电力行业知识性大语言模型构建方法研究  

Research on the Construction Method of Knowledge-Based Large Language Model in Power Industry

作　　者：赵必美 关文博 钟佳益 林全郴 ZHAO Bimei;GUAN Wenbo;ZHONG Jiayi;LIN Quanchen(Algorithm Division,Artificial Intelligence Technology Co.,Ltd,China Southern Power Grid,Guangzhou,510700,China;Key Laboratory of Speech Acoustics and Content Understanding,Institute of Acoustics,Chinese Academy of Sciences,Beijing,100190,China)

机构地区：[1]南方电网人工智能科技有限公司算法事业部,广州510700 [2]中国科学院声学研究所语音与智能信息处理实验室,北京100190

出　　处：《网络新媒体技术》2025年第2期50-55,共6页Network New Media Technology

摘　　要：随着大语言模型(LLM)的飞速发展,其在通用领域已经展现出强大的自然语言理解(NLU)、逻辑推理,以及自然语言生成(NLG)等能力。然而,大语言模型的“幻觉”问题,使其在垂直领域(例如,电力行业)的应用受到了限制。本文以电力行业为例,对构建该垂直行业知识性大语言模型的方法进行研究,提出一套完整的构建方案。该方案包含电力行业外部知识库构建方法,电力行业文档表征编码器训练方法以及大模型外部知识融合方法。实验表明,该方案在减少大模型“幻觉”现象,提高行业知识性问答准确率中起到了重要作用。With the rapid development of large language model(LLM),it has demonstrated excellent abilities in natural language understanding,logical reasoning,and natural language generation in the general field.However,the hallucination problem of LLM has limited its applications in vertical fields such as the power industry.This paper takes the power industry as an example to study the method of constructing a vertical industry knowledge-based LLM and proposes a complete construction scheme.This scheme includes the method for building external knowledge base,the training method for document representation encoder,and the method for integrating external knowledge into LLM in the power industry.The experimental results show that this scheme plays an important role in reducing the phenomenon of LLM hallucinations and improving the accuracy of industry knowledge-based question answering.

关 键 词：大语言模型 电力行业 知识增强 信息检索 幻觉现象 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]