The sparsity of short text data and the issue of repeated topics are significant challenges in topic modeling. Traditional topic models struggle to generate high-quality topics when dealing with short texts, due to the brevity of the text and the lack of sufficient context. To address this issue, this paper proposes a Short Text Topic Modeling (STTM) model based on vector quantization. The model converts short texts into low-dimensional representation vectors using an encoding module, enhances the distinction of topic distributions with a vector quantization module, and reduces the generation of repeated topics with a negative sampling decoding module. Experimental results show that the STTM model significantly outperforms traditional Latent Dirichlet Allocation (LDA) models and other neural network models in terms of topic coherence and diversity on the Chinese Twitter dataset. The model can more accurately extract users’ areas of interest, providing effective support for personalized recommendations, advertising targeting, and sentiment analysis. Future research can further optimize the automatic selection mechanism for the number of topics and explore its generalizability across other languages and platforms.