基于实体关系双向引导的中文少样本关系三元组抽取

Bidirectional Entity-Relation Guided Chinese Few-Shot Relation Triple Extraction

摘要: 关系三元组抽取是知识图谱构建中的关键任务。然而，实际应用中关系类别普遍呈现长尾分布，许多低频关系缺乏充足的标注样本，严重制约了监督学习方法的性能。为此，我们提出了一种基于实体关系双向引导的中文少样本关系三元组抽取方法，通过构建实体识别与关系分类之间的双向信息流，增强子任务间的信息交互与协同建模能力。为提升模型的表示能力和分类判别性能，我们设计了基于token置信度的加权原型构建方法，以动态评估各token的语义贡献并优化原型表示。同时，我们引入原型多样性约束损失，在提升类内一致性的基础上增强类间可分性，从而强化类别判别效果。实验结果表明，所提的方法在多个评价指标上显著优于主流基线模型，验证了其在少样本场景下的有效性与泛化能力。

Abstract: Relation triple extraction is a critical task in the construction of knowledge graphs. However, in real-world applications, the distribution of relation types often follows a long-tail pattern, where many low-frequency relations lack sufficient annotated examples, severely limiting the performance of supervised learning methods. To address this challenge, we propose a Chinese few-shot relation triple extraction method based on a bidirectional entity-relation guidance mechanism. By constructing a bidirectional information flow between entity recognition and relation classification, our method enhances the interaction and joint modeling between the two subtasks. To improve the model’s representation capability and classification performance, we further design a token-confidence-based weighted prototype construction method, which dynamically assesses the semantic contribution of each token to optimize prototype representations. In addition, we introduce a prototype diversity constraint loss to improve intra-class consistency while enhancing inter-class separability, thereby strengthening the model’s discriminative ability. Experimental results show that our method significantly outperforms mainstream baseline models across multiple evaluation metrics, demonstrating its effectiveness and generalization ability in few-shot scenarios.