Topic Modelling Soviet Periodicals: Methodological Cautions from a Ten-Year Corpus

Latent Dirichlet allocation has become a default tool for exploring large historical corpora, but its outputs are sensitive to preprocessing choices that are rarely reported. Using a corpus of 48,000 articles from regional Soviet periodicals published between 1961 and 1971, we show that stopword list selection alone shifts topic composition enough to reverse a substantive interpretive claim about the periodisation of agricultural discourse. We propose a minimal reporting standard for computational text analysis in historical research.