Improving short text topic modeling through fairness attention and semantic-enhanced

The deep engagement of online users has generated massive volumes of short text data, making it crucial to deeply understand the topical information in user-generated data. However, traditional topic modeling techniques face significant challenges in user data scenarios, including data irregularity, semantic sparsity, and lack of contextual information, which hinder their ability to effectively capture latent topic structures. To address these challenges, this paper proposes the Fairness Attention and Semantic-enhanced Topic Model (FASTM), a novel topic modeling approach that integrates Contextual Semantic Enhancement (CSE) and Fairness Attention Enhancement (FAE). On one hand, by constructing semantic neighbor networks, the model incorporates semantic associations into the topic assignment process, thereby alleviating data sparsity issues. On the other hand, through dynamic weight allocation and multi-objective optimization mechanisms, the model learns word importance while preventing excessive attention concentration, which enhances topic quality and ensures fairness in topic distribution. Experimental results on the THUCNews and Weibo datasets demonstrate that FASTM outperforms baseline methods in addressing semantic sparsity and semantic ambiguity, effectively improving the overall performance of topic modeling.