External information enhancing topic model based on graph neural network

Abstract

In the digital age, social media platforms have seen a surge in user-generated content, particularly short-form we-media content. Traditional topic modeling methods often struggle to effectively analyze such content due to their limited generalization ability and interpretability. To address this issue, we propose the Co-occurrence Graph Topic Model (COGTM), a novel approach designed to enhance topic modeling in the context of long-short text co-occurrence scenarios. COGTM leverages the inherent interconnectedness between short and associated long-texts, as well as semantically similar words, within the text corpus. By incorporating these associations into the modeling process, COGTM aims to capture richer semantic information and improve the interpretability of the learned topics. Empirical analysis demonstrates that COGTM outperforms baseline models in various text classification and clustering tasks. By effectively capturing the latent associations between different types of text elements, COGTM offers a promising approach to topic modeling in scenarios involving diverse and interconnected textual data.