基于话题知识增强的立场检测大模型提示学习框架

The Prompt Learning Framework for Stance Detection Using Topic Knowledge Enhancement in Large-scale Models

立场检测旨在分析观点性文本(例如支持、中立或反对)对给定目标的态度。随着预训练模型的发展,现有方法主要基于微调框架构建立场检测模型。近期,提示学习框架在自然语言处理任务中取得了成功。然而,在实际应用场景中,面向立场检测构建提示学习框架仍然具有如下挑战:推文文本可能不会明确地表达某种态度,而是使用各种话题标签(#hashtag)来表达立场观点。文中设计一种背景知识增强的提示学习框架(Background Knowledge Enhanced Framework,BKEF),在框架中首先提出了一个主题发现模型来学习主题表示其次,提出话题知识增强的提示学习网络构建立场预测器最后,选用三个公开数据集对本文所提的方法进行评测实验结果显示,文中提出的BKEF方法优于现有方法。

Stance detection aims to analyze the attitudes of opinionated texts towards a given target,such as support,neutrality,or opposition.With the development of pre-trained models,existing methods mainly construct stance detection models based on fine-tuning frameworks.Recently,the prompt learning framework has achieved success in natural language processing tasks.However,building a prompt learn-ing framework for stance detection still faces challenges in practical applications.Tweet texts may not ex-plicitly express a certain attitude but use various topic labels or background knowledge to convey stance views.In this paper,we propose a background knowledge-enhanced prompt learning framework(BKEF).Specifically,we first introduce a topic model to learn topic representations.Then a prompt-learning network is proposed for integrate topic knowledge.Finally,we evaluate our method on three publicly available data-sets,and experimental results demonstrate that our proposed BKEF method outperforms existing methods.