基于上下文学习和语义检索增强的零样本立场检测方法  

A Zero-Shot Stance Detection Method Based on In-Context Learning and Semantic Retrieval Enhancement

作　　者：卢银鹏 郭凯威 卢记仓[1] 周刚[1] 祝涛杰 LU Yinpeng;GUO Kaiwei;LU Jicang;ZHOU Gang;ZHU Taojie(Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《信息工程大学学报》2025年第2期196-202,共7页Journal of Information Engineering University

基　　金：河南省自然科学基金(222300420590)。

摘　　要：针对传统零样本立场检测方法需要依赖大量标注数据对模型微调的问题,提出一种基于上下文学习与语义检索增强的方法,以在不修改模型参数的条件下,提高大模型对未见对象的立场推理能力。通过语义相似度检索,从已有标注数据中筛选与待测文本相关的样例,并基于大模型的上下文学习范式,使用任务描述把相关样例和待测文本格式化为输入提示,以驱动大模型在更丰富的上下文语境下进行立场分类。实验结果表明,该方法在原始基础上提升了Flan-T5模型的零样本立场检测性能,并在SEM16数据集的细粒度零样本立场检测中显著优于原始模型。通过应用该方法,可以深化大语言模型对任务的了解,并激励模型参考相关样例中的关联知识以理解待测文本,从而更准确地推理关于未见对象的立场。To address the issue that traditional zero-shot stance detection(ZSSD)methods rely on a large amount of labeled data for model fine-tuning,a method based on in-context learning and semantic retrieval enhancement is proposed.The stance reasoning ability of large language models(LLMs)for unseen targets is improved without the need to tune model parameters.By retrieving semantically similar examples from existing labeled data and following the in-context learning paradigm of LLMs,relevant examples and test texts are formatted into input prompts with task descriptions,which enables models to perform stance classification in a more informative context.It is demonstrated through experiments that the ZSSD performance of the Flan-T5 model is improved by using this method,and the original model is significantly outperformed on the fine-grained task of the SEM16 dataset.The results indicate that the LLMs,understanding of the task can be deepened through this method,and relevant knowledge in examples is referenced to encourage LLM to gain better comprehension of test texts,thereby enabling more accurate inference of stances toward unseen targets.

关 键 词：零样本立场检测 上下文学习 语义检索 大语言模型 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]