Research On Text Sentiment Analysis Method Based On Deep Learning

With the development of digital economy in recent years,all walks of life are undergoing digital transformation,which often greatly reduces the cost of subjectively evaluating a commodity or a hot news,etc.Therefore,people leave a large number of comment text data on the Internet,and it is of great value to dig out the emotional information hidden behind these texts.It can help decision makers to adjust their decisions.In this thesis,a model is built based on the data of the comments text to dig out the emotional tendencies contained in the text.In view of the current research on sentence-level sentiment analysis,most models only capture the contextual semantic information of the sentence,but do not pay too much attention to the emotion part of speech of the word itself,ignoring the existence of emotion words,which may cause words with opposite emotion part of speech to be mapped into similar vectors in similar sentence structures,resulting in the deviation of emotion analysis results.In view of current researches on aspect-level sentiment analysis,the potential problem of semantic feature-based sentiment analysis models is that when the sentences are too long,such models may not be able to capture semantic dependencies,while the potential problem of syntactic feature-based sentiment analysis models is that the sentences are too short and the features are not obvious enough.To solve the above problems in sentence level and aspect level emotion analysis,this thesis proposes two models respectively.1.Aiming at sentence-level emotion analysis,this thesis proposes a sentence-level emotion analysis model which combines emotion word vector and ALBERT model.The comment text was input at the word embedding layer,the ALBERT model was used to obtain the word vector with semantic information,and the SSWE emotion word vector was introduced to obtain the word vector with emotion information,and the two word vectors were spliced before being sent to the coding layer.The word vector with rich semantic information is sent to the Bi-LSTM layer,and the overall features of the statements are captured in the forward and reverse directions by the upper and lower LSTM layers respectively.At the same time,the Attention mechanism is introduced to better focus on some important words in the statements.Finally,by applying this model to the three data sets of ChnSentiCorp,NLPCC14-SC and ASAP＿SENT,the accuracy values are all above80%,which is better than the results of other comparison models,and the performance of the model is fully measured.2.For aspect-level sentiment analysis,this thesis proposes an aspect-level sentiment analysis model based on Graph Attention Network(GAT),which integrates semantic and syntactic features.For semantic information,a fusion graph attention network and multihead attention mechanism were used to extract semantic features,and several emotion words with high weight were selected as the determining factors on the weight matrix.Syntactic information is analyzed by using syntactic dependency tree,and then features are extracted by using GAT.Finally,the features of the two parts are integrated for classification,so that the model has stronger robustness.Finally,comparative tests were carried out on Restaurant,Laptop and Twitter data sets,and the results were all better than other common models.The accuracy rate of the model in Restaurant data set was 85.13%.