方面级情感分类是一种细粒度的情感分析任务,旨在分类出文本中不同方面的情感.目前,现有方面级情感分类模型存在特征提取层次浅、泛化能力弱等问题.为此,该文提出一种基于融合对抗网络的方面级情感分类模型 ASFAN(Aspect-level Sentiment classification model based on Fusion Adversarial Networks).首先,从数据集中提取文本的方面词、位置、上下文信息表示.其次,将方面词、位置、上下文信息通过BERT编码.最后,通过多头注意力和局部注意力机制提取文本特征,将特征进行融合学习.此外,通过对抗学习算法生成对抗样本,将对抗样本作为一种文本数据增强样本,优化决策边界.实验结果表明,在SemEval 2014的Restaurant、Laptop数据集和ACL-2014的Twitter数据集上,ASFAN的准确率分别达86.54％、79.15％、76.16％,ASF AN对比大多数基线模型性能提升显著.
Abstract
Aspect-level sentiment classification is a fine-grained sentiment analysis task that aims to classify the sen-timent of different aspects of a text.,Existing aspect-level sentiment classification models have problems such as shallow feature extraction level and weak generalization ability.Therefore,we propose an aspect-level sentiment classification model based on fusion adversarial networks.Firstly,the aspect words,location,and contextual infor-mation representation of the text are extracted from the dataset.Secondly,the information of aspect words,posi-tion,and context is encoded by BERT.Finally,text features are extracted by multiple attention and local attention mechanisms,and then fused for learning.Moreover,the adversarial learning algorithm is used to generate adversari-al samples,which are used as a kind of textual data augmentation to optimize the decision boundary.The experimen-tal results show that the accuracy of the proposed method reaches 86.54％,79.15％,and 76.16％on the Restaurant and Laptop datasets of SemEval 2014 and the Twitter dataset of ACL-2014,respectively,outperforming most base-line models.
维普
万方数据