基于汉字多模信息与象形视觉对齐增强的古籍文本命名实体识别研究  

Named Entity Recognition of Ancient Texts Based on the Enhancement of Multimodal Information from Chinese Characters and Pictographic Visual Alignment

作　　者：郑旭辉 王昊[1,2] 裘靖文 Zheng Xuhui;Wang Hao;Qiu Jingwen(School of Information Management,Nanjing University,Nanjing 210023;Jiangsu Key Laboratory of Data Engineering and Knowledge Service,Nanjing 210023)

机构地区：[1]南京大学信息管理学院,南京210023 [2]江苏省数据工程与知识服务重点实验室,南京210023

出　　处：《情报学报》2025年第4期452-465,共14页Journal of the China Society for Scientific and Technical Information

基　　金：国家自然科学基金面上项目“关联数据驱动下我国非遗文本的语义解析与人文计算研究”(72074108)。

摘　　要：古籍的语义解析与人文计算是建设文化强国的重要组成部分,而古籍文本命名实体识别(named entity recognition,NER)是开展后续古籍知识发现与组织的前提和基础,设计一种适用于简体化文言文特性的命名实体识别模型具有重要的研究意义。汉字本身具有大量象形特征的视觉信息与发音信息,这些更贴合汉字发展历史的知识能够为识别古籍中的实体提供更多的信息以提高模型性能。基于此,本文构建了基于多模态汉字象形表示的GMAE-NER(guwen multi-information alignment enhanced NER)模型,该模型创新性地提出了汉字象形层面里图像与笔画信息的多模态特征处理和对齐方法,实现了将BERT(bidirectional encoder representations from transformers)表征与汉字视觉信息、发音信息相融合,有效增强了古籍文本命名实体识别的效果。本文将模型在纪传体史书《后汉书》上进行了大量的实验与对比,发现相较于基线模型,GMAE-NER在各个类别实体识别的F1指标上均得到了1.32~15.00个百分点的提升,并且能更好地识别出古籍文本中重叠表述的实体,消融分析结果也充分证明了该模型中视觉编码、发音编码与特征融合模块的有效性。The semantic analysis and digital humanities of ancient texts are crucial for cultural development.Named entity recognition(NER)is fundamental for the subsequent knowledge discovery and organization of these texts.Therefore,developing an NER model tailored to the characteristics of simplified classical Chinese is of significant research interest.Chinese characters inherently possess substantial visual and phonetic information with pictographic features,reflecting their historical development,which can enhance entity recognition in ancient texts.This study introduces the Guwen multi-information alignment enhanced NER(GMAE-NER)model,which leverages a multimodal representation of Chinese characters.The model employs a novel approach to process and align multimodal features and integrates bidirectional encoder representations from transformers(BERT)with visual and phonetic information regarding Chinese characters,thereby improving NER performance for ancient texts. Extensive experiments on the historical text Book of the Later Han Dynasty demonstrate that the GMAE-NER outperforms baseline models, achieving a 1.32-15.00 percentage points improvement in F1 scores across various entity categories and enhancing identification of entities with overlapping expressions. Ablation studies further validate the effectiveness of the visual encoding, phonetic encoding, and feature fusion modules of the model.

关 键 词：古籍文本 中文命名实体识别 汉字字形 汉字发音 跨模态交互融合 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]