«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年07期

页码:

133-139

栏目:

人工智能

出版日期:

2025-07-10

文章信息/Info

Title:

Research on Text Classification of Typical Hazardous Chemical Accidents Based on Large Language Models

文章编号:

1673-629X(2025)07-0133-07

作者:

张又元1; 2; 马新春1; 2; 赵军3

1. 新疆大学 软件学院,新疆 乌鲁木齐 830046;

2. 新疆电子研究所股份有限公司,新疆 乌鲁木齐 830013;

3. 新疆维吾尔自治区安全科学技术研究院,新疆 乌鲁木齐 830000

Author(s):

ZHANG You-yuan1; 2; MA Xin-chun1; 2; ZHAO Jun3

1. School of Software Engineering,Xinjiang University,Urumqi 830046,China;

2. Xinjiang Electronics Research Institute Co. ,Ltd. ,Urumqi 830013,China;

3. Xinjiang Uyghur Autonomous Region Institute of Safety Science and Technology,Urumqi 830000,China

关键词:

危化品事故; 大语言模型; 中长文本; 中间推理过程; 提示模板; 微调

Keywords:

hazardous chemical accidents;  large language models;  medium and long texts;  intermediate reasoning process;  prompt templates; fine-tuning

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0054

摘要:

为了实现对危化品事故案例的有效管理,首先需要对事故案例文本进行精确分类。 尽管当前的大语言模型(LLMs)在经过简单微调后能够在特定领域的中长文本分类任务中表现良好,但却忽视了生成式大语言模型强大的推理能力对于此类任务的重要促进作用。 生成式大语言模型不仅具有生成分类结果的能力,还具有生成推理过程的能力。 基于此,该文提出了一种基于推理模式的大语言文本分类模型,具体构建流程如下:首先,利用大型 LLMs“通义千问 2. 5”的推理能力,模拟生成连接案例文本和真实标签的中间推理过程;然后,将生成的推理过程编码为结构化的提示信息,并嵌入到用于分类任务的提示模板中;最后,在低资源条件下,选择小型 LLMs“Qwen1. 5-4B”作为该文采用的文本分类器,利用“通义千问 2. 5”LLMs 构建的提示模板进行微调。 实验证明,该方法在危化品事故案例的小样本数据集中表现优异,其F1 值达到了 90. 22% 。 此外,在公开新闻数据集上验证了该方法的泛化性能,其 F1 值也达到了 88. 04% 。

Abstract:

To effectively manage hazardous chemical accident cases,precise classification of the accident case texts is essential. Although current large language models (LLMs) perform well in specific domain’s medium and long text classification tasks after simple fine-tuning,they overlook the significant role of generative LLMs’ powerful reasoning capabilities in such tasks. Generative LLMs not only have the ability to generate classification results but also to generate the reasoning process. Based on this,we propose a large language text classification model based on reasoning patterns. The specific construction process is as follows:Firstly,the reasoning ability of the large LLM “Qwen 2. 5” is utilized to simulate and generate the intermediate reasoning process connecting the case text and the real label.Then,the generated reasoning process is encoded as structured prompt information and embedded into the prompt template used for the classification task. Finally,under low-resource conditions,the small LLM “ Qwen 1. 5 -4B” is selected as the text classifier,and the prompt template constructed by “ Qwen 2. 5” LLMs is used for fine - tuning. Experiments prove that the proposed method performs excellently in the small sample data set of hazardous chemical accident cases,with an F1 value reaching 90. 22% . Additionally,the gener-alization performance of the proposed method is verified on the public news dataset,with an F1 value reaching 88. 04%.

常用功能

更新日期/Last Update: 2025-07-10