Genre Classification of Russian Texts Based on Modern Embeddings and Rhythm

Abstract

This article investigates modern vector text models for solving the problem of genre classifying Russian-language texts. The models include ELMo embeddings, a pretrained BERT language model, and a set of numerical rhythmic characteristics based on lexico-grammatical tools. The experiments have been carried out on a corpus of 10 000 texts in five genres: novels, scientific articles, reviews, posts from the VKontakte social network, and news from OpenCorpora. Visualization and analysis of statistics for rhythmic characteristics have made it possible to distinguish both the most diverse genres in terms of rhythm (novels and reviews) and the least (scientific articles). It is these genres that are subsequently classified best using rhythm and the LSTM neural network classifier. Clustering and classifying texts by genre using the ELMo and BERT embeddings make it possible to separate one genre from another with a small number of errors. The multiclassification F-measure reaches 99%. This study confirms the effectiveness of modern embeddings in the tasks of computational linguistics and highlights the advantages and limitations of the set rhythmic characteristics on the genre classification material.