Reproduce “BP-Seg: A graphical model approach to unsupervised and non-contiguous text segmentation using belief propagation” on OpenResearch

About this paper

Fengyi Li, Kayhan Behdin, Natesh Pillai, Xiaofeng Wang, Zhipeng Wang, Ercan Yildiz

BP-Seg groups sentences by combining sentence embeddings’ similarities to candidate group representatives with semantic relatedness between sentence pairs through belief propagation. It permits non-adjacent sentences to share a group. On the Choi dataset’s 6–8 subset, it reports mean ARI 0.76 and NMI 0.89, above GraphSeg and k-means; however, GraphSeg leads in 3–5, and tests are limited to two synthetic English cases.

Key insights

• Graphical models, coupled with Belief Propagation, provide a flexible probabilistic framework for text segmentation that can effectively integrate both local coherence and global semantic similarity.
• Allowing for non-contiguous segmentation significantly improves the ability to capture complex semantic structures where related information may be scattered throughout a document.
• The unsupervised nature of BP-Seg makes it applicable without requiring labeled training data, a common challenge in text segmentation tasks.

Summary by alphaXiv.