Sub-document neural topic model

doi:10.1016/j.knosys.2025.114762

Sub-document neural topic model

2025-10-30

0

PRE

AI

DOI：10.1016/j.knosys.2025.114762

Original

Original request for help

Save

Abstract

En 中文

Neural Topic Models (NTMs) generally fall into two categories: Variational Autoencoder (VAE)-based and clustering-based frameworks. While recent advancements have improved topic quality and document-topic associations, significant challenges remain. Many state-of-the-art models now enhance performance by leveraging external contextual information, such as graph structures and pre-trained language models (PLMs), which in turn demand additional resources and auxiliary components. In this paper, we propose a novel framework called Sub-document Neural Topic Model (SubNTM), which enhances topic modeling by partitioning documents into sub-documents and learning topic distributions at both full-document and sub-document levels. This approach enriches contextual understanding and improves topic coherence. Furthermore, the dual-level inference mechanism effectively enhances the quality and granularity of document-topic distributions. Experimental evaluations on widely used benchmark datasets demonstrate that SubNTM consistently outperforms state-of-the-art topic modeling approaches across multiple evaluation metrics.

Sub-document neural topic model

Abstract

Journal

Organization