I’m excited to announce that our paper “REGENT: Relevance-Guided Attention for Entity-Aware Multi-Vector Neural Re-Ranking” has been accepted to SIGIR-AP 2025!
📄 Paper: https://arxiv.org/abs/2510.11592
💻 Code: https://github.com/shubham526/SIGIR-AP-2025REGENT
What’s REGENT?
When you search through long, complex documents, you naturally focus on key entities and concepts—names, places, organizations—that help you understand what’s relevant. But most neural search models don’t work this way. They process text in fixed windows, treating every word interaction equally and missing the semantic signals that actually matter.
REGENT changes this by teaching neural re-rankers to think more like humans. It uses entities as a “semantic skeleton” to guide the model’s attention, helping it identify what’s truly important in lengthy, multi-faceted documents. By integrating relevance guidance directly into the attention mechanism, REGENT combines precise term matching with high-level semantic reasoning—focusing on conceptually important content without losing sensitivity to exact matches.
Results
The results speak for themselves: REGENT achieves state-of-the-art performance across three challenging datasets, with improvements of up to 108% over BM25. It consistently outperforms strong baselines including ColBERT and RankVicuna.
This is the first work to successfully integrate entity semantics directly into neural attention, opening up a new paradigm for entity-aware information retrieval.
Looking forward to presenting this at SIGIR-AP 2025 and discussing with the IR community!