Title: Separating Similarity from Relatedness in Topic Modeling Frameworks

Abstract:

The rising prominence of large pre-trained language models (PLMs) has spurred their incorporation into topic modeling techniques. Yet, PLM-enhanced topic models diverge from traditional co-occurrence approaches like Latent Dirichlet Allocation (LDA) not merely in efficiency, but in the specific nature of the semantic relationships they identify. This study clarifies that distinction by mapping it onto two psycholinguistic dimensions: thematic relatedness (e.g., dog/bone) and taxonomic similarity (e.g., dog/wolf). To quantify these dimensions across topic vocabulary, we developed a substantial synthetic benchmark comprising word pairs annotated via large language models, subsequently training a neural classifier on this dataset. Our results, consistent across various corpora and model architectures, reveal that different topic modeling families occupy unique positions within the combined similarity-relatedness landscape. Furthermore, these metrics serve as predictors for downstream application success: tasks dependent on similarity yield better results with similarity-dense topics, while those relying on relatedness benefit from the opposite trait. However, overemphasizing one dimension can actually impair performance on tasks that require the other. Since neither axis offers universal advantages, assessing both provides a robust, model-independent diagnostic tool for analyzing the semantic composition captured by topic models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC