摘要
方面级情感分类是自然语言处理研究领域的一个热点问题,旨在分类出文本中不同方面的情感.目前,大多数深度神经网络情感分类模型都采用均值注意力机制,这导致情感词不能有效获得相应权重的问题.为此,提出一种基于对抗学习的自适应加权方面级情感分类模型AWSCM(Adaptive Weighted aspect-level Sentiment Classification Model based on adversarial learning),旨在自适应地学习文本权重.首先,将训练文本预处理成方面词、句子、句子对形式的输入,通过BERT对输入编码.然后,通过对抗学习算法和训练文本计算扰动生成对抗样本.最后,通过注意力机制提取训练文本和对抗样本编码后的深层文本特征和自适应权重,再进行联合学习.实验结果表明,和大多数深度神经网络情感分类模型相比,AWSCM能提升情感分类的准确性.同时,通过消融实验,证明了AWSCM结构设计的合理性.
Aspect-level sentiment classification is a hot topic in the field of natural language processing,which aims to classify different aspects of sentiment in text.At present,most deep neural network sentiment classification models use the mean attention mechanism,which leads to the problem that emotive word can not obtain corresponding weight.Therefore,adaptive weighted aspect-level sentiment classification model based on adversarial learning is proposed,which aims is to give text weight adaptively.Firstly,the training texts is preprocessed into the inputs of aspect word,sentence and sentence pairs,and the inputs is encoded by BERT.Secondly,the perturbations are calculated from adversarial learning algorithm and training texts to generate adversarial samples.Finally,the deep text features and the adaptive weights of the training texts and adversarial samples are extracted through attention mechanism,and the model is jointly to learn them.The experimental results show that,compared with most deep neural network sentiment classification models,AWSCM can improve the accuracy of sentiments classification.Meanwhile,the ablation experiment proves that the structural design of AWSCM is reasonable.
作者
张华辉
冯林
廖凌湘
刘鑫磊
王俊
ZHANG Hua-hui;FENG Lin;LIAO Ling-xiang;LIU Xin-lei;WANG Jun(College of Computer Science,Sichuan Normal University,Chengdu 610100,China;College of Business,Sichuan Normal University,Chengdu 610100,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第4期766-772,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(71971151)资助。
关键词
方面级情感分类
注意力机制
BERT
对抗学习
自适应学习
aspect-level sentiment classification
attention mechanism
BERT
adversarial learning
adaptive learning
