Scholarlink

<p>Relation extraction aims to identify semantic relations between entities in unstructured text. However, existing methods still face challenges such as relation redundancy, overlapping relations, and error propagation in pipeline frameworks. To address these issues, this paper proposes a joint extraction framework that combines multi-branch dilated convolutions with a multi-head cross-attention mechanism. The method uses BERT and PERCNet to obtain contextual and enhanced semantic features, respectively, and employs multi-head cross-attention together with a learnable gated fusion mechanism for feature interaction and adaptive fusion, thereby improving relation representation. Subsequently, the model filters redundant relations through latent relation prediction, performs joint extraction of entities and relations, and further refines triplets with a global correspondence mechanism. Experimental results show that the proposed method achieves F1 scores of 93.5%, 94.5%, and 75.0% on the NYT, WebNLG, and DuIE datasets, respectively, demonstrating its effectiveness.</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

Entity-relation joint extraction model based on multi-branch dilated convolutions and multi-head cross-attention mechanism

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • LiYun Kang,
  • HuiJun Zhang,
  • YanJun Lu

摘要

Relation extraction aims to identify semantic relations between entities in unstructured text. However, existing methods still face challenges such as relation redundancy, overlapping relations, and error propagation in pipeline frameworks. To address these issues, this paper proposes a joint extraction framework that combines multi-branch dilated convolutions with a multi-head cross-attention mechanism. The method uses BERT and PERCNet to obtain contextual and enhanced semantic features, respectively, and employs multi-head cross-attention together with a learnable gated fusion mechanism for feature interaction and adaptive fusion, thereby improving relation representation. Subsequently, the model filters redundant relations through latent relation prediction, performs joint extraction of entities and relations, and further refines triplets with a global correspondence mechanism. Experimental results show that the proposed method achieves F1 scores of 93.5%, 94.5%, and 75.0% on the NYT, WebNLG, and DuIE datasets, respectively, demonstrating its effectiveness.