Vector-Quantization-Based Topic Modeling

With the purpose of learning and utilizing explicit dense topic embeddings, we propose three variations novel vector-quantization-based models (VQ-TMs): (1) Hard VQ-TM, (2) Soft (3) Multi-View VQ-TM. The model family capitalize on vector quantization techniques, embedded input documents, viewing words as mixtures topics. Guided by a comprehensive set evaluation metrics, conduct systematic quantitative qualitative empirical studies, demonstrate superior performance VQ-TMs compared to important baseline models. Through unique case study code generation from natural language descriptions, further illustrate power in downstream tasks.