Automated topic analysis for restricted scope health corpora: methodology and comparison with human performance

Abstract

This paper addresses the problem of identifying topics which describe information content, in restricted size sets of scientific papers extracted from publication databases. Conventional computational approaches, based on natural language processing using unsupervised classification algorithms, typically require large numbers of papers to achieve adequate training. The approach presented here uses a simpler word-frequency-based approach coupled with context modeling. An example is provided of its application to corpora resulting from a curated literature search site for COVID-19 research publications. The results are compared with a conventional human-based approach, indicating partial overlap in the topics identified. The findings suggest that computational approaches may provide an alternative to human expert topic analysis, provided adequate contextual models are available.