Towards Interpreting Topic Models with ChatGPT

Abstract

Topic modeling has become a popular approach to identify semantic structures in text corpora. Despite its wide applications, interpreting the outputs of topic models remains challenging. This paper presents an initial study regarding a new approach to better understand this output, leveraging the large language model ChatGPT. Our approach is built on a three-stage process where we first use topic modeling to identify the main topics in the corpus. Then, we ask a domain expert to assign themes to these topics and prompt ChatGPT to generate human-readable summaries of the topics. Lastly, we compare the human-and machine-produced interpretations. The domain expert found half of ChatGPT's descriptions useful. This explorative work demonstrates ChatGPT's capability to describe topics accurately and provide useful insights if prompted accurately.