Topic Modeling Literary Quality

3.1. Quantitative analysis

We aim to gain insight into the distribution of topics in relation to the literary ratings of the novels (predicting literary ratings is not the main aim here). In order to interpret the topic distributions, we introduce the concept of mono-topicality. ² A mono-topical novel contains little diversity in topic distribution, which means that one or two topics are dominant throughout the novel. A novel which shows more variation in topics has a more even distribution of topics, i.e., such a novel has a larger topic diversity. Fig. 2 shows an example of both cases.

The x-axis shows the distribution of topics, sorted from least to most prevalent. In John Grisham’s The Appeal, topic 5 (“lawsuits”) has a proportion of 47.8 % of all 50 topics. This novel is more mono-topical than the Franzen’s Corrections, which has a more balanced distribution of topic proportions.

We hypothesize that the less mono-topical a novel is, the higher the perceived literariness by readers will be. And indeed, Fig. 3 shows that there is a statistically significant correlation between the diversity of topics of a book and its perceived literariness. Books with a single, highly prominent topic, such as Grisham’s, tend to be seen as less literary.

Figure 2: Distribution of the top 15 topics in novels with high (left) and low (right) mono-topicality

Table 1: Themes in fifteen highly literary novels, all of which are originally Dutch Table 2: Six topics from the model that address themes present in the fifteen highly literary novels, but which are not the most prominent as topics in those novels

Figure 3: Correlation between share of themost prominent topic per book and mean literariness ratings

Figure 4: Correlation between topic 29 proportion and mean literariness ratings

3.2. Interpretation

There are several possible explanations for the correlation. Genre novels could have a tendency to single out certain topics, as they deal with more ‘typical’ or genre-specific subject matter than do general novels. If this were the case, we would simply be finding that genre novels are considered to be less literary than general novels, and this would tell us little about literary quality in a more general sense. General novels in the other hand, deal with all sorts of subjects and themes, across and beyond ‘genre’ subjects, and therefore a topic model may not be able to single out thematic groups of words common to these novels, and thus may not find one single prominent topic. A third explanation could be that highly literary novels do deal with specific themes and subjects which are also part of genre novels, but that these are described in wordings that are more implicit or subtle, and therefore do not come up as single, clear topics. If this were the case, that would bring us closer to an explanation of what topics have to do with literary quality. These explanations are not mutually exclusive and we will explore the topic model here to examine the value of the second and third explanation.

The topic that shows the highest correlation (r=0.49) with literary appreciation is topic 29; cf. Fig. 4. This topic is most prominent in fifteen originally Dutch general novels. The twenty words in topic 29 with the highest weights are begin, music, play, occasion, first, the first, sing, only, year, one, stay, sometimes, even, new, own, always, high, exact(ly), bike, appear. They show little coherence, making it hard to interpret their context, although ‘music and performance’ appears to be part of it. To find out more about the novels in which this topic is prominent, we consult a Dutch website maintained by librarians called Literatuurplein, which provides information on the themes and content of Dutch novels.

Most of these novels show similarities in themes, such as family relationships. In ten of the novels the protagonist has an artistic profession: a couple of writers, a painter and a stand-up comedian. None of them has a musical or acting career, despite the 'music and performance' words; and vice versa, none of the twenty most prominent words concern writing. All in all, at first glance topic 29 seems not to address the themes and content of the novels, whereas most other topics in the model do concern specific themes (cf. Fig. 1 and Table 2).

For instance, topic 2 and 11 address family relations, topic 12 and 41 are about writing novels, and topic 6 and 33 concern health issues. These topics are present, but as smaller topics. This shows that the second explanation, of the general novels not sharing themes, is not valid. It could be an indication though that the highly literary novels indeed use a more subtle way of describing themes similar to other novels in our corpus, our third explanation. As a final note, in topic 29 there are proportionally more adverbs than in the other topics mentioned, which contain more nouns. Perhaps this shows that style is a more shared element in literary novels than the choice of words. In other words, this brief analysis shows that there is merit to our third explanation. This will therefore become a new hypothesis for further research.