Topic modeling for short texts: comparative analysis of algorithms

Abstract

The steady increase in the popularity of social media as a means of communication actualizes methodological issues related to processing of short texts with less semantic context than large corpora, which are widely used for training and testing machine learning models for textual data. Topic modeling, an unsupervised machine learning technique aimed at aggregating texts into topic clusters, has many academic and practical applications where information on true groupings of texts is not available. However, the performance of topic modeling algorithms may be limited by requirement of a sufficient semantic context for a high-quality numerical representation of a unit of text, which may not be derived effectively from a short document. This paper is dedicated to discussing 6 different approaches to topic modeling, comparing their performance on a set of Russian-language comments on TikTok and formally evaluating their performance based on speed and coherence of the resulting topics.