Scholarlink

<p>In addition to flat and nested entities, there are discontinuous entities in biomedical Named Entity Recognition (NER) datasets. This type of entity consists of discontinuous spans, and there is also nesting between different entities, making it more challenging to identify entities with this structure than flat and nested entities. The existing span-based discontinuous NER methods still suffer from severe class imbalance and high computational costs, which affect the model’s overall performance. We propose a candidate span classification based on intra-entity token prediction enhancement and an entity span relationship prediction model based on contrastive learning to address the above issues. To reduce the number of negative samples without losing long entities, we use sequence labeling to predict intra-entity token labeling and generate candidate spans based on the intra-entity token labeling. In addition to improving the recognition performance of discontinuous entities, we also introduce contrastive learning in the entity span relationship prediction module to improve the similarity between entity spans with Connection relationships. Multi-task learning is used to capture the dependencies of entity content token prediction, candidate span classification, and entity span relationship prediction to improve the recognition performance of named entities. To enhance word representations with syntactic information, this paper employs Attention Graph Convolutional Networks (A-GCN) to incorporate dependency structures and their types, thereby improving model performance. The experimental results show that in most cases, our model outperforms the baseline models, with F1-values of 74.73%, 84.55%, and 83.68% on the CADEC, ShARe13, and ShARe14 datasets.</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

A discontinuous NER model based on token prediction and contrastive learning to enhance span

  • 期刊论文

  • 发表日期：

    2025年6月3日

查看全文
  • Yaodi Liu,
  • Kun Zhang,
  • Dianying Chen,
  • Chenxi Cai,
  • Xiaohe Wu,
  • Rong Tong

摘要

In addition to flat and nested entities, there are discontinuous entities in biomedical Named Entity Recognition (NER) datasets. This type of entity consists of discontinuous spans, and there is also nesting between different entities, making it more challenging to identify entities with this structure than flat and nested entities. The existing span-based discontinuous NER methods still suffer from severe class imbalance and high computational costs, which affect the model’s overall performance. We propose a candidate span classification based on intra-entity token prediction enhancement and an entity span relationship prediction model based on contrastive learning to address the above issues. To reduce the number of negative samples without losing long entities, we use sequence labeling to predict intra-entity token labeling and generate candidate spans based on the intra-entity token labeling. In addition to improving the recognition performance of discontinuous entities, we also introduce contrastive learning in the entity span relationship prediction module to improve the similarity between entity spans with Connection relationships. Multi-task learning is used to capture the dependencies of entity content token prediction, candidate span classification, and entity span relationship prediction to improve the recognition performance of named entities. To enhance word representations with syntactic information, this paper employs Attention Graph Convolutional Networks (A-GCN) to incorporate dependency structures and their types, thereby improving model performance. The experimental results show that in most cases, our model outperforms the baseline models, with F1-values of 74.73%, 84.55%, and 83.68% on the CADEC, ShARe13, and ShARe14 datasets.