Dual-View Reranking Framework for Multi-Hop Document Retrieval
A new cascaded reranking framework, DualView, addresses the challenge of multi-hop question answering by efficiently identifying minimal relevant document sets from retrieved candidates. The architecture features a Local Scorer using stacked cross-attention for fine-grained query-document relevance and a Global Scorer that models inter-document dependencies via Transformer-based context aggregation. An Adaptive Gate dynamically fuses these views based on query semantics. Operating as a lightweight post-retrieval stage over E5-base-v2 candidates, the model achieves competitive results under fixed candidate set reranking with offline cached embeddings, particularly outstanding on the MuSiQue dataset with a score of 99. The work is published on arXiv with ID 2605.18767.
Key facts
- DualView is a dual-view cascaded reranking framework for multi-hop document reranking.
- It comprises a Local Scorer with stacked cross-attention and a Global Scorer with Transformer-based context aggregation.
- An Adaptive Gate dynamically fuses the local and global views conditioned on query semantics.
- The model operates as a lightweight post-retrieval stage over E5-base-v2 candidates.
- It uses offline cached embeddings under a fixed candidate set reranking setting.
- Achieves competitive results, particularly outstanding on MuSiQue with a score of 99.
- Published on arXiv with ID 2605.18767.
- Addresses multi-hop question answering requiring information aggregation from multiple documents.
Entities
Institutions
- arXiv