Genres in CBF
The Corpus of British Fiction operates with eight genres of contemporary, adult fiction: adventure, crime, general, historical, humour, romance, spy and war. These genres are quite well established (see Wikipedia), together with other genre labels such as fantasy, horror or science fiction and more recently, gay and/or LGBTQ fiction. Such labels are are not however, as far as I know, defined in such a manner as to clearly distinguish them from each other programmatically.
(Note, moreover, that within these fairly broad genre labels, there are, of course also, several sub-genres, not least for crime (detective fiction, mystery, thrillers, cosy crime, etc.).)
In what follows, I will discuss the CBF genre labels in light of how they compare to each other, i.e., is there any evidence to suggest that the texts labelled e.g. crime belong together? Remember that the genre labels are based on information about the individual texts (books, novels) found on the web, mostly on Wikipedia (see e.g. how The Crime at Black Dudley is defined).
How I tried to find this out
The first step in my procedure was to calculate the CS scores of all the texts in the corpus compared to all the other texts. I then removed the scores where a text had been compared to itself and where it had been compared to other texts by the same author, as this would boost the score for those texts, as shown in an earlier post. In the next step, I removed all comparisons with scores below 0.3 (moderate similarity), since such low scores indicate that the texts compared have little in common (this is interesting in itself, especially if a text labelled, let's say, crime, never even got an even moderate similarity to other crime texts in the corpus).
Which texts that were removed in this manner will be different depending on whether the basis for the CS scores is word, lemma, parts of speech (POS) or semantic tag (SEM). The final step consisted of counting the number of times the genre of the current text was similar to the genres in each of the other texts that had reached a CS score of >= 0.3 and do this for all the texts. So, for instance, if texts labelled crime reached the score of 0.3 when compared to other texts labelled crime more often than when compared to texts labelled, e.g., spy or general, then we may claim that texts labelled crime in this corpus have something in common that groups them together (more often than when grouped with texts of other genres). This "something" is, as we have seen the words, lemmas, POS or SEM tags.
I will show these scores in tables consisting of lemmas, parts of speech, semantic tags and words.
Results
Table 1 shows the result of the first comparison based on lemmas. The percentages in bold show, with one exception, that all the genres have more in common with texts with the same label than with texts with other genre labels, when counting, as I have explained, texts comparisons with a CS score of >= 0.3. Disregarding the exception, humour, the percentages range from 52.1 (historical) to 24.1 (war). Note, however, that war reaches a percentage of 24.1 when compared to general as well.
Table 1. Genre comparisons based on lemmas: percentages
The humour genre stands out as being the genre that can be argued to have more in common with all the other genres, especially crime and general, than with itself. The lemma inventory of this genre is not very coherent or similar it seems.
At the other end we have historical, which, it seems, is a quite coherent genre, and which have little in common with the other genres. Pretty much the same can be said of romance.
These percentages, however, only tell part of the story and an important factor is the raw frequencies behind them.
Table 2. Genre comparisons based on lemmas: raw frequencies
The rather low numbers / raw frequencies for adventure, humour and war make the comparisons including these genres less certain and more speculative than, e.g. comparisons including crime and general. This caveat also applies to the next tables, where I will only show percentages. Tables 3-5 contain similar kinds of percentages when comparing semantic tags, words and parts of speech.
Table 3 shows tendencies similar to those for the lemma comparisons; humour seems to have even more in common with general when we compare the CS scores for semantic tags; note also that adventure and historical seems more similar than adventure compared to itself.
Table 3. Genre comparisons based on semantic tags: percentages
Using words as input to the model and as basis for the CS comparisons confirm the content of tables 1 and 3, but now humour seems to conform more to the general picture of being more like itself than other genres, even though general is again a strong contender for the top percentage/most similar.
Table 4. Genre comparisons based on words: percentages
Finally, Table 5 shows how using parts of speech as the basis for comparison falls out. This is, as perhaps would be expected, looking a bit different (see the percentages in blue). In this table crime has more in common with humour, adventure with itself and spy novels, general and romance with adventure and war with humour. Only (adventure), historical, humour and spy seem to follow the normal patterns of being most like itself.
Table 5. Genre comparisons based on parts of speech: percentages
To sum up and coming back to the question we posed in the title, there seems to be something to be said for the genre labels we have attached to our texts, that is, the genres we have chosen seem to be internally stable in the sense that they are more similar to other texts of the same genre than to texts of other genres (see Tables 1, 3, and 4).
A finale note to ponder: Can it be that words rather than e.g. lemmas are better at distinguishing genres / grouping similar texts?