Research And Implementation Of Text Sentiment Snalysis Based On Deep Learning

In recent years,deep learning technology plays an important role in natural language processing,and it has achieved remarkable result in research and application in this field.The framework based on deep learning has surpassed human beings in some aspects.Microblog has become an important platform for the public to express their opinions and emotion because it is simply and easily to use,rapidly to spread and other characteristics.Therefore,a large number of text data with emotion have been produced.It is of great significance to analyze the emotional tendency of these text data with emotional information,and dig out the emotional tendency contained in them.Therefore,this paper uses deep learning technology to analyze the text data of microblog,designs and implements a microblog emotion analysis system.In order to improve the level of sentiment analysis technology,this paper studies GRU and BERT,proposes a text sentiment analysis model based on BERT hidden layers,and proves the effectiveness of the model through experiments.The specific work contents are as follows:CNN is used to extract the local feature information of the input text.In the part of text feature extraction,aiming at the problem that the feature vector obtained by BERT can not fully represent the text feature information,this paper proposes to use convolutional neural network to obtain the local feature of the input text,which is a supplement to the emotional information of the text encoded by BERT.A CNN semantic extraction layer is added to the representation layer of BERT model to extract the local features of input text through convolution and pooling.The 12 hidden layers of BERT are fused to extract the global feature information of the input text.The features learned by each hidden layer of BERT are different.When using BERT pre-training model,the output of the last hidden layer is usually used as the feature representation vector of the input data,which can not fully represent the features of the input data.To solve this problem,this paper proposes a fusion model of 12 hidden layer features based on BERT.Linear,LSTM and GRU are used to connect 12 hidden layers respectively to get 12 feature representations of input data,which are spliced with the feature vectors from CNN feature extraction layer to get the complete semantic information of input data.The complete feature vector of input data is sent to the downstream model for training emotion analysis model.Through the experiment,it is found that using GRU to connect 12 hidden layers has the best effect.On this basis,the paper proposes to use GRU to group the hidden layers,and the experiment shows that the model with 2 GRU to divide 12 hidden layers into 2 groups,the first 6 hidden layers as a group,and the last 6 hidden layers as a group has the best performance,and the accuracy of the model has been effectively improved.Design and implement a sentiment analysis system.The model proposed in this paper is applied to microblog,and designes a microblog sentiment analysis system.The system crawls relevant topic data on microblog social platform,uses the model proposed in this paper to analyze the emotional tendency of the data,and displays the analysis results to users.