Scholarlink

Currently, there is limited research on ancient Chinese named entity recognition, primarily due to the scarcity of publicly available datasets for model training. We constructed a CMAG-NER dataset based on the “Comprehensive Mirror for Aid in Government”. Addressing the challenges faced by existing models in identifying person entities with omitted surnames and determining entity boundaries in ancient Chinese texts, we integrated the LEBERT-CRF model with a general domain lexicon and the “Comprehensive Mirror for Aid in Government Dictionary” to fuse external statistical information and rule-based knowledge, thereby enhancing the performance of ancient Chinese named entity recognition. Additionally, to improve the model’s comprehension of ancient Chinese, we have compiled a substantial corpus of ancient Chinese literature for incremental pre-training of BERT-Ancient-Chinese. Experimental results demonstrate that the proposed method effectively mitigates the aforementioned challenges, but the performance of large models on this task still requires improvement.

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

Research on Named Entity Recognition in Ancient Chinese Based on Incremental Pre-training and Domain Lexicon

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Wenjun Kang,
  • Jiali Zuo,
  • Qili Dai,
  • Yiyu Hu,
  • Mingwen Wang

摘要

Currently, there is limited research on ancient Chinese named entity recognition, primarily due to the scarcity of publicly available datasets for model training. We constructed a CMAG-NER dataset based on the “Comprehensive Mirror for Aid in Government”. Addressing the challenges faced by existing models in identifying person entities with omitted surnames and determining entity boundaries in ancient Chinese texts, we integrated the LEBERT-CRF model with a general domain lexicon and the “Comprehensive Mirror for Aid in Government Dictionary” to fuse external statistical information and rule-based knowledge, thereby enhancing the performance of ancient Chinese named entity recognition. Additionally, to improve the model’s comprehension of ancient Chinese, we have compiled a substantial corpus of ancient Chinese literature for incremental pre-training of BERT-Ancient-Chinese. Experimental results demonstrate that the proposed method effectively mitigates the aforementioned challenges, but the performance of large models on this task still requires improvement.