Computational thematics: comparing algorithms for clustering the genres of literary fiction

This paper compares various algorithms for unsupervised learning of thematic similarities between literary texts, focusing on their application in automatically clustering book genres. The algorithms are categorized into three steps: text preprocessing, feature extraction, and distance measurement. The study tests numerous combinations of these steps using a corpus of books from four pre-tagged genres (detective fiction, science fiction, fantasy, and romance). Clustering performance is validated against the known genre labels, identifying the best and worst combinations. The difference between the best and worst methods is illustrated by clustering 5000 random novels from the HathiTrust corpus.

Oleg Sobchuk, Artjoms Šeļa