Dynamic topic models for retrospective event detection: A study on soviet opposition-leaning media

Abstract

In recent years, there has been an increasing interest in digital humanities. This interest is justified by the development of natural language processing tools and the emergence of digitized text collections of documents in different fields of knowledge, for example, literature, art, philosophy, and history. In this paper, we applied unsupervised topic modeling to the Bulletin of Opposition, the journal of Soviet opposition published by Trotskyists in Paris from 1929 to 1941, to analyze the main trends in the Russian opposition-leaning media. We identified topic classes using models based on Latent Dirichlet Allocation and examined Dynamic Topic Models as a tool to single out the main issues of interest for historical research. Applying topic modeling and statistical methods, we proposed an approach to Retrospective Event Detection that was evaluated on a human-annotated set of historical news items. The present study may help to improve event detection on smaller text corpora.