GraphCycle-CLTM: A Cycle-Consistent Graph Contrastive Framework for Neural Topic Modeling

Abstract

Neural topic models have recently advanced through contrastive learning and graph-structured semantic priors, yet several fundamental limitations remain. First, existing contrastive objectives often rely on frequency-based or heuristic negative sampling, which introduces semantic bias and weakens latent topic separation. Second, contrastive learning is typically restricted to the encoder, creating a structural mismatch between discriminative document representations and the generative decoder that produces topic-word distributions. Third, although graph-based topic models effectively propagate semantic information over document-word or word-word graphs, they lack an explicit mechanism to enforce consistency between propagated graph semantics and the decoder's topical structure. To address these challenges, we propose GraphCycle-CLTM, a unified cycle-consistent graph contrastive framework for neural topic modeling that tightly couples representation learning with generative topic modeling. The framework integrates three mutually reinforcing components. First, a graph-guided contrastive sampler constructs semantically balanced positive and negative pairs via multi-hop propagation over document-word and word co-occurrence graphs, enabling structure-aware and unbiased contrastive learning. Second, a cycle-consistent topic reconstruction module enforces bidirectional agreement between original and reconstructed topic-word distributions, effectively bridging the gap between encoder-level representations and decoder-level generation. Third, a topic-aware graph propagation layer injects explicit topical priors into document embeddings through a tri-partite semantic graph, ensuring that message passing remains aligned with learned topic semantics. Extensive experiments on eight benchmark datasets demonstrate that GraphCycle-CLTM consistently outperforms state-of-the-art neural, graph-based, and contrastive topic models across topic coherence, topic diversity, document clustering quality, and downstream classification accuracy.