UniDRE: Universal method for document-level relation extraction with knowledge retrieval and reasoning augmentation on small and large language models

Document-level relation extraction (Doc-RE) seeks to identify relation triples among entity pairs across different sentences in a document. Unlike sentence-level RE, Doc-RE demands more comprehensive reasoning capabilities. Despite recent advances, most existing Doc-RE methods focus on isolated or partial reasoning abilities, leaving the comprehensive reasoning required for Doc-RE insufficiently explored. To solve these problems, the Uni versal knowledge retrieval augmentation methods for D oc- RE ( UniDRE ) were proposed with comprehensive reasoning abilities based on both S mall and L arge L anguage M odels, denoted as U n i D R E S L M and U n i D R E L L M , respectively. For SLM-based U n i D R E S L M , we augmented the comprehensive reasoning abilities by meticulously designing and jointly optimizing multi-stage reasoning models. Specifically, we constructed a document graph with the co-reference resolution model for semantic encoding and co-reference reasoning. Then, the document graph was further expanded into a document knowledge graph with the axial attention mechanism by retrieving the external knowledge for the logical and common-sense reasoning. For LLM-based U n i D R E L L M , we retrieved relevant relation triples from external knowledge bases to guide the LLM to alleviate the problems of hallucination generation and adaptation to small samples in vertical domains. Comprehensive experiments conducted on two benchmark datasets verified the effectiveness of our proposed method and also found that: (1) A finely designed SLM-based method for Doc-RE can achieve outstanding performance over the existing state-of-the-art baseline models. (2) The LLM-based Doc-RE method augmented with knowledge retrieval was more suitable for small sample scenarios.

Authors

Institutions

Publication Details

Journal
Journal of King Saud University - Computer and Information Sciences
Published
2026-11-19
DOI
https://doi.org/10.1007/s44443-026-01453-4
Primary Topic
Topic Modeling
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

UniDRE: Universal method for document-level relation extraction with knowledge retrieval and reasoning augmentation on small and large language models

Yuxiang Wang, Yihua Huang, 麦丞程, Hanxiang Wang et al.
Journal of King Saud University - Computer and Information Sciences
Topic Modeling
article

UniDRE: Universal method for document-level relation extraction with knowledge retrieval and reasoning augmentation on small and large language models

Yuxiang Wang, Yihua Huang, 麦丞程, Hanxiang Wang, Tongtong Xie, Genlin Ji
article en

Abstract

Document-level relation extraction (Doc-RE) seeks to identify relation triples among entity pairs across different sentences in a document. Unlike sentence-level RE, Doc-RE demands more comprehensive reasoning capabilities. Despite recent advances, most existing Doc-RE methods focus on isolated or partial reasoning abilities, leaving the comprehensive reasoning required for Doc-RE insufficiently explored. To solve these problems, the Uni versal knowledge retrieval augmentation methods for D oc- RE ( UniDRE ) were proposed with comprehensive reasoning abilities based on both S mall and L arge L anguage M odels, denoted as U n i D R E S L M and U n i D R E L L M , respectively. For SLM-based U n i D R E S L M , we augmented the comprehensive reasoning abilities by meticulously designing and jointly optimizing multi-stage reasoning models. Specifically, we constructed a document graph with the co-reference resolution model for semantic encoding and co-reference reasoning. Then, the document graph was further expanded into a document knowledge graph with the axial attention mechanism by retrieving the external knowledge for the logical and common-sense reasoning. For LLM-based U n i D R E L L M , we retrieved relevant relation triples from external knowledge bases to guide the LLM to alleviate the problems of hallucination generation and adaptation to small samples in vertical domains. Comprehensive experiments conducted on two benchmark datasets verified the effectiveness of our proposed method and also found that: (1) A finely designed SLM-based method for Doc-RE can achieve outstanding performance over the existing state-of-the-art baseline models. (2) The LLM-based Doc-RE method augmented with knowledge retrieval was more suitable for small sample scenarios.

Journal of King Saud University - Computer and Information SciencesVol. 38(10)
Nanjing Normal University (CN), Nanjing University (CN)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon