Fine-grained named entity recognition of ancient texts based on multi-view semantic fusion and deep learning

To accurately extract the named entities present in ancient texts, an entity recognition method based on multi-view semantic fusion and deep learning is designed. This approach enables a comprehensive analysis of multi-view features and semantic relationships within ancient texts, thereby effectively advancing the research on automatic annotation of ancient texts. The Zuozhuan was selected as the experimental corpus, and four feature extraction methods were developed based on ShuoWenJieZi, aiming to capture fine-grained textual features from four distinct perspectives: Pronunciation (Shuo), Original Text (Wen), Cangjie Character Method (Jie), and Glyph Structure (Zi). Additionally, we proposed four feature representation methods that leverage multi-view semantic fusion to generate sequential vectors to represent ancient texts. Then, we constructed the MHA-CNN-BiGRU-CRF model, aiming to achieve high-precision identification of named entities in ancient texts. Our model could effectively identify three types of entities in the Zuozhuan: person, time, and location—achieving an F1-score 0.8837, surpassing the performance of existing deep learning models. The multi-view semantic fusion approach can capture the semantic knowledge of ancient texts from multiple dimensions and discover deep correlation information. The model proposed in this paper can accurately identify the named entities in ancient texts, effectively promoting the research on the digitization of ancient texts and digital humanities.