摘要
持续关系抽取模型使用包含新关系的数据集进行重新训练时,模型参数的相关先验关系类别会被重新加权,导致灾难性遗忘问题,尤其在相似关系的决策边界附近。因此,本文提出一种基于边缘对比学习的持续关系抽取模型MCL-CRE。该模型通过边缘对比学习提升模型区分现有关系的能力,并扩大决策边界附近相似关系表示之间的距离。实验结果表明,该模型可稳定区分相似关系,且在不平衡的TACRED数据集上取得了显著效果。
The Continual Relation Extraction model is re-trained with datasets containing new relations,it easily leads to the catastrophic forgetting problem caused by re-weighting the model parameters of relevant prior relation types,especially near the decision borderline of analogous relation pairs.Therefore,we propose the Marginal Contrastive Learning-based Continual Relation Extraction model(MCL-CRE).Our model adopts marginal contrastive learning to enhance the ability to distinguish existing relations from new relations and enlarges the distance between analogous relation representations near the decision borderline.Experimental results show that our model can stably distinguish analogous relations,and achieve significant results on the unbalanced TACRED dataset.
作者
左洋
王聪聪
葛宝泉
Zuo Yang;Wang Congcong;Ge Baoquan(School of Artificial Intelligence,Xinjiang Vocational University of Technology,Kashgar,Xinjiang 844210,China)
出处
《计算机时代》
2025年第11期13-18,共6页
Computer Era
关键词
持续关系抽取
灾难性遗忘
相似关系
边缘对比学习
Continual Relation Extraction
Catastrophic Forgetting
Analogous Relations
Marginal Contrastive Learning
