Scholarlink

<p>Knowledge entities are the minimal independent units that summarize crucial concepts and essential elements in scientific literature. This paper aims to automatically extract scientific knowledge entities that researchers are concerned about from Chinese scientific literature in the general domain. However, there are two main challenges: the lack of high-quality general-domain annotated corpora and inevitable annotation errors during manual labeling; and the absence of explicit entity length modeling in most NER models, often leading to overlong or implausible span predictions, especially in nested or structurally complex texts. To address these issues, we define nine fine-grained categories of scientific knowledge entities and construct a high-quality annotated corpus of 1100 Chinese scientific abstracts across eight major disciplines, totaling 22,184 labeled entities. We propose a novel span-based model, StructBERT-AT-DMGP, which integrates adversarial training (AT) and a Dual-Masked Global Pointer (DMGP). In the model, AT serves as an implicit regularization method to improve robustness against annotation noise, while DMGP incorporates length constraints into span prediction, enhancing boundary recognition in complex scenarios. Experimental results show that StructBERT-AT-DMGP achieves a macro-average F1-score of 70.23% on our dataset, outperforming several mainstream open-source baselines. Ablation studies validate the effectiveness of AT and DMGP. Furthermore, evaluations on three public Chinese NER benchmarks–MSRA, Resume, and CLUENER–yield F1-scores of 96.25%, 96.62%, and 83.03%, respectively, demonstrating strong generalizability across domains.</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

Named entity recognition for chinese scientific literature based on adversarial training and dual-masked global pointer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月10日

查看全文
  • Mengting Zhang,
  • Yang Li,
  • Zhixiong Zhang,
  • Yajiao Wang,
  • Qianqian Yu,
  • Liangping Ding

摘要

Knowledge entities are the minimal independent units that summarize crucial concepts and essential elements in scientific literature. This paper aims to automatically extract scientific knowledge entities that researchers are concerned about from Chinese scientific literature in the general domain. However, there are two main challenges: the lack of high-quality general-domain annotated corpora and inevitable annotation errors during manual labeling; and the absence of explicit entity length modeling in most NER models, often leading to overlong or implausible span predictions, especially in nested or structurally complex texts. To address these issues, we define nine fine-grained categories of scientific knowledge entities and construct a high-quality annotated corpus of 1100 Chinese scientific abstracts across eight major disciplines, totaling 22,184 labeled entities. We propose a novel span-based model, StructBERT-AT-DMGP, which integrates adversarial training (AT) and a Dual-Masked Global Pointer (DMGP). In the model, AT serves as an implicit regularization method to improve robustness against annotation noise, while DMGP incorporates length constraints into span prediction, enhancing boundary recognition in complex scenarios. Experimental results show that StructBERT-AT-DMGP achieves a macro-average F1-score of 70.23% on our dataset, outperforming several mainstream open-source baselines. Ablation studies validate the effectiveness of AT and DMGP. Furthermore, evaluations on three public Chinese NER benchmarks–MSRA, Resume, and CLUENER–yield F1-scores of 96.25%, 96.62%, and 83.03%, respectively, demonstrating strong generalizability across domains.