A disentangled multimodal neural topic model

This study focuses on multimodal topic modeling and attempts to separate public topics (shared across modalities) from private topics (unique to each modality) hidden in text and image data. To address this issue, we propose a novel Disentangled Multimodal Neural Topic Model (DMNTM). Specifically, we design the modality-specific encoder with an independence constraint to capture private topics, and the public encoder with a product-of-experts module to extract cross-modal shared topics. We conduct extensive experiments on six public datasets, including multimodal online reviews from Amazon, posts from Flickr, tweets from Twitter, and webpages from Wikipedia. Compared with state-of-the-art methods, we find that DMNTM significantly improves topic modeling performance in terms of perplexity, coherence, diversity, and topic quality over the best baseline. In two downstream tasks, including recommendation and sentiment classification, DMNTM further improves the performance. These results show that disentangling public and private topics effectively enhances both the quality and utility of multimodal representations. © 2026 Elsevier Ltd.