Scholarlink

<p>Continual few-shot relation extraction aims to enable models to incrementally learn new relation types from limited labeled instances while retaining knowledge of previously learned relations. This task presents two critical challenges: catastrophic forgetting caused by the continual learning process, and overfitting due to the scarcity of supervision in few-shot scenarios. Moreover, many existing approaches treat each task independently, overlooking the underlying relationships among tasks and thereby limiting effective knowledge transfer. To overcome these difficulties, we propose an innovative model, <b>ECKT</b> (<b>E</b>nhancing <b>C</b>ross-Task <b>K</b>nowledge <b>T</b>ransfer). The model consists of three key modules: (1) A memory strategy to preserve critical information across tasks, (2) A prompt representation module to enhance task-specific features, (3) A cross-task cohesion module that minimizes distribution overlap while strengthening cross-task connections. Specifically, the cross-task cohesion module encourages the model to distinguish between various sample distributions, thereby alleviating catastrophic forgetting and mitigating overfitting. At the same time, it strengthens task-to-task relationships, improving knowledge transfer. We do thorough analyses on two popular benchmark datasets, showing that <b>ECKT</b> outperforms most existing CFRE methods, thus verifying the efficacy of our model. Our code is publicly available at <a href="https://github.com/moonfelix427/ECKT-main">https://github.com/moonfelix427/ECKT-main</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

ECKT: enhancing cross-task knowledge transfer in continual few-shot relation extraction

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Yifan Yao,
  • Juan Zhang,
  • Zhijun Fang,
  • Yongbin Gao

摘要

Continual few-shot relation extraction aims to enable models to incrementally learn new relation types from limited labeled instances while retaining knowledge of previously learned relations. This task presents two critical challenges: catastrophic forgetting caused by the continual learning process, and overfitting due to the scarcity of supervision in few-shot scenarios. Moreover, many existing approaches treat each task independently, overlooking the underlying relationships among tasks and thereby limiting effective knowledge transfer. To overcome these difficulties, we propose an innovative model, ECKT (Enhancing Cross-Task Knowledge Transfer). The model consists of three key modules: (1) A memory strategy to preserve critical information across tasks, (2) A prompt representation module to enhance task-specific features, (3) A cross-task cohesion module that minimizes distribution overlap while strengthening cross-task connections. Specifically, the cross-task cohesion module encourages the model to distinguish between various sample distributions, thereby alleviating catastrophic forgetting and mitigating overfitting. At the same time, it strengthens task-to-task relationships, improving knowledge transfer. We do thorough analyses on two popular benchmark datasets, showing that ECKT outperforms most existing CFRE methods, thus verifying the efficacy of our model. Our code is publicly available at https://github.com/moonfelix427/ECKT-main.