This study applies non-negative matrix factorization (NMF) and latent Dirichlet allocation (LDA) to extract topics from interviews with Russian civil servants. This study aims to identify and analyze the predominant topics in Russian officials’ public oral speech during the first quarter of the 21st century. Using a corpus-based approach, the research examines both the unified corpus and its four sub-corpora – each representing distinct occupational groups – to detect overarching topics as well as unique, role-specific ones. The analysis seeks to determine the most effective model for topic extraction and comparative assessment across these datasets. The material of the study is represented by transcripts of 90 randomly sampled interviews containing 252.225 words. We conducted lemmatization via pymystem3, developing two configurations for each method: one incorporating all denominative parts of speech except numerals and another restricted to nouns. To estimate the optimal number of topics and training epochs, we used the logarithm of perplexity for LDA and the coherence score for both methods. We concluded that both NMF and LDA perform better when trained on denominative parts of speech than on nouns only. For both data preprocessing techniques, the strongest results were achieved through the utilization of NMF. The average coherence score increased by 14.6% (from 0.363 to 0.425) for LDA and by 4% (from 0.611 to 0.638) for NMF. The application of both models allowed us to identify 6 main topics within the entire interview corpus: ‘citizenship’, ‘justice’, ‘real estate’, ‘arbitration’, ‘state body operations’ and ‘finance’.