Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CDER: Collaborative Evidence Retrieval for Document-level Relation Extraction

Khai Phan Tran, Xue Li

TL;DR

This work tackles document-level relation extraction (DocRE) with a focus on evidence retrieval (ER). It introduces CDER, a collaborative ER framework that builds a document-level bipartite graph with three sub-graphs and employs Entity Pair-aware GATs (EP-GATs) to model cross-pair collaboration and evidence dynamics. Key innovations include a dynamic Entity Pair sub-graph $G_{PP}$ guided by relation similarity, EP-GAT layers for $G_{PS}$ and $G_{PP}$, and a focal loss to handle severe evidence imbalance, which together yield strong ER performance and improve downstream DocRE. Experiments on DocRED show that CDER not only achieves state-of-the-art ER metrics but also provides substantial gains when plugged into existing DocRE systems, underscoring its practical impact for robust, scalable information extraction.

Abstract

Document-level Relation Extraction (DocRE) involves identifying relations between entities across multiple sentences in a document. Evidence sentences, crucial for precise entity pair relationships identification, enhance focus on essential text segments, improving DocRE performance. However, existing evidence retrieval systems often overlook the collaborative nature among semantically similar entity pairs in the same document, hindering the effectiveness of the evidence retrieval task. To address this, we propose a novel evidence retrieval framework, namely CDER. CDER employs an attentional graph-based architecture to capture collaborative patterns and incorporates a dynamic sub-structure for additional robustness in evidence retrieval. Experimental results on the benchmark DocRE dataset show that CDER not only excels in the evidence retrieval task but also enhances overall performance of existing DocRE system.

CDER: Collaborative Evidence Retrieval for Document-level Relation Extraction

TL;DR

This work tackles document-level relation extraction (DocRE) with a focus on evidence retrieval (ER). It introduces CDER, a collaborative ER framework that builds a document-level bipartite graph with three sub-graphs and employs Entity Pair-aware GATs (EP-GATs) to model cross-pair collaboration and evidence dynamics. Key innovations include a dynamic Entity Pair sub-graph guided by relation similarity, EP-GAT layers for and , and a focal loss to handle severe evidence imbalance, which together yield strong ER performance and improve downstream DocRE. Experiments on DocRED show that CDER not only achieves state-of-the-art ER metrics but also provides substantial gains when plugged into existing DocRE systems, underscoring its practical impact for robust, scalable information extraction.

Abstract

Document-level Relation Extraction (DocRE) involves identifying relations between entities across multiple sentences in a document. Evidence sentences, crucial for precise entity pair relationships identification, enhance focus on essential text segments, improving DocRE performance. However, existing evidence retrieval systems often overlook the collaborative nature among semantically similar entity pairs in the same document, hindering the effectiveness of the evidence retrieval task. To address this, we propose a novel evidence retrieval framework, namely CDER. CDER employs an attentional graph-based architecture to capture collaborative patterns and incorporates a dynamic sub-structure for additional robustness in evidence retrieval. Experimental results on the benchmark DocRE dataset show that CDER not only excels in the evidence retrieval task but also enhances overall performance of existing DocRE system.

Paper Structure

This paper contains 31 sections, 12 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Evidence Retrieval in DocRE.
  4. Methodology
  5. Problem Formulation
  6. Encoder Module
  7. Document Encoding
  8. Entity Embedding
  9. Sentence Embedding
  10. Document-level Bipartite Graph
  11. Entity Pair Sub-graph ($G_{PP}$)
  12. Entity Pair-Sentence Sub-graph ($G_{PS}$)
  13. Sentence Sub-graph ($G_{SS}$)
  14. Entity Pair-aware GATs
  15. EP-GAT Motivation
  16. ...and 16 more sections

Figures (2)

  • Figure 1: An example document from the DocRED dataset yao2019docred. Entity pairs are given with corresponding supporting evidence sentences (i.e.,Evidence) and relations.
  • Figure 2: Overall framework of CDER: We obtain entity pairs and sentence embeddings and construct a bipartite graph from the document. Then, the graph is divided into Entity Pair ($G_{PP}$), Sentence-Entity Pair ($G_{PS}$), and Sentence ($G_{SS}$) sub-graphs. We enhance the embeddings of entity pairs and sentences using our proposed EP-GAT layers on $G_{PP}$ and $G_{PS}$, and GCNs kipf2016semi on $G_{SS}$. The final refined embeddings are employed for ER tasks.