Research On Text Sentiment Classification Based On BERT Model

With the rapid development of Internet,the text data on the network is becoming more and more and better value of text data mining as one of the requirements of people more and more concern Text data mining in the field of sentiment classification has become a hot research field This paper mainly study the outbreak of weibo comments sentiment classification problem,through deep emotional learning model to realize intelligent text sentiment classification.Based on BERT model,this paper firstly studies the function of feature selection method in improving the learning efficiency of BERT model.Traditional feature selection method is calculated on the data of training set.Due to the lack of comprehensive training set,some important words are not included in the feature words,which makes the model ineffective Aiming at the problem of text emotion classification,the method designed in this paper integrates the emotion score as the weight into the traditional feature selection,effectively realizes the optimization of the original feature selection algorithm,and can increase the weight of the emotion words,so as to better represent the text emotion features In this way,the obtained feature dictionary can reduce the over-dependence on training data,improve the generalization ability of the model,and reduce the over-fitting phenomenon.Meanwhile,feature selection,as a plug and play component,can be used according to the actual needs.Secondly,this paper adds mask operation to attention mechanism,The BERT-Masked Att-BLSTM model is proposed,and two mask methods,random mask and learning mask,are designed.Traditional attention mechanism in after reaching the appropriate depth,with the increase of depth,increase the training error.The mask attention mechanism increases the diversity of network structure and can effectively alleviate this problem.Mask attention mechanism also has augmented the effect of the sample,because after the mask is a different from the original text of the text data,this is equivalent to indirectly increase the sample The mask attention mechanism can achieve the same effect as increasing samples in the case of fewer training samples.Moreover,the mask attention mechanism can reduce the excessive mutual adaptation between elements,improve the generalization ability of the model,and reduce the over-fitting phenomenon And after the mask operation,the model will learn to classify incomplete and non-standard sentences,which has practical significance especially in the text data of weibo and other online platforms.The experimental results show that the feature selection algorithm can make the model learning more centralized and reduce the computational cost.Moreover,the feature selection method based on emotion score designed in this paper is superior to the traditional feature selection method in the data set used in this paper.And the Mutual Information-Sentiment Score feature selection method designed in this paper is superior to the traditional feature selection method in the data set used in this paper.The BERT-Masked Att-BLSTM model proposed in this paper can more easily obtain improved accuracy from a large increase in depth,and produces better results than the traditional BERT-Attention-BLSTM model.