Scholarlink

<p>Recognizing Textual Entailment (RTE) is a fundamental task in natural language processing with broad downstream applications. In low-resource scenarios, existing methods struggle due to their reliance on external knowledge, sensitivity to semantic distribution bias, and limitations in capturing fine-grained entity distinctions. To address these challenges, we propose a novel framework that enhances semantic awareness of sequence entities through two key strategies: Sample-based Entity Replacement Augmentation and Integration of Entity Features. First, we extract and categorize entities from the dataset, then augment the data by replacing entities with others of the same type, leveraging semi-supervised learning to enhance model training. Second, we introduce the Advanced Contrastive Learning Encoder (ACE) model, which integrates filtered entity features into sequence representations and employs sequence-entity contrastive learning to improve entity distinction. Our approach significantly reduces dependence on external knowledge, mitigates semantic distribution bias, and enhances entity-awareness in RTE. Experimental results demonstrate that our methods consistently achieve state-of-the-art performance across all benchmark datasets for low-resource Chinese RTE 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

Entity injection with contrastive learning encoder for Chinese few-shot natural language inference

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Peichao Lai,
  • Feiyang Ye,
  • Yanggeng Fu,
  • Ruiqing Wang,
  • Yingjie Wu,
  • Yilei Wang

摘要

Recognizing Textual Entailment (RTE) is a fundamental task in natural language processing with broad downstream applications. In low-resource scenarios, existing methods struggle due to their reliance on external knowledge, sensitivity to semantic distribution bias, and limitations in capturing fine-grained entity distinctions. To address these challenges, we propose a novel framework that enhances semantic awareness of sequence entities through two key strategies: Sample-based Entity Replacement Augmentation and Integration of Entity Features. First, we extract and categorize entities from the dataset, then augment the data by replacing entities with others of the same type, leveraging semi-supervised learning to enhance model training. Second, we introduce the Advanced Contrastive Learning Encoder (ACE) model, which integrates filtered entity features into sequence representations and employs sequence-entity contrastive learning to improve entity distinction. Our approach significantly reduces dependence on external knowledge, mitigates semantic distribution bias, and enhances entity-awareness in RTE. Experimental results demonstrate that our methods consistently achieve state-of-the-art performance across all benchmark datasets for low-resource Chinese RTE tasks.