Unsupervised Learning of Plot Structure: A Study in Category Romance
July 19, 2013, 08:30 | Long Paper, Burnett 115
There are two broad approaches to machine plot analysis: annotation-based systems (Lendvai et al. 2010) and formal models of plot (Lakoff and Narayanan 2010). Annotation-based systems are inspired by markup languages such as XML, while formal models of plot are offshoots of artificial intelligence research. This paper proposes a new approach, based on gene sequencing, and derives a model of plot directly from a very large corpus of novels without training or a pre-defined model. The technique reduces novels to their narrative components, classifies these components according to type, then recombines these constituent elements to typify the plots of a group of texts. This technique is applied to an entire genre, the category romance imprint Harlequin Presents.
Harlequin Presents publishes roughly eight books every month, and is probably the most commercially successful fiction genre in the world (Harlequin Company History). The genre can be characterized by recourse to a limited number of types of plot, although there are distinct sub-categories. Most importantly, the genre is available as an ebook, so each novel in the imprint has a definitive edition that is easily subjected to machine analysis. This study uses 1500 novels — over 15 years of Harlequin Presents. This is one of the first studies of popular culture to use machine analysis on an entire genre.
Although the conscription of machines to the task is relatively recent, the study of narrative is not. Traditional narratology can be traced back to Propp’s work on folklore in the early twentieth century (Propp 1968). Propp collected a set of functions that described all possible actions in his collection of folk tales. The plot of any single folk tale could be described as a subset of these functions laid end-to-end. Propp’s work was rejected (Lévi-Strauss 1976a), integrated (Dundes 1997, 47) and then conflated with that of the structuralists, whose work with myths extended Propp’s ideas to cover much more than folklore.
Romance novels have two important parallels with Propp’s folk tales and Lévi-Strauss’ myths. Firstly, all three genres are, or were, contemporary. Propp’s folk tales were a living art form in the early twentieth century (Haney 2009, xiii). Lévi-Strauss recorded many of the oral myths that he later integrated into his theories (Lévi-Strauss 1976b, 35-65). While stretching back 15 years, the most recent Harlequin Presents novels in our sample have been published this month. Secondly, all three genres are curated by others. Propp used a standard edition of folktales and Lévi-Strauss tapped indigenous traditions to define his myths. In our case, Harlequin Presents has been categorized by the publisher. Yet, unlike either folk tales or myths, romance novels have never had an oral form — which makes them ideal for machine analysis.
The technique itself is a modified version of Weighted Gene Co-Expression Network Analysis (Zhang and Hovarth 2005). This technique has been developed to allow mining of gene sequencing information, although the application to written language is a natural extension. Like words, genes are typically redundant, in that many genes signal at once to achieve a desired effect, similar to the manner in which words are collocated when expressing an idea. Natural language data is transformed to resemble gene sequencin information by seg- menting novels into bins and counting the words in each segment. A correlation matrix is then computed, giving the strength of relationship between each word to each other. Words are then clustered together into co-expression networks based on their frequency of co-occurrence.
Networks of genes that frequently co-occur are known as modules, and this terminology is used here to describe collocated words. The behaviour of a module throughout the genre is then typified, giving a cardinal behaviour for all words in the module. External factors, such as author and date of publication can then be related to the modules, to see how they effect the genre. It is this relationship between modules and external data that reveals the most interesting patterns within the genre. Some modules, such as those relating to the status of the hero, are correlated with the beginning of the novel. Other modules, such as those relating to pregnancy or marriage, are strongly correlated with the final segments of a novel. Other modules are related to authorship, and others can be used to classify the entire genre according to narrative strategy.
Unlike purely stylometric studies, modules are typically closely related to theme and incident - concerns directly under the control of an author. Corre- lation of modules to individual authors is not truly useful for authorship dis- crimination, but reflects preferences that an author can be expected to show as they specialise in particular narrative forms or explore certain themes. Similarly, changes in a genre over time can be seen as a direct reaction to external events rather than changes in an author’s internal mental state.
One criticism of traditional narratology is the difficulty it has relating abstract categories back to the mechanics of the writing (Shen 2005, 146). Machine analysis based on annotations or artificial intelligence research both go some way to alleviating this problem. Deriving a model directly from the text eliminates this problem entirely, although it introduces another: modules of words do not always tie closely into our received notions of narrative. In particular, the abstract categories structuralists leveraged to study the similarities between cultures (Lévi-Strauss 1981, 64-66) are not found by this technique. While modules are illustrative of the texts and genres at hand, they do not really generalize beyond them, providing an insight that is deep but not broad.
Broad insights are the specialty of mark-up based and artificially intelligent narrative systems. These other systems have recourse to categories not derived from the texts at hand, and are much more able to draw links between different groups of texts. Mark-up based systems, although they cannot easily scale to working with thousands of novels as we do here, are able to leverage the (often formidable) skills and intelligence of their users. The more formalistic systems, with their pre-programmed categories are also able to generalize from a single genre. This reflects the very different design goals of these approaches: we are concerned here with mere analysis, whereas markup tools are often a form of scholastic augmentation and artificially intelligent systems typically have plot generation as an ultimate aim (Gervás 2012).
Stylometry has typically focused on high-frequency function words to show the mechanics of language at work. Techniques derived from computational biology allow the extraction of thematic and narrative components, and allows these to be related to authorship, date of publication or other external factors. Other approaches to modeling narrative structure have their strengths, but frequently have broader objectives than the analysis of the texts at hand. Weighted Gene Co-Expression Networks sacrifice these goals but provide a flexible method of unsupervised learning of narrative structure.