Scholarlink

<p>Continual few-shot relation extraction aims to incrementally learn new relations from limited samples while avoiding forgetting old ones. However, current methods fail to fully leverage entity information, making it difficult for the model to capture the correlation between relations and entities. This correlation becomes increasingly chaotic during continual learning process, leading to severe catastrophic forgetting. Furthermore, the challenge of overfitting caused by limited data has long hindered this task. In this paper, we propose a <b>D</b>ynamic <b>E</b>ntity-<b>R</b>elation Interaction (DERI) module aimed at modeling the entity-relation correlation. DERI performs adaptive fusion on cross-attention of relations and entities to capture correlation of both. In this process, the fusion weights are dynamically adjusted to obtain stable entity-relation correlation and adapt to the emergence of new relations during continual learning procedure, which alleviates catastrophic forgetting. To address the issue of overfitting, we propose a Multi-<b>P</b>rompts <b>E</b>nhancement strategy that combines prompts containing different levels of information, helping the model capture relation feature from multiple dimensions. Experimental results on two public datasets show that our method outperforms other competitive baselines.</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

PE-DER: a dynamic entity-relation interaction method based on multi-prompts enhancement for continual few-shot relation extraction

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Weibo Chen,
  • Yanyan Feng,
  • Zehong Lin,
  • Fenghuan Li,
  • Yun Xue,
  • Chengbo He

摘要

Continual few-shot relation extraction aims to incrementally learn new relations from limited samples while avoiding forgetting old ones. However, current methods fail to fully leverage entity information, making it difficult for the model to capture the correlation between relations and entities. This correlation becomes increasingly chaotic during continual learning process, leading to severe catastrophic forgetting. Furthermore, the challenge of overfitting caused by limited data has long hindered this task. In this paper, we propose a Dynamic Entity-Relation Interaction (DERI) module aimed at modeling the entity-relation correlation. DERI performs adaptive fusion on cross-attention of relations and entities to capture correlation of both. In this process, the fusion weights are dynamically adjusted to obtain stable entity-relation correlation and adapt to the emergence of new relations during continual learning procedure, which alleviates catastrophic forgetting. To address the issue of overfitting, we propose a Multi-Prompts Enhancement strategy that combines prompts containing different levels of information, helping the model capture relation feature from multiple dimensions. Experimental results on two public datasets show that our method outperforms other competitive baselines.