 3,000 MFW and tf-idf resulted in the highest mean accuracy, that combination of parameters was kept for the analysis of further results. Figure 76 visualizes the classification results for the word n-gram features.

733The results for word n-grams are similar to the ones for thematic subgenres in that word 2-grams produce better results than word 3-grams or 4-grams. Here, too, the best mean accuracy of 0.87 is reached with 3,000 MFW for tf-idf and also for z-scores. As for word features, also for word n-grams, the results get much better with 1,000 MFW or more. It is of interest that the word 2-grams are almost as good as the word features which reached 0.88 because interpreting word 2-gram features might reveal further relevant aspects about the subgenres. Next, the classification results for normal character n-gram features are given in figure 77.

734For the character n-grams, the highest mean accuracy of 0.86 is achieved with character 4-grams, 4,000 MFW, and tf-idf, so it is slightly worse than for word 2-grams or basic MFW. Other n-gram units lead to very similar results, but interestingly, with more than 3,000 MFW, the scores drop a bit for character 3-grams but rise for the 4-grams and also the 5-grams. As more content words are included with higher numbers of MFW, maybe it is advantageous for the classification of subgenres when a bigger part of these words is captured with the larger character n-grams. In contrast to word units or word 2-grams, the results for character n-grams already stabilize largely with 500 MFW. The results for the special n-gram type “word”, which only contains character n-grams from the inside of words and n-grams that include the end of one word and the beginning of another, are shown in figure 78.

735Here the highest mean accuracy of 0.88 is reached with 500 MFW, character 4-grams, and tf-idf values or z-scores. This is higher than for the normal character n-grams, but this special character n-gram type does not outperform other feature types completely because it is equal to the highest score that was achieved with basic MFW. The other special n-gram type, “affix-punct”, which only includes prefixes, (i.e., word beginnings), and punctuation marks, yields 0.87 as the best score. The same is achieved with character 4-grams, this time with 2,000 MFW and tf-idf values, as shown in figure 79. That 4-grams are the best token unit for character n-grams was also found out by Hettinger et al. (2015 Hettinger, Lena, Martin Becker, Isabella Reger, Fotis Jannidis, and Andreas Hotho. 2015. “Genre Classification on German Novels.” In Proceedings of the 26th International Workshop on Database and Expert Systems Applications (DEXA), 249–253. Valencia. https://doi.org/10.1109/DEXA.2015.62.), so they seem to be generally useful for subgenre classification.

736In table 43, the classification results for the different constellations of literary currents are given, based on the SVM classifier, 3,000 MFW, and tf-idf values.⁶⁰⁸

737With MFW features, the subgenre constellation for which the highest mean accuracy was reached in the classification is the novela romántica versus the other novels with 0.93. The same constellation also had the highest value with topic features, but here it is even slightly higher. The second best result is 0.92 for the novela romántica versus the novela realista, followed by 0.90 for the novela romántica versus the novela naturalista. The contrast of the realist novel with all others yields a mean accuracy of 0.87, the realist novel against the naturalistic novel 0.86, and the worst result is 0.83 for the naturalistic novel versus the other novels. Compared to the topic features, here, all values are higher except the ones involving the naturalistic novel. This result is intriguing because it seems that the naturalistic novel is better captured with features related to themes than with word features. The change goes in the other direction for the realist novel. Here, the contrast between the novela realista versus the other novels is 7 % higher than with topic features, and the opposition of the novela realista and the novela naturalista is 9 % higher here than with topic features. The novela romántica versus the novela realista is only 1 % better with MFW than with topic features, though. So the realist novel wins if word features are used instead of topics. The result is so interesting because the label “novela naturalista” is used in subtitles of several novels in the corpus. In addition, its status as a (thematic) subgenre versus a literary current or movement has been discussed, for example, by Schlickers, who confirms that the naturalistic novel was a concept that was consciously applied and communicated by contemporary writers in the nineteenth century.⁶⁰⁹

738To get an insight into the kind of MFW features that are relevant for the classification by literary current, the two contrasts of the novela romántica versus the novela realista and the novela realista versus the novela naturalista are analyzed here. Figure 80 shows the top feature weights for the romantic versus the realist novels.

739The features that are distinctive for the realist novels are on the left side and the ones for the romantic novels on the right side. Among the top 25 features, there are 17 which are important for the realist novel and only 8 for the romantic novel, which is a sign that the realist novel is stylistically more homogeneous than the romantic novel. Among the top words for realist novels, there are several adjectives, of which the most important one is “rojo”. The others are “duro”, “gran”, “duros”, “excelente”, and “listo”. In the list of the romantic novels, there is no adjective. A more frequent use of adjectives could be due to more descriptive passages in the realist novels. Nouns, too, are only present in the list for realist novels: “corriente”, “grupo”, “política”, “envidia”, “hoja”, “importancia”. So politics is an important topic in realist novels. Groups of people are also relevant, and “envy” points to their materialist orientation. The other three nouns are not so easily interpreted at first sight. Furthermore, there are some verbs in the past tense: “mostraba”, “habló”, and “escribió”. The last one, together with “hoja”, seems to indicate that people frequently write things down in realist novels. The words that are distinctive for the romantic novel include the first person plural verb form “hemos”, the participle “quedado”, the adverbs “verdaderamente” and “entonces”, the demonstrative pronouns “esta” and “este”, and the adjective “solo”. The fact that the word “lonely” is typical for the romantic novel fits well with the concept of the individual romantic hero or heroine. Furthermore, it is noticeable that the words with the top weights for the two different literary currents belong to different grammatical categories. The feature importances of the naturalistic versus realist novel are shown in figure 81.

740Here, the words that are distinctive for the naturalistic novels are on the left side, and the ones that are typical for the realist novels are on the right. Again, the proportion of top features is 8:17, with more features for the realist novels. The nouns that are typical for the naturalistic novels are “mitad”  and “pasajeros”. The first one is an indication of quantity. The passengers could either be connected to transport in a city or to travel. The word “costado” is here probably the participle “cost” and not the noun “side” and alludes to the importance of money in naturalistic novels. The adverbs “hasta” and “recién” indicate that time also matters in novels of this literary current. The verb “convencer” is the only infinitive in the whole list. To convince someone means the need to discuss and the will to achieve a certain goal. The words “cual” and “fuesen” are not easily interpreted. As to the words that are distinctive for the realist novels, there are some which did already appear in contrast with the romantic novels. Again, there are several verbs in the past tense: “traía”, “dieron”, and “oyeron”. The nouns “figura”, “gracia”, and “voces” are new. The verbs in past tense could mean that narrative passages in which the actions of characters are described are more frequent in realist novels than in romantic and naturalistic ones. Altogether, the MFW features that are weighted highly by the SVM classifier can be interpreted well as stylistic cues of the literary currents in the same way as the topic features could be explained for the thematic subgenres. To conclude the discussion of the classification results for literary currents, the classification profiles with the proportions of true positives, false positives, and false negatives are visualized in figures 82 to 84. On the one hand, they serve to see how many novels were always classified correctly and are part of the conventional as well as the textual genres. On the other hand, it is checked whether there are any cases of novels that are misclassified very frequently and are only part of either the text type or the conventional genre.

741What is striking is that there are no cases of novels that are false positives or false negatives in more than 20 % of the cases for none of the literary currents. This was very different for the thematic subgenres. It means that there are more considerable discrepancies between the conventional genres and the text types for the thematic subgenres than for the literary currents. A hypothesis is that the literary currents that the novels belong to are mostly determined by literary historians. Therefore, they could be closer to the textual characteristics of the texts than, for example, explicit historical labels of thematic subgenres that might be misleading. Another hypothesis is that the style of novels that are part of a certain literary current is easier to recognize than the primary theme of a novel. At least in some cases of the novels, there are diverging opinions also among literary historians about which primary topic they have. Furthermore, the primary themes of the novels are not necessarily congruent with the surface topics, which can lead to misclassifications. All in all, the classification results for the literary currents showed that these are indeed categories with a stylistic unity.

4.2.2.2.1 Method

745To create the network for the family resemblance analysis, first, the similarities between all the individual novels were calculated for the chosen feature set using cosine similarity.⁶¹⁰ After that, the resulting textual similarities were mapped onto a network structure. The novels themselves constitute the nodes in the network. The network relationships (or edges) were determined using the three nearest neighbors of each text, which were selected from a ranking of the text similarities. The strength (or weight) of the edges was calculated by summing up the similarity values of the neighbors.⁶¹¹

746In the overall similarity matrix, there are relationships between all the texts. Reducing the number of connections for the network to the three nearest neighbors makes the network less complex and the closest relationships in it more salient. This, in turn, enhances the interpretability of the network. The choice of three is arbitrary and could be varied. However, using more than one nearest neighbor makes the results of the network more stable, as Eder has shown (Eder 2017, 56–60Eder, Maciej. 2017. “Visualization in stylometry: Cluster analysis using networks.” Digital Scholarship in the Humanities 32 (1): 50–64. https://doi.org/10.1093/llc/fqv061.). Eder introduced the idea of visualizing nearest neighborships based on textual similarities in a network structure, intending to make the results of stylometric cluster analysis more reliable. This technique is adapted here with a different aim: to formalize the family resemblance concept for genre analysis.

747In addition to the creation of the basic network structure, community detection was used to explore the families of novels in the network. Communities are sets of nodes in a network that are more densely connected to each other than to nodes outside.⁶¹² Different algorithms for the detection of network communities exist.⁶¹³ Here, the Louvain modularity algorithm was used. It is based on modularity optimization, which means that possible divisions of the network are checked and optimized to reach high modularity so that the nodes which share a community are more likely to be connected with each other than with other nodes that are not community members. The Louvain algorithm optimizes local communities iteratively until the best global modularity of the network is reached. Here, non-overlapping communities are built, meaning that each node is only part of one community in the result. As it looks for densely connected data points, the Louvain algorithm is suitable to detect families of novels. To that end, the novels are represented as vectors in a space of textual features, in which they can be closer to each other or further away, and these similarities (or distances) are interpreted in terms of network links. Furthermore, the algorithm is also comparatively efficient and has been implemented in Python, which is used to create and visualize the network here.⁶¹⁴

748Reflecting on how the concept of family resemblance is formulated by Wittgenstein and in literary genre theory and how it is implemented here, the following observations can be made. Using similarity relationships between novels based on feature distributions means that not the presence or absence of a trait or a set of traits determines the connection between members of a family and the difference to other families, but the numerical strengths of the features in combination and this only between pairs of individual novels.⁶¹⁵ The similarities of the various pairs are summed up and interpreted in terms of neighbor rankings to build the families. This transfers the idea of partial and overlapping similarities to a quantitative approach because nearest neighborship between novel A and another novel B, on the one hand, and novel B and C, on the other, does not mean that A and C must also be closest in a direct comparison.

749Second, when communities are calculated and interpreted as families, the boundaries of the categories are retroactively sharpened because communities are clusters, which have members and non-members. This is an advantage that balances out the looseness of the original family resemblance concept. However, there is a significant difference between these  communities and classes in a logical sense because the former emerge from a network of similarities and not from the condition of shared common features. The communities mark a boundary between one group of dense relationships and another, they cut off the family at a certain point but they do not lever out the basic idea of family resemblance, because the links to other families and family members in the network can still be explored.⁶¹⁶

4.2.2.2.2 Data

750The subcorpus of novels used for the family resemblance analysis includes 83 novels first published between 1840 and 1910. Of these, 40 are historical novels, and 43 are sentimental novels. 32 of the novels are Argentinian, 35 are Mexican, and 16 are Cuban. The novels were written by 74 different authors, 11 of them female and 72 male. To prevent the authorial signal from interfering too much with the genre signal, for each subgenre, only one novel per author was chosen. Nevertheless, if authors wrote novels in both subgenres, these are included. There are nine authors with both a historical and a sentimental novel in the subcorpus. Figure 85  shows the distribution of the novels by decade and subgenre.⁶¹⁷

751As features, a topic model with 100 topics and an optimization interval of 100 was used, which is considered a medium degree of specification, given that the overall corpus contains 256 novels. The feature set was generated for the whole corpus Conha19 with the goal of having more stable topics that represent the novel of the time in a better way than if they had been based on the smaller subcorpus. For the network analysis, only the features for the novels in the subcorpus are used.⁶¹⁸

752In figure 86, the top 40 words of four of the resulting 100 topics are visualized. They exemplify the range of themes covered in the novels. The first topic is about love and feelings (“amor”, “corazón”, “alma”, “amor”, “pasión”); the second topic is dominated by politics (“gobierno”, “ministro”, “guerra”, “poder”), the third one is about crime and banditry (“bandido”, “jefe”, “ladrón”, “robo”), and the fourth one about religion and colonization (“sacerdote”, “dios”, “español”, “guerrero”). The first number in parentheses indicates the rank of the topic by its probability in the whole corpus, so the topics are of different importance for the whole collection of texts. The lower the topic rank, the more important the topic is, so the love topic is a very general one, the politics and crime topics are still rather common, and the colonization topic is more special.

4.2.2.2.3 Results

753With the approach outlined in the previous section, three kinds of networks were produced, two for the individual subgenres and one for the two subgenres combined, as shown in table 44.⁶¹⁹

754The last column indicates how many families, that is, clusters based on the communities in the network, were produced. The number of clusters is identical for both historical and sentimental novels when they are analyzed separately. Given that the number of novels doubles when the two subgenres are combined, the number of resulting clusters does not grow proportionally, indicating that there is an overlap between the subgenres. The discussion of the results focuses on the historical novels.⁶²⁰ Figure 87 shows the first network for historical novels and topics (HIST-topics). The communities detected are indicated by the different colors of the nodes.

755An important question for the interpretation of the network is which kinds of novels constitute the different families. Before looking at different clusters in detail, an overview of the cluster sizes was generated, and the possible influence of some text-external and -internal factors on the clusters was calculated, as displayed in figures 88 and 89.

756Four of the resulting six clusters are evenly sized, with 8 novels each, and the other two are smaller. Cuban novels are only contained in clusters 1, 2, 3, and 5. Clusters 1, 4, and 5 are dominated by Mexican novels and clusters 2 and 3 by Argentine novels. Cluster 3 is an Argentine-Cuban cluster, and cluster 5 is a Mexican-Cuban cluster. Even if there are some tendencies regarding the distribution of novels by country in the different clusters, there is no cluster consisting only of novels from one country, and it should also be kept in mind that the overall number of novels in the individual clusters is quite small. The narrative perspective is not significant for the historical novels because there is only one novel with a homodiegetic narrator, the others all have a heterodiegetic narrator. The five historical novels written by female authors are distributed over the three clusters 2, 3, and 4, so there is no clear female cluster. Regarding the distribution of the novels over the years, there is also much overlap, as all the clusters have earlier and later novels. Apart from one outlier, cluster 2 is rather late, and cluster 4 is mostly filled with earlier works.

757When one looks at one cluster in detail, it is possible to retrace the family resemblance relationships. In table 45, the novels contained in cluster 3 are listed together with their nearest neighbors (N-1, N-2, N-3), including the weight of the edge to the respective neighbor. The strongest relationship exists between “La novela de la sangre” (1903, AR) by Carlos Octavio Bunge and “Los misterios del Plata” (1868, AR) by Juana Manso de Noronha because they are mutually closest to each other. Other bilateral nearest neighborships between novels in the cluster are highlighted in lighter orange. The novel “Pepa Larrica” (1884, AR) by Rafael Barreda has two nearest neighbors in the cluster, but the relationships are only unilateral. Boxes that are not highlighted show which nearest neighbors are outside of the current cluster. It becomes clear that some novels are central members of the family while others are rather distant relatives.

758The topic distributions for the five novels are visualized in figure 90 to see what topics are decisive for the relationships in this cluster of historical novels. The axis on the top shows the absolute value that the topic achieved in each novel, and the axis to the left shows the individual 100 topics.⁶²¹ In addition to the lines for the five novels in the cluster, a black dashed line indicating the mean topic values for all the historical novels in the network is added. The topics are ordered by importance in the whole corpus of 256 novels from top to bottom so that more general topics are at the top and more special topics are further down. Some topics of interest are labeled, the black ones being particularly important for this cluster and the red ones less important when compared to all the historical novels in the corpus. What makes the family approach visible is that not all the decisive topics are equally relevant for the individual novels in the cluster. For example, the topics “sacerdote-dios-español” and “fortaleza-batería-plaza” are underrepresented in the whole cluster, but “amor-corazón-alma” and “soldado-fuego-columna” are only partly less relevant. The first corresponds to the mean for the novel “Amalia” and the second reaches almost the mean for “Vía Crucis”. Topics that are overrepresented in several novels in the cluster are “voz-palabra-brazo”, “idea-espíritu-instante”, “pueblo-ley-país”, “calle-puerta-voz”, “agua-cuerpo-sangre”, “gobierno-ministro-guerra”, “puerta-espíritu-cabeza”, and “cabeza-rosa-asesino”. They stand for the general characteristics of the family: historical novels that are not so much mixed with love stories, not focused on military actions and not about the Conquest or colonial history, but about political ideas and conditions, and about (inter)personal contacts and states, about voices, words, and bodies. However, as specific topic values are not necessary conditions, some of the novels have their own special topics. The topic “mar-buque-puerto” is specific for “Los misterios del Plata”, “negro-esclavo-amo”  for “Vía Crucis”, the only Cuban novel in this cluster, and “capitán-voz-revolución” for “Pepa Larrica”.

759A more general overview of the topics that are distinctive for the different clusters in the network of historical novels is given in figure 91. In the heatmap, the brighter the boxes, the more important the topics are for the cluster, and the darker, the lesser important they are. The distinctiveness was calculated by normalizing the topic values to z-scores. Here, only the top 30 most distinctive topics are shown. The values in parentheses at the end of the topic labels indicate the ranks of the topics in the whole corpus, so the topic “voz-palabra-brazo”, for example, is much more general than “fortaleza-batería-plaza”.

760The distinctive topics of cluster 3 that were already discussed can be recognized in the heatmap. The smallest cluster 0 seems to be about the Conquest and colonial history, as the most distinctive topics are concerned with Indians and Spaniards (“indio-español-tierra”), rural church (“cura-fraile-pueblo”), and seafaring (“mar-buque-puerto”). In Cluster 1, topics about military campaigns and rural life prevail, making one think about internal struggle, bandits, and gauchos. The topics that are distinctive for this cluster are about conversation and social roles of bandits (“palabra-asunto-razón”, “bandido-jefe-razón”, “bandido-jefe-ladrón”), provincial police forces (“cabellero-comisario-provincia”), horses (“caballo-amo-instante”), soldiers and military actions (“manera-soldado-muerte”, “ejército-prisionero-jefe”), and ranches (“hacienda-compadre-pueblo”). Cluster 2 is not so easy to interpret. It is also about military action (“soldado-fuego-columna”, “sargento-cerro-gruta”, and “ejército-guerra-ciudad”), but there are other, individual topics. Cluster 4 is clearly romantic, as it contains topics about love (“corazón-alma-lágrima”, “amor-corazón-alma”), about the missionary work, colonial administration and aristocracy (“sacerdote-dios-español”, “alcalde-dama-barón”), and illness (“instante-doctor-sitio”). This fits well with the observation that it contains mostly earlier novels. The last cluster is politico-historical with top topics about the military (“soldado-jefe-coronel”, “fortaleza-batería-plaza”), government (“gobierno-ministro-guerra”), and the time of the French intervention in Mexico (“francés-emperador-estudiante”).

761The results for the network of sentimental novels and the one containing both types of subgenres are only summarized briefly here. Regarding the metadata, the cluster sizes vary more for the sentimental novels. The biggest cluster has 14 novels, and the smallest one has only two. All the clusters are mixed by country. Among the sentimental novels, there are more with an autodiegetic and a homodiegetic narrator, and the narrative perspective has an influence on the results. The smallest cluster, for instance, consists solely of autodiegetic texts featuring topics related to inner life and landscape. For the sentimental novels, there are also clearer tendencies of topic changes over time, as figure 92 shows. The early cluster 0 is romantic with letters, dance, aristocracy, and much emotionality. The three later clusters are the ones dominated by interiorization, and the mid-century cluster 5 is worldly about food, marriage, business, and money.

762When both subgenres are analyzed together, they are not neatly sorted into different families. As can be seen in figure 93, there are clusters dominated by one subgenre – clusters 1, 3, and 6 by sentimental novels and clusters 2, 4, and 7 by historical novels – but there is no cluster containing only novels of one subgenre and the clusters 0 and 5 are entirely mixed. In the combined network, the cluster sizes vary moderately from 7 to 13 novels. Here, too, there is no clear tendency for countries. Different narrative perspectives are not concentrated in single clusters, so this aspect observed for the sentimental novels alone disappears when they are analyzed in the more general setup. Regarding the distribution by years, cluster 1 is early, clusters 3, 4, and 6 are late, and the others are mixed. The topics that are distinctive for the different families reflect the relative purity or mixture of subgenres as well as the preferences of the early versus late nineteenth century.

763To conclude, with the analysis of topics in nineteenth-century Spanish-American historical and sentimental novels in a network-based approach, a proposal was made here how the concept of family resemblance that has been introduced into genre theory in the 1960s and argued for by several genre theorists also recently, can be applied in a digital genre stylistics approach. When one looks at the current strategies to categorize genres in this field, the majority focuses on classificatory groupings based on the assumption of features that are common to all members of a class. Yet, there are also alternative ways to analyze genres in digital stylistics. Especially stylometric network analyses implicitly contain the idea of overlapping similarities and unsharp boundaries characteristic of the family resemblance approach. Here these two scenarios were brought together. With the chosen approach to compare feature distributions of the novels in terms of nearest neighborships and to organize the resulting network of similarities into communities interpreted as families, the original idea of family resemblance is adjusted for the digital analysis. First, because rather than the presence or absence of individual textual features, the degree of their joint presence in individual pairs of novels is decisive, and second, because communities or clusters found in the similarity network constitute a way to delimit the “families” retroactively, without changing the underlying concept of intertwining shared characteristics of individual members of the groups.

764For the Argentine, Mexican, and Cuban historical and sentimental novels, the analysis confirmed that there are subtypes of the subgenres that have been described in literary-historical approaches, such as a novel with a historical setting and a sentimental plot or a historical novel focusing on contemporary political conditions in contrast to novels about historical events about colonial times. In addition, influences of the narrative perspective on subtypes of the sentimental novel became visible. Analyzing both types of subgenres together resulted in mixed groups as well as some that are dominated by one subgenre. While the country the novels were published in does not have a clear impact on the resulting families of novels, the year of publication has an influence in some cases when the preferred and avoided topics reflect the literary development in the nineteenth century.

765The family resemblance analysis is a categorization method that is more open than classification, and if it is applied to more than one type of novel, it allows the novels to be grouped based on criteria other than only their conventional subgenre. All kinds of factors that have an influence on the surface features – as the topics in the analysis conducted here – potentially have an impact on the structure of the family resemblance network if these factors are not controlled beforehand through the composition of the corpus, as was done here with authorship. So depending on how many and which determining factors of style are permitted to enter the network, the family resemblance analysis reveals different stylistic connections between the works, conventional subgenre being one of them. All in all, the results show that features common to all novels of a conventional subgenre cannot be expected and that the textual factors that influence the subgroups or families of subgenres are diverse. There is not one decisive factor, each family has its own traits that hold it together, and inside of each one, there are additional individual traits as well as connections to other families.

766The algorithm producing the family resemblance network and the resulting data offer an empirical ground on which literary historians can look for sense in genre historical terms. The idea of family resemblance that is implemented here is different from Wittgenstein’s metaphor because it is not about resemblances caused by biological kinship or because of the use of the same word for concepts that do not have necessary common semantic features but partial similarities in meaning. It is also not about conventional relationships between the literary works and does not say anything about the historico-cultural and communicative relevance of the connections inside the network. Here the concept is interpreted in terms of textual families. As the family resemblance network is based on stylistic features, it might reveal previously unrecognized textual similarities in addition to confirming known ones on a broader textual basis. By not presupposing strict uniformity inside and strict boundaries between the text types, it might come closer to the multi-faceted genre that the novel is.