Research On Text Sentiment Classification Method Based On Deep Learning

With the rapid development of Internet technology,commentary texts have gradually increased,and text sentiment analysis based on the emotional semantic features extracted from the text has become one of the hot tasks in the field of NLP.The feature form extracted by traditional methods is too simple to dig deep into the hidden information of the text.In recent years,deep learning has achieved good results in the field of NLP,but the neural network model is single and independent,and there are still problems of context semantics and long-distance dependence.Based on deep learning,this paper first aims at the problem that a single neural network model cannot fully aggregate its own advantages.,proposed a method of fused neural network model for sentiment classification;in view of the problem of losing sentiment information in the traditional word vector representation process,a method of obtaining fused part-of-speech vectors was proposed for feature extraction.The specific results are as follows:(1)Aiming at the problem that a single neural network model cannot fully extract sentence sequence features,this paper proposes an improved fusion model(Att-CLSTM),which can first obtain highly abstract text features,and use the convolution operation of CNN to extract textual features.Abstract features of neighboring words;adding LSTM layer to extract context information features,and obtaining more text features during deep model training;in order to further distinguish the information in the text that affects the classification results,the Attention model is integrated to change the weight of sequence information,which is an important text feature Give greater weight to focus on the textual information that affects the classification,so as to make full use of the textual information.In order to further verify the performance of the model,this paper conducts experiments on the public IMDB movie review dataset,and the results show that the classification accuracy of the model has been improved to a certain extent.(2)Aiming at the problem of losing emotional information in the traditional word vector representation process,this paper proposes a fusion model(PB-CLSTM),which pays attention to part-of-speech features in the process of word vector processing,and splices the traditional word vector and part-of-speech vector.Obtain the fusion part-of-speech vector;use the BERT model to obtain continuous and related input sequences,and input the results obtained by the pre-training model into the fusion model to further extract features and output classification results.This paper conducts experiments based on Tan Songbo Chinese hotel evaluation data set and Jingdong platform commodity evaluation data set,verifies and explores the relationship between the text word vector dimension of the proposed model and the experimental results,and finally determines the optimal parameter value;at the same time,set Dropout The strategy alleviates the overfitting problem during the experiment,and the experimental results show that the classification accuracy of the model has been improved to a certain extent.