Thrilled to announce that our paper “QDER: Query-Specific Document and Entity Representations for Multi-Vector Document Re-Ranking” has been accepted to SIGIR 2025—the premier conference in information retrieval!
📄 Paper: https://dl.acm.org/doi/pdf/10.1145/3726302.3730065
💻 Code: https://github.com/shubham526/SIGIR2025-QDER
Bridging Two Worlds in Neural IR
Neural information retrieval has evolved along two parallel paths:
- Entity-oriented approaches that leverage knowledge graphs for semantic understanding
- Multi-vector models that capture fine-grained token-level semantics
What if we could combine the best of both worlds? That’s exactly what QDER does.
The Key Innovation: Late Aggregation
Most neural re-rankers compute similarity scores on aggregated embeddings—collapsing rich semantic information into single vectors too early in the process. QDER takes a fundamentally different approach we call “late aggregation”:
Instead of aggregating first and comparing later, QDER maintains individual token and entity representations throughout the entire ranking process, performing aggregation only at the final scoring stage. This preserves fine-grained semantic signals that matter for complex queries.
How It Works
QDER transforms query-document relationships through three key steps:
- Fine-grained representations: Maintain separate token and entity embeddings
- Learned attention patterns: Transform these representations to capture relevance signals
- Precise mathematical operations: Apply carefully chosen aggregation functions for accurate matching
By integrating knowledge graph semantics directly into a multi-vector framework, QDER captures both entity-level concepts and token-level precision.
Results That Speak Volumes
QDER achieves significant performance gains across five standard benchmarks:
- 36% improvement in nDCG@20 over the strongest baseline on TREC Robust 2004
- Exceptional performance on difficult queries: Achieves nDCG@20 of 0.70 where traditional approaches completely fail (nDCG@20 = 0.0)
- Consistent improvements across all tested datasets
Why This Matters
QDER demonstrates that entity-aware, fine-grained semantic modeling can push the boundaries of neural re-ranking. By preserving semantic richness throughout the ranking process and only aggregating at the end, we can build systems that better understand complex information needs.
This work sets a new foundation for entity-aware retrieval and opens exciting directions for future research in neural IR.
Looking forward to presenting at SIGIR 2025 and continuing the conversation with the IR community!