自动化学报2025，Vol.51Issue(2)：327-342,16.DOI:10.16383/j.aas.c240069
基于大语言模型的中文实体链接实证研究
An Empirical Study of Chinese Entity Linking Based on Large Language Model
摘要
Abstract
Large language models(LLMs)have recently made significant advancements in natural language pro-cessing.When scaled sufficiently,large language models exhibit reasoning capabilities that traditional pre-trained language models(PLMs)lack.In order to explore how to apply the emergent capabilities of large language models to the Chinese entity linking task,the following four methods are adapted:Knowledge augmentation,adapter fine-tuning,prompt learning,and in-context learning.Empirical studies on the Hansel and CLEEK datasets show that supervised learning methods based on Qwen-7B/ChatGLM3-6B outperform PLM-based methods.It achieves im-provements ranging from 3.9％to 11.8％on the Hansel-FS dataset,0.7％to 4.1％on the Hansel-ZS dataset,and 0.6％to 3.7％on the CLEEK dataset.When scaled to 72 billion parameters,Qwen-72B's unsupervised methods yield results comparable to the supervised fine-tuning of Qwen-7B,with a performance range of-2.4％to+1.4％.Furthermore,the large language model Qwen has a clear advantage in the long-tail entity scenario(11.8％),and as the number of parameters increases,the advantage will become more obvious(13.2％).The analysis of the error cases(hereinafter referred to as error analysis)found that the errors related to entity granularity and entity type accounted for a high proportion,36％and 25％respectively.This shows that in the entity linking task,accurately dividing entity boundaries and correctly judging entity types are the key to improving system performance.关键词
实体链接/大语言模型/知识增强/适配器微调/提示学习/语境学习Key words
Entity linking/large language model(LLM)/knowledge augmentation/adapter fine-tuning/prompt learning/in-context learning(ICL)引用本文复制引用
徐正斐,辛欣..基于大语言模型的中文实体链接实证研究[J].自动化学报,2025,51(2):327-342,16.基金项目
国家自然科学基金(62172044)资助Supported by National Natural Science Foundation of China(62172044) （62172044）
