Classical Tricks for Better Topics: Can Traditional Text Mining Techniques Improve Generative Topic Modeling?

Abstract

Topic modeling is widely used to discover and organize latent themes in large text collections. While recent embedding-based methods have improved topic discovery by leveraging dense semantic representations, they typically represent topics as lists of keywords, often providing limited insights into the underlying theme. More recently, Large Language Model (LLM)-based approaches, such as TopicGPT, have introduced generative topic modeling methods capable of producing human-readable topic labels and descriptions. However, the effectiveness of these methods depends heavily on the quality and representativeness of the documents selected to guide topic generation. In this paper, we propose Core-Periphery Sampling (CPS), a hybrid strategy that combines classical text representation and clustering methods with a TopicGPT-inspired pipeline. CPS selects both central documents and boundary instances that capture transitional and potentially informative cases, providing the generative model with an informed and diverse subset for topic induction. Experimental results show that CPS outperforms random sampling, particularly when the underlying document representation yields well-separable clusters. Although BERTopic remains stronger on clustering-oriented metrics, CPS offers an effective, computationally simple mechanism for guiding LLM-based topic generation via informed document selection, resulting in coherent and contextually grounded topic descriptions.