A Comparative Study of Topic Models for Student Evaluations

Deciphering valuable insights from unstructured written comments in student evaluations within higher education poses a significant challenge, calling for advanced analytical tools.Previous research has underscored the importance of employing topic modeling as a vital tool for unraveling complexities in large volumes of unstructured textual data and extracting dominant topics.While Latent Dirichlet Allocation (LDA) has been a longstanding choice in past studies, the emergence of new and modern topic modeling techniques demands a comparative analysis of their performance against LDA when used with student evaluations.This paper evaluates four topic modeling methods -Latent Dirichlet Allocation (LDA), Nonnegative Matrix Factorization (NMF), BERTopic, and Top2Vec -based on two key performance metrics: topic coherence and topic diversity.Using data from Rate My Professors, our assessment overwhelmingly endorses BERTopic, showcasing its superiority in producing the most informative topics, unmatched easeof-use, and overall functional excellence for extracting valuable insights from written comments in student evaluations.