面向文本情感分析的BiTCN-SA模型的研究

Research on BiTCN-SA Model for Text Emotion Analysis

为解决传统神经网络模型不能完整获取文本特征表示以及上下文信息使用不足的问题,提出一种面向文本情感分析的BiTCN-SA模型架构。首先采用Word2Vec中的Skipgram词向量表示方法将词向量序列输入模型;然后借助BiTCN双向提取文本全局特征,同时很好地联系上下文序列特征;最后添加自注意力层以捕捉文本特征的内部相关性,帮助模型优化特征,提高情感分析的准确率。以BiTCN为基线模型,在大众点评餐厅评论数据集上对比了6种模型的情感分类结果,实验结果显示,BiTCN-SA模型的准确率可达95.1%。结果表明,与基线模型相比所提模型的准确率提升了2.3%;与主流深度学习方法CNN,LSTM,BiLSTM,TCN相比,准确率有了8.2%,14.4%,4.8%,9.6%不同程度的提升,均证明了BiTCN-SA模型在文本情感分析中的有效性。

In order to solve the problem that the traditional neural network model can′t acquire the text feature completely and the context information is not used sufficiently,this paper proposes a BiTCN-SA model framework for text sentiment analysis.Firstly,through the Skip-gram word vector representation method in Word2Vec,the word vector sequence is input into the model.Then,with the help of BiTCN network layer,the global features of the text can be extracted directionally,and the context sequence features can be well connected at the same time.Finally,a self attention layer is added after the network layer to capture the internal correlation of the features in the text to help the model optimize the feature vector to improve the accuracy of emotion analysis.BiTCN is used as a baseline model to compare the emotion classification results of six models.The results of experiments show that the accuracy of BiTCN-SA model in the restaurant review dataset is 95.1%.The results of experiments show that the accuracy of the proposed model is improved by 2.3%compared with the baseline model,and by 8.2%,14.4%,4.8%and 9.6%compared with the mainstream deep learning methods CNN,LSTM,BiLSTM and TCN,the effectiveness of BiTCN-SA model in text sentiment analysis is proved.