Scholarlink

<p>Relation extraction (RE) plays a vital role in transforming unstructured text into structured knowledge. Person-centric relations are prevalent in large-scale user-generated content and real-world applications. Extracting such relations requires fine-grained semantic modeling and high computational efficiency. Despite their importance, person-centric relations remain underexplored in current RE research. Existing models often miss implicit semantic cues linked to person entities, leading to poor performance. To address this, we propose SELF, a multi-task learning framework, which enhances extraction performance for person-centric relationships. SELF introduces a simple, parameter-free auxiliary task that models the semantic space of person-related entities. This enables effective capture of sparse, crucial semantic information in person-centric contexts. Additionally, SELF incorporates a parallel hierarchical feature retainer module. This module adaptively maintains shallow and deep semantic representations from the auxiliary task, enriching the understanding of person-centric relationships. Extensive experiments on TACRED, Re-TACRED, and SemEval-2010 Task-8 show that SELF outperforms existing models in classifying person-centric relations and maintains computational efficiency in multi-task settings.</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

SELF: a multi-task learning framework for robust person-centric relation extraction

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Hailin Wang,
  • Wentong Niu,
  • Hangyi Ren,
  • Jiahao Li,
  • Jingxuan Tian,
  • Dan Zhang

摘要

Relation extraction (RE) plays a vital role in transforming unstructured text into structured knowledge. Person-centric relations are prevalent in large-scale user-generated content and real-world applications. Extracting such relations requires fine-grained semantic modeling and high computational efficiency. Despite their importance, person-centric relations remain underexplored in current RE research. Existing models often miss implicit semantic cues linked to person entities, leading to poor performance. To address this, we propose SELF, a multi-task learning framework, which enhances extraction performance for person-centric relationships. SELF introduces a simple, parameter-free auxiliary task that models the semantic space of person-related entities. This enables effective capture of sparse, crucial semantic information in person-centric contexts. Additionally, SELF incorporates a parallel hierarchical feature retainer module. This module adaptively maintains shallow and deep semantic representations from the auxiliary task, enriching the understanding of person-centric relationships. Extensive experiments on TACRED, Re-TACRED, and SemEval-2010 Task-8 show that SELF outperforms existing models in classifying person-centric relations and maintains computational efficiency in multi-task settings.