曾庆禹,洪蒙纳,李继庚.基于低秩适配微调大语言模型构建断纸信息知识图谱[J].中国造纸,2026,45(4):189-197 本文二维码信息
二维码（扫一下试试看！）
基于低秩适配微调大语言模型构建断纸信息知识图谱
Construction of a Web Break Information Knowledge Graph Based on Large Language Models Fine-tuned with Low-rank Adaptation
收稿日期：2025-09-24  修订日期：2025-11-14
DOI：10.11980/j.issn.0254-508X.2026.04.024
关键词:  大语言模型  知识图谱  断纸故障  参数高效微调
Key Words:large language models  knowledge graph  web break fault  parameter-efficient fine-tuning
基金项目:
作者单位邮编
曾庆禹* 华南理工大学先进造纸与纸基材料全国重点实验室,广东广州,510640 510640
洪蒙纳* 华南理工大学先进造纸与纸基材料全国重点实验室,广东广州,510640 510640
李继庚 华南理工大学先进造纸与纸基材料全国重点实验室,广东广州,510640 510640
摘要点击次数: 470
全文下载次数: 77
摘要:为实现对造纸过程断纸故障的精准预测与知识结构化，本研究针对断纸数据特点与造纸设备影响，结合低秩适配（LoRA）和思维链提示工程技术，系统对比了LoRA、权重分解低秩适配（DoRA）、注入适配器（IA3）、量化低秩适配（QLoRA）4种参数高效微调（PEFT）策略对大语言模型的微调效果，并基于优选的LoRA微调大语言模型ChatGLM3-6B，构建了适配造纸领域实体、属性与关系的断纸信息知识图谱。结果表明，LoRA算法在断纸信息的实体识别与关系抽取任务中综合表现最优，召回率达100%、F1分数达92.31%，显著优于未微调ChatGLM3-6B及其他主流大语言模型（讯飞星火MAX、通义千问2.5-7B）。
Abstract:To achieve precise prediction and knowledge structuring of web break faults during the papermaking process, this study focused on the characteristics of web break data and the impact of papermaking equipment. By combining low-rank adaptation (LoRA) and chain-of-thought prompting engineering techniques, this research systematically compared the fine-tuning effects of four parameter-efficient fine-tuning (PEFT) strategies on large language models, including LoRA, weight-decomposed low-rank adaptation (DoRA), infused adapter by inhibiting and amplifying inner activations (IA3), and quantized low-rank adaptation (QLoRA). Based on the optimally LoRA fine-tuned large language models, ChatGLM3-6B, it constructed a web break information knowledge graph adapted to the entities, attributes, and relations within the papermaking domain. The results indicated that the LoRA algorithm achieved the best comprehensive performance in the entity recognition and relation extraction tasks for web break information, reaching a recall rate of 100% and an F1 score of 92.31%. Its performance significantly outperformed the non-fine-tuned ChatGLM3-6B and other mainstream large language models (such as iFLYTEK Spark MAX and Qwen2.5-7B).
查看全文   HTML   查看/发表评论  下载PDF阅读器