Scholarlink

<p>The lack of extensive annotated corpora, domainspecific language, and complex entity boundaries make Biomedical Named Entity Recognition (BioNER) a difficult undertaking. This study introduces E6, a unique method that combines structured prediction utilizing Conditional Random Fields (CRF) with BIOES labeling, domain-adapted language models (PubMedBERT), and a new span-level contrastive loss that specifically addresses boundary detection problems. E6 outperforms the previous state-of-the-art by 2.56 points, achieving 91.56% F1 on the NCBI Disease corpus. This study contributions include: (1) an empirical analysis of loss weighting strategies for CRF-based models; (2) a boundary-aware training objective that produces hard negatives through systematic span perturbations; and (3) a fully reproducible implementation that is ready for production. The suggested model is especially successful in minimizing ± 1 token border errors by 70%, proving that explicit boundary modeling is an important but little-studied portion of biological NER pipelines.</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

The E6 model: span-level contrastive learning for enhancing biomedical named entity recognition

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Nandita Goyal,
  • Navdeep Singh

摘要

The lack of extensive annotated corpora, domainspecific language, and complex entity boundaries make Biomedical Named Entity Recognition (BioNER) a difficult undertaking. This study introduces E6, a unique method that combines structured prediction utilizing Conditional Random Fields (CRF) with BIOES labeling, domain-adapted language models (PubMedBERT), and a new span-level contrastive loss that specifically addresses boundary detection problems. E6 outperforms the previous state-of-the-art by 2.56 points, achieving 91.56% F1 on the NCBI Disease corpus. This study contributions include: (1) an empirical analysis of loss weighting strategies for CRF-based models; (2) a boundary-aware training objective that produces hard negatives through systematic span perturbations; and (3) a fully reproducible implementation that is ready for production. The suggested model is especially successful in minimizing ± 1 token border errors by 70%, proving that explicit boundary modeling is an important but little-studied portion of biological NER pipelines.