面向事件因果关系分析的提示词构建方法  

Prompt words construction method for event causality analysis

作　　者：岳一峰 范嘉薇 张昊[1] 任祥辉 YUE Yifeng;FAN Jiawei;ZHANG Hao;REN Xianghui(The 15th Research Institute of China Electronics Technology Group Corporation,Beijing 100083,China)

机构地区：[1]中国电子科技集团公司第十五研究所网络安全部,北京100083

出　　处：《电子设计工程》2025年第9期7-11,共5页Electronic Design Engineering

基　　金：国家重点研发计划(2022YFB3103605)。

摘　　要：大语言模型在通用任务上具备优秀的理解和表达能力。然而在事件因果关系分析任务中,需要综合两个句子中的事件信息来确定因果关系的类别,这方面还存在很大的探索空间。为了降低大模型输出的不确定性,并且考虑到其受到提示词的约束,设计了一种面向事件因果关系分析的提示词构建方法——事件因果提示词方法。通过合理构建提示词,该方法可以约束大语言模型的反馈结果,提高对事件因果关系预测的准确率。实验结果显示,该方法既能帮助大语言模型进行解答思路分析和结果推断,也能引导大语言模型给出较准确的答案。其中,二分类推断效果相较于传统的预训练模型增长约40%,三分类分析结果增长约20%。Large Language Models(LLM)possess excellent language comprehension and expressive abilities in general tasks,but they are insufficient for analyzing causal relationships between events in the task,which requires the integration of event information from two sentences to determine the category of causality.Considering the constraints of carefully designed prompt words on the output of large models to reduce uncertainty,this study designs an eventcausality prompt word construction method.By constructing prompt words appropriately,this method can constrain the feedback results of LLM and improve the accuracy of predicting eventcausality.Experimental results demo-nstrate that this method enables Large Language Models to analyze answer strategies and infer results.Additionally,the proposed prompting word method in this study can prompt LLM to provide more accurate answers.Compared with the traditional pretraining model,the results of two-classification inference increased by about 40%,and the results of three-classification analysis increased by about 20%.

关 键 词：大语言模型 提示词 分类 动词因果关系 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN915.5[自动化与计算机技术—控制科学与工程]