标签先验知识增强的方面类别情感分析方法研究  

Aspect-Category Based Sentiment Analysis Enhanced by Label Prior Knowledge

作　　者：吴任伟 李琳[1] 何铮 袁景凌[1] WU Renwei;LI Lin;HE Zheng;YUAN Jingling(School of Wuhan University of Technology,Wuhan,Hubei 430070,China;Deloitte Consulting(Shanghai)Co.,Ltd,Shanghai 510623,China)

机构地区：[1]武汉理工大学计算机与人工智能学院,湖北武汉430070 [2]德勤咨询(上海)有限公司,上海510623

出　　处：《中文信息学报》2024年第12期127-136,147,共11页Journal of Chinese Information Processing

基　　金：国家自然科学基金(62276196)。

摘　　要：当前,基于方面类别的情感分析研究集中于将方面类别检测和面向类别的情感分类两个任务协同进行。然而,现有研究未能有效关注情感数据集中存在的噪声标签,影响了情感分析的准确率。基于此,该文提出一种标签先验知识增强的方面类别情感分析方法(AP-LPK)。首先该文为面向类别的情感分类构建了自回归提示训练方式,可以有效利用预训练语言模型的学习能力。同时该方式通过自回归生成标签词,以期获得比非自回归更好的语义一致性。其次,每个类别的标签分布作为标签先验知识引入,并通过伯努利分布对其做进一步精炼,以减轻噪声标签的干扰。然后,AP-LPK将上述两个步骤分别得到的情感类别分布进行融合,以获得最终的情感类别预测概率。最后,该文提出的AP-LPK方法在五个数据集上进行评估,包括SemEval 2015和SemEval 2016的四个基准数据集和AI Challenger 2018的餐饮领域大规模数据集。实验结果表明,该文提出的方法在F1指标上优于现有方法。Current aspect-category based sentiment analysis aims at performing joint aspect category detection and category-oriented sentiment classification.To deal with the noisy labels often occurred in sentiment datasets,we propose an aspect-category based sentiment analysis approach with label prior knowledge(AP-LPK).Specifically,we firstly construct an autoregressive prompting learning that can effectively utilize the learning ability of pre-trained language models.And then label words are generated through autoregression for better semantic consistency than non-autoregression.Secondly,we introduce the label distribution of each category as label prior knowledge that is refined through Bernoulli distribution to mitigate the interference of noisy labels.And then,the output labels from the autoregressive prompting and label prior knowledge refinement jointly decide the prediction based on the distributions of sentiment polarities.Evaluated on five datasets including the four benchmark datasets from SemEval 2015 and 2016 and the Restaurant-domain dataset from AI Challenger 2018,the proposed approach outperforms existing baselines in terms of F 1.

关 键 词：基于方面类别的情感分析 提示学习 标签先验知识 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]