大语言模型下古籍语间跨语言时间表达式抽取研究

• 作者： 贾智 ¹ 赵志枭 ² 沈思 ¹
• 作者单位：
  1.  南京理工大学-经济管理学院

  2.  南京农业大学-信息管理学院
• 发布时间：2025-02-12

摘要: [目的/意义]古籍中的时间表达对于语义理解具有重要意义，由于大模型在各种自然语言处理任务上表现出色，因此本文探究了大模型在古籍时间表达式抽取任务上的性能。 [方法/过程]对先秦古籍语料分别进行处理，作为SikuBERT-BiLSTM-CRF进行训练和对Baichuan2-13B-Base、Baichuan2-7B-Base、Xunzi-Baichuan2-7B进行指令微调的数据集，并使用准确率、召回率、F1值作为指标验证模型的性能，最后在《汉书》等古籍上验证模型的泛化能力。[结果/结论]实验结果表明，在古籍时间表达式抽取任务中，总体来看SikuBERT-BiLSTM-CRF模型表现最优， Baichuan2-13B-Base等大模型也都具备较好的能力和不错的泛化能力，展现出大模型在该任务上的潜力。

Abstract: [Purpose/significance] The temporal expressions in ancient texts hold significant importance for semantic understanding. Given the outstanding performance of large models in various natural language processing tasks, this paper explores the performance of large models in the task of extracting temporal expressions from ancient texts. [Method/process] The pre-Qin ancient texts were processed separately to serve as datasets for training SikuBERT-BiLSTM-CRF and for fine-tuning Baichuan2-13B-Base, Baichuan2-7B-Base, and Xunzi-Baichuan2-7B with instruction tuning. The performance of the models was validated using precision, recall, and F1 score as metrics, and the generalization capability of the models was further verified on ancient texts such as the "Book of Han." [Result/conclusion] The experimental findings reveal that, in the task of extracting temporal expressions from ancient texts, the SikuBERT-BiLSTM-CRF model outperforms others overall. Large language models, including Baichuan2-13B-Base, also exhibit commendable capabilities and robust generalization abilities, showcasing the potential of large language models in this domain.

• 来自： 王东波
• 分类： 信息资源管理
• 稿件状态： 已投稿期刊
• 引用： PSSXiv:202502.00522 (或此版本 PSSXiv:202502.00522V1)
  DOI:10.12451/202502.00522
  CSTR:32012.36.PSSXiv.202502.00522
• 推荐引用方式： 贾智,赵志枭,沈思.大语言模型下古籍语间跨语言时间表达式抽取研究.哲学社会科学预印本平台:https://zsyyb.cn/abs/202502.00522.[PSSXiv:202502.00522V1] (点此复制)