Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

QDER: Query-Specific Document and Entity Representations for Multi-Vector Document Re-Ranking

Shubham Chatterjee, Jeff Dalton

TL;DR

QDER proposes a unified, query-specific, entity-aware multi-vector re-ranking framework that preserves token- and entity-level representations through a late-aggregation approach. It leverages attention-guided interactions, bilinear cross-pattern scoring, and external lexical signals to produce discriminative query-focused embeddings, significantly improving ranking, especially for difficult queries. Empirical results across five benchmarks show substantial gains over strong baselines, with notable improvements on challenging topics and case studies illustrating robust entity attention dynamics. This work demonstrates that dynamically adapting document representations to each query, and integrating knowledge graph signals within a multi-vector architecture, can substantially advance neural IR performance with practical relevance for retrieval and RAG systems.

Abstract

Neural IR has advanced through two distinct paths: entity-oriented approaches leveraging knowledge graphs and multi-vector models capturing fine-grained semantics. We introduce QDER, a neural re-ranking model that unifies these approaches by integrating knowledge graph semantics into a multi-vector model. QDER's key innovation lies in its modeling of query-document relationships: rather than computing similarity scores on aggregated embeddings, we maintain individual token and entity representations throughout the ranking process, performing aggregation only at the final scoring stage - an approach we call "late aggregation." We first transform these fine-grained representations through learned attention patterns, then apply carefully chosen mathematical operations for precise matches. Experiments across five standard benchmarks show that QDER achieves significant performance gains, with improvements of 36% in nDCG@20 over the strongest baseline on TREC Robust 2004 and similar improvements on other datasets. QDER particularly excels on difficult queries, achieving an nDCG@20 of 0.70 where traditional approaches fail completely (nDCG@20 = 0.0), setting a foundation for future work in entity-aware retrieval.

QDER: Query-Specific Document and Entity Representations for Multi-Vector Document Re-Ranking

TL;DR

QDER proposes a unified, query-specific, entity-aware multi-vector re-ranking framework that preserves token- and entity-level representations through a late-aggregation approach. It leverages attention-guided interactions, bilinear cross-pattern scoring, and external lexical signals to produce discriminative query-focused embeddings, significantly improving ranking, especially for difficult queries. Empirical results across five benchmarks show substantial gains over strong baselines, with notable improvements on challenging topics and case studies illustrating robust entity attention dynamics. This work demonstrates that dynamically adapting document representations to each query, and integrating knowledge graph signals within a multi-vector architecture, can substantially advance neural IR performance with practical relevance for retrieval and RAG systems.

Abstract

Neural IR has advanced through two distinct paths: entity-oriented approaches leveraging knowledge graphs and multi-vector models capturing fine-grained semantics. We introduce QDER, a neural re-ranking model that unifies these approaches by integrating knowledge graph semantics into a multi-vector model. QDER's key innovation lies in its modeling of query-document relationships: rather than computing similarity scores on aggregated embeddings, we maintain individual token and entity representations throughout the ranking process, performing aggregation only at the final scoring stage - an approach we call "late aggregation." We first transform these fine-grained representations through learned attention patterns, then apply carefully chosen mathematical operations for precise matches. Experiments across five standard benchmarks show that QDER achieves significant performance gains, with improvements of 36% in nDCG@20 over the strongest baseline on TREC Robust 2004 and similar improvements on other datasets. QDER particularly excels on difficult queries, achieving an nDCG@20 of 0.70 where traditional approaches fail completely (nDCG@20 = 0.0), setting a foundation for future work in entity-aware retrieval.

Paper Structure

This paper contains 24 sections, 11 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Approach
  4. QDER: Conceptual Foundation
  5. QDER: Technical Architecture
  6. Multi-Channel Representation Processing
  7. Dynamic Attention-Guided Interaction
  8. Multi-Granular Interaction Modeling
  9. Relevance Signal Integration
  10. Cross-Pattern Interaction Scoring
  11. End-to-End Training
  12. Final Hybrid Scoring
  13. Complementary Integration of Relevance Signals
  14. Experimental Methodology
  15. Datasets
  16. ...and 9 more sections

Figures (2)

  • Figure 1: Difficulty test on Robust04 title queries. 5% most difficult queries for BM25+RM3 to the left and the 5% easiest ones to the right. Performance reported as macro-averages across queries. For the most difficult queries (0-5%), relevant documents are promoted to the top of the ranking by QDER.
  • Figure 2: Visualization of document embeddings using t-SNE for QDER (left) and SBERT (right) on TREC Robust 2004 (Title). Relevant documents (red, 19.8%) and non-relevant documents (gray, 80.1%) are shown. QDER produces clear, coherent clusters of relevant documents, while SBERT embeddings display significant mixing. This highlights QDER’s ability to create query-specific representations that naturally separate by relevance, supporting its superior retrieval performance.