A Comparison of Topic Models Based on LDA, STM and NMF for Qualitative Studies of Russian Short Prose

The paper describes the results of topic modelling of short prose fiction based on three methods, namely Latent Dirichlet Allocation (LDA), the Structural Topic Model (STM), and the Non-Negative Matrix Factorization (NMF), combined with different text preprocessing options (all parts of speech vs. only nouns). The experimental design is tested on the basis of the Corpus of Russian Short Stories of 1900–1930s. The research made it possible to determine the specifics of the algorithms under consideration and to assess the effectiveness of their application for the qualitative analysis of fiction texts.