基于双输入和BiLSTM-MHSA的评论文本方面情感分类方法

Aspect-level sentiment classification method for comment texts basedon dual input and BiLSTM-MHSA

【目的】针对方面情感分类输入类别在不同领域之间差异较大,汽车用户评论文本语义信息不全,语义特征难以提取等问题,提出基于双通道输入的并行双向编码表征(bidirectional encoder representation from transformers,BERT)双向长短期记忆多头自注意力模型的方面情感分类方法。【方法】首先采用了方面情感和方面抽取的双重标签进行标注;其次通过并行的方面抽取和方面情感分类任务通道,分别使用BERT、双向长短期记忆网络(bidirectional long and short-term memory networks,Bi-LSTM)及多头注意力机制(multihead self-attention,MHSA)提取更深层次的语义信息及近距离和远距离特征信息;最后采用条件随机场(conditional random field,CRF)分类器和Softmax分类器进行分类。【结果】在相关的汽车用户评论文本数据集和多语言混合数据集上,本研究提出的模型相较于主流的方面情感分类方法,具有同步抽取方面词和判断情感极性的能力,且有效提高了方面词抽取和方面情感分类的准确率和F_(1)值。【结论】本研究提出的模型更有利于汽车销售者分析用户评论,同时对识别用户评论文本的情感极性的研究也有一定的参考价值。

[Objective]Aiming at the problems of great differences among different fields for the input categories of aspect sentiment classification,the incomplete semantic information of car user comment texts,and the difficulty in extracting semantic features,a parallel BERT bidirectional long-short-term memory multihead self-attention model based on dual-channel input was proposed for aspect-level sentiment classification.[Method]First,the dual labels of aspect sentiment and aspect extraction were adopted for labeling;then,the parallel aspect extraction and aspect sentiment classification task channels were mobilized,by means of BERT,bidirectional long and short-term memory networks(Bi-LSTM)and multihead self-attention(MHSA)respectively,to extract deeper semantic information and short-distance and long-distance feature information;finally,the conditional random field(CRF)classifier and SoftMax classifier were employed to classify.[Result]Compared with the mainstream aspect sentiment classification methods,the model can simultaneously extract aspect words and judge sentiment polarity on related car user comment text datasets and multilingual mixed datasets,and effectively improve the accuracy and F_(1)value of aspect word extraction and aspect sentiment classification.[Conclusion]The model proposed in this study is more conducive to analyzing user comments by car sellers,and also has certain reference value for identifying the sentiment polarity of user comment texts.