基于prompt和知识增强的方面级情感分析  

Aspect-based Sentiment Analysis Based on Prompt and Knowledge Enhancement

作　　者：李阳 唐积强[3] 朱俊武[1] 梁明轩[1,2] 高翔 LI Yang;TANG Jiqiang;ZHU Junwu;LIANG Mingxuan;GAO Xiang(College of Information Engineering,Yangzhou University,Yangzhou,Jiangsu 225000,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China;National Computer Network Emergency Response Technical Team/Coordination Center of China,Beijing 100029,China)

机构地区：[1]扬州大学信息工程学院,江苏扬州225000 [2]中国科学院计算技术研究所,北京100190 [3]国家计算机网络应急技术处理协调中心,北京100029

出　　处：《计算机科学》2023年第S01期67-73,共7页Computer Science

基　　金：国家242信息安全计划项目(2021A008);北京市科技新星计划交叉学科合作课题(Z191100001119014);国家重点研发计划重点专项(2017YFC1700300,2017YFB1002300);国家自然科学基金(61702234);江苏省(扬州大学)研究生科研与实践创新计划项目(SJCX21_1551)。

摘　　要：方面级情感分析是一种新兴的细粒度情感分析任务,旨在根据给定句子和方面词判断情感极性。目前广泛使用的预训练语言模型由于训练目标和方面级情感分析的目标有差异,分析结果不好。为了缓解预训练语言模型和情感分析目标的差异,prompt被引入到方面级情感分析中,采用伪标签加方面词和意见词的方式创建prompt连续模板,并使用prompt-encoder训练伪标签使其拥有语义信息;然后,使用主题图注意力机制融合关于方面词和意见词的外部知识,根据融合外部知识的隐藏向量预测由情感词典组成的候选标签词;最后,采用求和置信度分数的方式将候选标签词的概率映射到情感极性分布空间上。实验表明,该模型在SemEval 2014任务的笔记本电脑数据集和餐厅数据集上将正确率分别提高了1.53%和3.5%。sentiment analysis is an emerging fine-grained sentiment analysis task that aims to judge sentiment polarity based on given sentences and aspect words.Currently widely used pre-trained language models are different due to their training objectives and those of aspect-based sentiment analysis,resulting in poor analysis results.In order to alleviate the difference between the pre-trained language model and the sentiment analysis target,prompt is introduced into aspect-based sentiment analysis,using pseudo-labels plus aspect words and opinion words to create prompt continuous templates,and using prompt-encoder to train pseudo-labels to have Semantic information;then,use the topic graph attention mechanism to fuse external knowledge about aspect words and opinion words,and predict candidate label words composed of sentiment dictionaries according to the hidden vector fused with external knowledge;The probabilities of label words are mapped onto the sentiment polarity distribution space.Experiments show that the model improves the accuracy by 1.53%and 3.5%on the Laptops dataset and Restaurants dataset of the SemEval 2014 task.

关 键 词：方面级情感分析 预训练语言模型 PROMPT 情感词典 知识增强 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]