Speeding up calibration of latent Dirichlet allocation model to improve topic analysis in software engineering

Extraction of topics from large text corpuses helps improve Software Engineering (SE) processes. Latent Dirichlet Allocation (LDA) represents one of the algorithmic tools to understand, search, exploit, and summarize a large corpus of data (documents), and it is often used to perform such analysis. However, calibration of the models is computationally expensive, especially if iterating over a large number of topics. Our goal is to create a simple formula allowing analysts to estimate the number of topics, so that the top X topics include the desired proportion of documents under study. We derived the formula from the empirical analysis of three SE-related text corpuses. We believe that practitioners can use our formula to expedite LDA analysis. The formula is also of interest to theoreticians, as it suggests that different SE text corpuses have similar underlying properties.