Scholarlink

<p>Relation extraction (RE) is crucial for advancing information extraction, with overlapping RE being a significant sub-task. This scenario is generally more complex than normal RE. While normal RE involves two entities, overlapping RE involves two or three distinct entities, each with multiple mentions. This increases the relation ambiguity, particularly for single entity overlapping (SEO). To address this issue, we propose a multi-task learning approach, imbibing a syntactic perception task into the RE with two supervisory signals: mention-cross matrix and syntactic binary signal. The former supervises this model to predict an asymmetric product matrix, emphasizing different mentions interactions. The latter utilizes the Steiner minimal tree to identify token-level semantic paths across all mentions in syntactic trees and guides the attention to capture semantic dependencies. These signals collectively optimize the extraction process. The results on four datasets confirm our approach achieves competitive performance, demonstrating the effectiveness of disambiguating SEO scenarios in RE.</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

Harnessing multi-task learning to improve overlapping relation extraction

  • 期刊论文

  • 发表日期：

    2025年6月20日

查看全文
  • Hailin Wang,
  • Song Zhang,
  • Dan Zhang,
  • Ao Ma

摘要

Relation extraction (RE) is crucial for advancing information extraction, with overlapping RE being a significant sub-task. This scenario is generally more complex than normal RE. While normal RE involves two entities, overlapping RE involves two or three distinct entities, each with multiple mentions. This increases the relation ambiguity, particularly for single entity overlapping (SEO). To address this issue, we propose a multi-task learning approach, imbibing a syntactic perception task into the RE with two supervisory signals: mention-cross matrix and syntactic binary signal. The former supervises this model to predict an asymmetric product matrix, emphasizing different mentions interactions. The latter utilizes the Steiner minimal tree to identify token-level semantic paths across all mentions in syntactic trees and guides the attention to capture semantic dependencies. These signals collectively optimize the extraction process. The results on four datasets confirm our approach achieves competitive performance, demonstrating the effectiveness of disambiguating SEO scenarios in RE.