Scholarlink

<p>Currently, the mainstream methods for relational triple extraction mainly employ joint models and have achieved significant results. However, these methods still face challenges in dealing with the complexity of unstructured text. In joint models based on labeled relational triple extraction, the BiRTE model suffers from the issue of redundant entity pairs, leading to the generation of incorrect triples. To address this issue, This paper introduces a Ground Entity Extractor which is designed to aid the Dual Entity Linker (DEL), in addition, adversarial training methods are introduced during the DEL training process. Experimental results demonstrate that the DEL model extracts entity pairs from two directions, when compared to the previous BIRTE model, it generates more accurate entity pairs. On the WebNLG dataset, The accuracy increased by 2.9%, and F1 improved by 1.1%. On the NYT10 dataset, accuracy increased by 1.0%, recall increased by 0.5%, and F1 score improved by 0.7%. This brings significant enhancements to the triple entity extraction task, and in comparison with 8 baseline models across all datasets, the DEL model achieves better results.</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

DEL: a strategy for resolving redundancy in entity pairs within dual entity linker for relational triple extraction

  • 期刊论文

  • 发表日期：

    2025年1月25日

查看全文
  • Songhua Hu,
  • Hengxin Wang

摘要

Currently, the mainstream methods for relational triple extraction mainly employ joint models and have achieved significant results. However, these methods still face challenges in dealing with the complexity of unstructured text. In joint models based on labeled relational triple extraction, the BiRTE model suffers from the issue of redundant entity pairs, leading to the generation of incorrect triples. To address this issue, This paper introduces a Ground Entity Extractor which is designed to aid the Dual Entity Linker (DEL), in addition, adversarial training methods are introduced during the DEL training process. Experimental results demonstrate that the DEL model extracts entity pairs from two directions, when compared to the previous BIRTE model, it generates more accurate entity pairs. On the WebNLG dataset, The accuracy increased by 2.9%, and F1 improved by 1.1%. On the NYT10 dataset, accuracy increased by 1.0%, recall increased by 0.5%, and F1 score improved by 0.7%. This brings significant enhancements to the triple entity extraction task, and in comparison with 8 baseline models across all datasets, the DEL model achieves better results.