Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FCDS: Fusing Constituency and Dependency Syntax into Document-Level Relation Extraction

Xudong Zhu, Zhao Kang, Bei Hui

TL;DR

This paper tackles document-level relation extraction by integrating two complementary syntactic views: constituency syntax for sentence-level aggregation and dependency syntax for cross-sentence relational reasoning. The proposed FCDS architecture fuses a constituency-tree representation (via Tree-LSTM) with a dependency-graph backbone (via GCN), augmented by a document node to bridge sentences and a dynamic fusion mechanism that balances both signals. Empirical results on DocRED and biomedical datasets show state-of-the-art performance and robust improvements over baselines, with ablations confirming the contribution of both syntactic components. The approach demonstrates that multi-granularity syntax can significantly enhance cross-sentence reasoning in DocRE and suggests further gains from smarter sentence selection guided by constituency structures.

Abstract

Document-level Relation Extraction (DocRE) aims to identify relation labels between entities within a single document. It requires handling several sentences and reasoning over them. State-of-the-art DocRE methods use a graph structure to connect entities across the document to capture dependency syntax information. However, this is insufficient to fully exploit the rich syntax information in the document. In this work, we propose to fuse constituency and dependency syntax into DocRE. It uses constituency syntax to aggregate the whole sentence information and select the instructive sentences for the pairs of targets. It exploits the dependency syntax in a graph structure with constituency syntax enhancement and chooses the path between entity pairs based on the dependency graph. The experimental results on datasets from various domains demonstrate the effectiveness of the proposed method. The code is publicly available at this url.

FCDS: Fusing Constituency and Dependency Syntax into Document-Level Relation Extraction

TL;DR

This paper tackles document-level relation extraction by integrating two complementary syntactic views: constituency syntax for sentence-level aggregation and dependency syntax for cross-sentence relational reasoning. The proposed FCDS architecture fuses a constituency-tree representation (via Tree-LSTM) with a dependency-graph backbone (via GCN), augmented by a document node to bridge sentences and a dynamic fusion mechanism that balances both signals. Empirical results on DocRED and biomedical datasets show state-of-the-art performance and robust improvements over baselines, with ablations confirming the contribution of both syntactic components. The approach demonstrates that multi-granularity syntax can significantly enhance cross-sentence reasoning in DocRE and suggests further gains from smarter sentence selection guided by constituency structures.

Abstract

Document-level Relation Extraction (DocRE) aims to identify relation labels between entities within a single document. It requires handling several sentences and reasoning over them. State-of-the-art DocRE methods use a graph structure to connect entities across the document to capture dependency syntax information. However, this is insufficient to fully exploit the rich syntax information in the document. In this work, we propose to fuse constituency and dependency syntax into DocRE. It uses constituency syntax to aggregate the whole sentence information and select the instructive sentences for the pairs of targets. It exploits the dependency syntax in a graph structure with constituency syntax enhancement and chooses the path between entity pairs based on the dependency graph. The experimental results on datasets from various domains demonstrate the effectiveness of the proposed method. The code is publicly available at this url.
Paper Structure (19 sections, 21 equations, 5 figures, 6 tables)

This paper contains 19 sections, 21 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Relation Works
  3. Document-level Relation Extraction
  4. Constituency and Dependency Syntax
  5. Methodology
  6. Text Encoding
  7. Constituency Tree Construction
  8. Dependency Graph Construction
  9. Dynamic Fusion and Classification
  10. Experiments
  11. Datasets
  12. Implementation details
  13. Results on DocRED
  14. Results on Biomedical Datasets
  15. Ablation Study
  16. ...and 4 more sections

Figures (5)

  • Figure 1: A sentence-level and a document-level relation instance from DocRED. Entity pairs are colored differently according to relation. To identify irrelevant or even noisy relations, unrelated entities are uniformly labeled with one single color.
  • Figure 2: Syntactic parsing results of evidence sentence "Louis Chollet ...” mentioned in the previous relation instance. (a) and (b) represent the corresponding dependency and constituency tree, respectively.
  • Figure 3: The overview of our architecture. Note that we use the result of constituency syntax to enhance the dependency graph and obtain relations between entity pairs with dynamic weighted fusion.
  • Figure 4: The learning curve of $\eta$ on DocRED, CDR and GDA datasets.
  • Figure 5: Two different cases in DocRED.