Research on ancient text entity extraction based on domain-specific large language models
Digital Scholarship in the Humanities · 14.11.2026
DOI: 10.1093/llc/fqag127 ↗Kurz gefasst
Fachartikel in der Zeitschrift Digital Scholarship in the Humanities, online erschienen am 14.11.2026. Verfasst von Wenhua Zhao und 3 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben. Gefördert von: National Social Science Foundation of China.
Abstract (Originalsprache)
Large language models (LLMs) are well-known for their natural language generation abilities, but their performance in domain-specific text comprehension, particularly for Classical Chinese and English, has been less explored. This study focuses on bilingual data from Zizhi Tongjian in Classical Chinese and English. We construct an entity extraction instruction dataset by manually annotating official titles in fifteen volumes of Zizhi Tongjian. The annotations are verified using a BERT-based model. Based on this, we build an instruction dataset for entity extraction to fine-tune LLMs. The study involves fine-tuning two types of LLMs: the Xunzi series (domain-specific models for ancient texts) and general open-source LLMs using the LoRA framework. We experiment with different prompt formats to assess their impact on performance. Results show that, under identical conditions, the Xunzi-Qwen2-7B model outperforms other models. The best results were achieved by constructing instruction data from annotated original text, with entity-only instruction data also showing improvements. After fine-tuning with 5,000 samples, the Xunzi-Qwen2-7B model achieved an accuracy of 85.90 per cent, recall of 86.75 per cent, and F1 score of 86.32 per cent for official title extraction in Zizhi Tongjian.
Förderung
- National Social Science Foundation of China
Zitieren
APA
Zhao, W., Zhao, Z., Ye, W., & Wang, D. (2026). Research on ancient text entity extraction based on domain-specific large language models. Digital Scholarship in the Humanities. https://doi.org/10.1093/llc/fqag127
Vancouver
Zhao W, Zhao Z, Ye W, Wang D. Research on ancient text entity extraction based on domain-specific large language models. Digital Scholarship in the Humanities. 2026. doi:10.1093/llc/fqag127
BibTeX
@article{Zhao2026,
author = {Zhao, Wenhua and Zhao, Zhixiao and Ye, Wenhao and Wang, Dongbo},
title = {Research on ancient text entity extraction based on domain-specific large language models},
journal = {Digital Scholarship in the Humanities},
year = {2026},
doi = {10.1093/llc/fqag127}
}
EUVIMED