DiVATM: Unsupervised Neural Topic Modeling using Disentangled Variational Autoencoders

Topic modeling has been pivotal in NLP for extracting semantic structures from text corpora. Traditional methods like LDA often struggle with coherence and diversity. We propose DiVATM (Disentangled Variational Autoencoder for Topic Modeling), a neural architecture using VAEs with disentangled latent representations. The DiVATM encoder-decoder framework captures the semantic structure and reconstructs documents from disentangled variables using β -TCVAE, improving interpretability and coherence. Extensive experiments on 20 newsgroups and Reuters-21578 show DiVATM outperforms state-of-the-art models in perplexity and topic coherence. DiVATM achieves a perplexity of 150.2 on 20-Newsgroups and 85.7 on Reuters-21578, with coherence scores of 0.75 and 0.82, respectively. Qualitative evaluations reveal DiVATM generates more distinct and interpretable topics. Ablation studies confirm β-TCVAE contributes to a 20% increase in topic diversity. DiVATM advances unsupervised topic modeling, offering a robust framework for future neural representation learning research.