The Cryptic Novel: A Computational Taxonomy of the Eighteenth-Century Literary Field

Overview
This project employs machine learning and other text-mining based clustering techniques to study the relationship between taxonomic systems for categorizing prose fiction in the eighteenth century. Our goal is to challenge narratives of the so-called “rise” of the novel (Watt, 19571; McKeon, 19872; Richetti 19993) that trace a continuity between the prose fiction that emerged in the eighteenth-century and the novel as a fully realized critical object in the late nineteenth century. Our project instead uses statistical modeling methods to theorize an alternate history for the novel: not, as proposed in Franco Moretti’s Graphs, Maps, Trees4, as a history of genres, but instead as a history of genre labels--that is, a history formed by the interaction between paratextual, self-identified literary labels that circulated in the eighteenth-century marketplace and the texts that they defined. Our project, which is currently in its final stages at Stanford University’s Literary Lab, bridges the historical and computational to enact a form of literary history that puts pressure on traditional theories focused on charting the rise of the novel. The project is deeply invested in using digital and statistical methods to excavate the nuances in the ways in which eighteenth-century genres evolved. Instead of analyzing texts via contemporary genre categorizations, we recover the relationship between the self-applied taxonomy of eighteenth-century prose fiction and the lexical and semantic features of the texts themselves to recuperate a historically-determined literary field that has been largely overlooked (or noted and elided) by modern criticism. Drawing from a corpus of 2,385 digitized texts from the Eighteenth Century Collections Online (ECCO) database, we ask how new computational methods can aid us in uncovering what kinds of work self-identified genre labels in titles (specifically “story,” “history,” “tale,” “letters,” “romance,” “life,” “adventure,” and “novel”) accomplished within the literary marketplace of the long eighteenth-century, and how that kind of work is distinct from critically- or retroactively-designated genres (“gothic,” “jacobin,” “epistolary,” “historical,” “it-narrative,” “oriental tale.”)

Aims
Our project seeks to answer a series of critical questions about the taxonomic systems of eighteenth-century fiction using computer generated models. Do generic labels formally differentiate separate kinds of writing in a useful way (such that a “novel” is lexically, semantically or formally different from a “romance”) or do they merely function as signals within the marketplace itself, so that there is no textual or formal differences between a “tale” or a “story” beyond mere marketplace convention? How does the relationship between different genre labels change over time, and how can it assist us in understanding the evolution of titles as a representative labeling system? This approach takes advantage of recent advances in probabilistic modeling to recover the meaning of labels that have been homogeneously condensed under the rubric of novel; it simplifies the complexity of the word “novel,” with all its attendant genres and subgenres, to make the field that “novel” inhabits more complex. Developing from these research questions, our method-driven aims are: (1) to detect and assess large-scale trends in the development of and relation between genre labels across time (2) to isolate formal differences in the corresponding full texts identified under these categorical labels, and (3) to compare these differences to a corpus of texts categorized according to modern genre designations.

Methods and Narrative
In the early stages of our project, we began with an exploratory approach to our data, using statistical, unsupervised learning techniques to identify word patterns and clusters within the texts, prior to classifying the texts’ content onto the label categories. Our initial approach sought to determine if some underlying structure existed among these texts that we could then map onto a taxonomic system, such as genre, or in our case, title labels. Using a series of clustering techniques built off of different feature sets (most frequent words, most distinctive words, etc), we found that the assumptions that provided the foundation for our project were vindicated: not only did the titles under each label cluster together based on the Most Frequent Words in each text, but there were intriguing differences between clusters. For instance, though a coherent cluster of “history” is present from the first quarter of the century, definitively clustered labels “novel” and “romance” don’t emerge until later in the century. These differences opened our research to a new set of concerns, such as the relative stability of each category, or the ways in which the labels individually and collectively change across time. Indeed, a major finding of this phase was the dramatic influence time itself had over every other variable we considered. Such a finding strengthened the presence of a literary-historical component of our research and prompted us to pay closer attention to the divisions of time we were employing.

In our current work, we have employed supervised learning techniques and machine classification to specifically interrogate the relationship between text and label: which labels have a high level of cohesion, and thus predictability, and which labels are less cohesive, and were perhaps more tolerant of formal deviation and experimentation? To answer this question, we employed a discriminant function analysis to determine if texts could be reliably algorithmically classified into their labels, treated in our analysis as a taxonomic system. These results were validated through a leave-one-out cross validation to determine the strength of our predictive model. . The results confirmed the findings of the PCA: our model performed better than chance, categorizing each of our labels, on average, 75% correctly. Another dimension of our study was to investigate differences on a label-by-label basis: what does it mean, textually, that a “life” is harder to algorithmically classify than a “romance”? Did difficulties in categorization have anything to do with the textual heterogeneity within the labels themselves? These questions prompted us toward another set of analyses designed to evaluate the self-similarity of each label. We divided each text in our corpus into ten equal parts, measuring the distance from each part to the other in a matrix that, when averaged, resulted in a distance score for each. The findings from this activity helped to both clarify and complicate the story that was emerging regarding labels over the course of the century.

Our most recent work has taken us in a more broadly comparative direction: to examine the textual and larger structural differences between our current corpus of texts, classified according to generic labels, and another corpus we have compiled from modern, genre-focused bibliographies. Using the same techniques of machine learning and self-similarity, we have started to interrogate the logics by which generic labels organize and describe the literary field of the eighteenth century in ways we have yet to discover.

Results
Our initial results proved the value and worth of studying this type of paratextual literary categorization. The temporalization of data allowed by our method lets us observe, first, the simultaneous dominance and differentiation of “novel” and “tale” at the end of the century, leading us to speculate, as Anthony Jarrells (2012)5 has, on twin “rises” of genre in this century as opposed to one rise; that is, the eighteenth-century appears to be as much a story of the rise of “novel” as it is of the rise of “tale.” The result of our machine learning classifications and self-similarity form a compelling argument for the ways in which, throughout the eighteenth-century, the field of prose fiction underwent a transformation through the processes of both differentiation and consolidation. From these results, we argue that the novel can be viewed as merely one branch, rather than the single formal end to which all eighteenth-century prose fiction trends. Other important results, observed from our PCA charts, indicate a general movement across all generically-labelled texts from the usage of words that indicate exteriority (such as “law,” “city,” “army,” “money,” “order,” and “public”) to words that indicate interiority (such as “lover,” “marry,” “imagine,” “woman,” “write” and “read”). In the final stages of the project, we are working to apply similar computational tools to genres conventionalized by modern criticism -- such as the Gothic novel or Oriental tale -- to examine how such categorizations relate to self-applied genre labels. Preliminary results indicate that, after sampling our texts and controlling for time period and author, generic labels classify with a success rate comparable to texts designated as such by critics. The question then becomes not whether genre labels organize more successfully than conventional genres, but whether and how they organize differently--if generic labels can tell us something about the trajectory and organization of formal literary history that genre cannot. Our presentation will include a more detailed discussion of our methods and results and of the implications of the latter for traditional teleological narratives of eighteenth-century prose fiction

References
1. Watt, Ian (1957). The Rise of the Novel: Studies in Defoe, Richardson, Fielding. London: Chatto and Windus.

2. McKeon, Michael (1987). The Origins of the English Novel: 1600-1740. Baltimore: Johns Hopkins UP

3. Richetti, John (1999). The English Novel in History, 1700-1780. London: Routledge.

4. Moretti, Franco (2004). Graphs, Maps, Trees: Abstract Models for Literary History. New York: Verso

5. Jarrells, Anthony (2012). After Novels: Short Fictional Forms and the Rise of the Tale. The Oxford History of the Novel in English. 12 vols. Vol. 2. Eds. Peter Garside and Karen O'Brien. Oxford: Oxford UP.