Reproduce “REGENT: Relevance-Guided Attention for Entity-Aware Multi-Vector Neural Re-Ranking” on OpenResearch

About this paper

Shubham Chatterjee

Researchers at Missouri University of Science and Technology developed REGENT, a neural re-ranker that integrates token-level lexical signals and query-specific entity representations via a novel relevance-guided attention mechanism. This approach significantly outperforms existing baselines, achieving a 108% relative MAP improvement over BM25 on TREC Robust04 by enabling deeper semantic reasoning for complex queries and long documents.

Key insights

• Explicitly injecting lexical and semantic relevance signals directly into the attention mechanism is crucial for enabling deeper content prioritization, surpassing the limitations of uniform token interaction or post-hoc score fusion.
• Query-specific entities serve as a powerful 'semantic skeleton,' with their removal leading to a catastrophic 74% drop in MAP, indicating their central role in guiding the model's understanding of complex documents.
• Integrating traditional IR signals like BM25 at the token level significantly enhances performance, demonstrating that fine-grained integration is superior to coarse document-level fusion.

Summary by alphaXiv.