Scholarlink

<p>Joint entity and relation extraction aims to simultaneously identify entity boundaries and their semantic relations, providing a reliable foundation for building high-quality knowledge graphs and enhancing downstream reasoning and applications. However, existing methods still suffer from redundant representations, weak interaction modeling between entities and relations, and unstable decision boundaries in complex sentences. To address these issues, we propose LATTE, a novel joint extraction model integrating three key components: a low-rank attention mechanism that compresses irrelevant features while retaining essential semantics; a multi-head relation-aware interaction that aligns relation prototypes to capture overlapping triples; and a residual task-aware bottleneck that performs differentiated compression to stabilize decision boundaries. Experiments on NYT and WebNLG show that LATTE achieves F1 scores of 93.8% and 95.1%, respectively, outperforming mainstream baselines and demonstrating superior robustness and effectiveness for joint extraction tasks.</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

LATTE: a joint entity and relation extraction model based on low-rank attention mechanism and residual task-aware bottleneck

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Xiaoqin Zhang,
  • Shuhan Nie,
  • Yanjun Lu

摘要

Joint entity and relation extraction aims to simultaneously identify entity boundaries and their semantic relations, providing a reliable foundation for building high-quality knowledge graphs and enhancing downstream reasoning and applications. However, existing methods still suffer from redundant representations, weak interaction modeling between entities and relations, and unstable decision boundaries in complex sentences. To address these issues, we propose LATTE, a novel joint extraction model integrating three key components: a low-rank attention mechanism that compresses irrelevant features while retaining essential semantics; a multi-head relation-aware interaction that aligns relation prototypes to capture overlapping triples; and a residual task-aware bottleneck that performs differentiated compression to stabilize decision boundaries. Experiments on NYT and WebNLG show that LATTE achieves F1 scores of 93.8% and 95.1%, respectively, outperforming mainstream baselines and demonstrating superior robustness and effectiveness for joint extraction tasks.