Skip to main navigation Skip to search Skip to main content

TriFusion: A Triple-View Fusion Framework for Document-Level Relation Extraction

  • Beihang University

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Document-level Relation Extraction (DocRE) aims to identify relationship labels between multiple entities across an entire document, differing from traditional sentence-level extraction by requiring reasoning across multiple sentences. Existing methods rely on logical inference and entity-context relationships but often underutilize rich linguistic information in documents. This paper proposes a triple-view fusion framework (Dependency syntax relation context fusion, DSRC) to enhance syntax aggregation by integrating dependency trees. It leverages graph-structured syntax and dependency paths while addressing long-tail and multi-label challenges through relation relevance and entity completion. Evaluated on two benchmark datasets (DocRED, Re-DocRED), our method significantly outperforms state-of-the-art models.

Original languageEnglish
Title of host publicationProceedings - 2025 IEEE Smart World Congress, SWC 2025, 2025 IEEE Ubiquitous Intelligence and Computing, Autonomous and Trusted Computing, Digital Twin, Metaverse, Scalable Computing and Communications
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1238-1243
Number of pages6
ISBN (Electronic)9798331575984
DOIs
StatePublished - 2025
Event2025 IEEE Smart World Congress, SWC 2025 - Calgary, Canada
Duration: 18 Aug 202522 Aug 2025

Publication series

NameProceedings - 2025 IEEE Smart World Congress, SWC 2025, 2025 IEEE Ubiquitous Intelligence and Computing, Autonomous and Trusted Computing, Digital Twin, Metaverse, Scalable Computing and Communications

Conference

Conference2025 IEEE Smart World Congress, SWC 2025
Country/TerritoryCanada
CityCalgary
Period18/08/2522/08/25

Keywords

  • Dependency Graph
  • Dynamic Gating
  • Entity Completion
  • Information Extraction

Fingerprint

Dive into the research topics of 'TriFusion: A Triple-View Fusion Framework for Document-Level Relation Extraction'. Together they form a unique fingerprint.

Cite this