Scholarlink

<p>This paper proposes CKHG, an innovative document-level relation extraction model that achieves explicit path reasoning by integrating fine-grained global heterogeneous graphs with external contextual knowledge. Specifically, we integrate Wikidata as an external knowledge source and construct a refined graph structure through representation fusion and global heterogeneous graph construction. This structure effectively captures multilevel interactions among word embeddings, mentions, entities, sentences, and documents. Furthermore, to effectively infer relational paths between entity pairs, we explicitly design three path reasoning tasks, combining entity context with document-level logical reasoning, significantly improving link prediction accuracy. Experimental results demonstrate that CKHG achieves superior performance on both DocRED and DWIE datasets, surpassing baseline models in F1 and Ign F1 scores. The enhanced global heterogeneous graph and external contextual knowledge effectively improve CKHG’s relation extraction capability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Document-level relation extraction model guided by global information and external contextual knowledge

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Hongli Yu,
  • Han Cao,
  • Yachao Cui,
  • Chenxi Dong

摘要

This paper proposes CKHG, an innovative document-level relation extraction model that achieves explicit path reasoning by integrating fine-grained global heterogeneous graphs with external contextual knowledge. Specifically, we integrate Wikidata as an external knowledge source and construct a refined graph structure through representation fusion and global heterogeneous graph construction. This structure effectively captures multilevel interactions among word embeddings, mentions, entities, sentences, and documents. Furthermore, to effectively infer relational paths between entity pairs, we explicitly design three path reasoning tasks, combining entity context with document-level logical reasoning, significantly improving link prediction accuracy. Experimental results demonstrate that CKHG achieves superior performance on both DocRED and DWIE datasets, surpassing baseline models in F1 and Ign F1 scores. The enhanced global heterogeneous graph and external contextual knowledge effectively improve CKHG’s relation extraction capability.