Scholarlink

<p>Entity Relation Extraction (ERE) aims to identify semantic relations between entities from unstructured texts. Despite achieving promising results, existing methods face significant efficiency challenges due to the overwhelming presence of irrelevant candidate entity pairs–with over 96% of candidate pairs lacking meaningful relations in typical datasets. This computational burden severely limits the practical applicability of ERE systems. To address these limitations, we propose FastERE, a fast framework that introduces three key innovations: (1) a pruner-based Entity Pairs Selection Module that dynamically filters irrelevant entity pairs before relation extraction, reducing computational overhead while improving sample quality; (2) an efficient task reformulation that frames NER as sequence tagging with adjacent attention mechanisms and RE as grouped triplet prediction with masked parallel packing; and (3) a multi-task learning framework with prefix-enhanced shared encoders that enables effective feature interaction while maintaining task independence. Experimental results on benchmark datasets demonstrate that FastERE achieves 6-20<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> speedup over state-of-the-art methods while maintaining competitive accuracy, with <i>F</i>1 scores of 69.9% and 67.0% on ACE05 and ACE04 respectively. Our code is available at <a href="https://github.com/xerrors/FastERE">https://github.com/xerrors/FastERE</a>.</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

Fastere: a fast framework for entity relation extractions

  • 期刊论文

  • 发表日期：

    2025年8月25日

查看全文
  • Wenjie Zhang,
  • Tianyang Xu,
  • Yang Hua,
  • Zhenhua Feng,
  • Xiaoning Song

摘要

Entity Relation Extraction (ERE) aims to identify semantic relations between entities from unstructured texts. Despite achieving promising results, existing methods face significant efficiency challenges due to the overwhelming presence of irrelevant candidate entity pairs–with over 96% of candidate pairs lacking meaningful relations in typical datasets. This computational burden severely limits the practical applicability of ERE systems. To address these limitations, we propose FastERE, a fast framework that introduces three key innovations: (1) a pruner-based Entity Pairs Selection Module that dynamically filters irrelevant entity pairs before relation extraction, reducing computational overhead while improving sample quality; (2) an efficient task reformulation that frames NER as sequence tagging with adjacent attention mechanisms and RE as grouped triplet prediction with masked parallel packing; and (3) a multi-task learning framework with prefix-enhanced shared encoders that enables effective feature interaction while maintaining task independence. Experimental results on benchmark datasets demonstrate that FastERE achieves 6-20 \(\times \) speedup over state-of-the-art methods while maintaining competitive accuracy, with F1 scores of 69.9% and 67.0% on ACE05 and ACE04 respectively. Our code is available at https://github.com/xerrors/FastERE.