Research On Text Sentiment Analysis Based On Deep Learning

With the development of information technology,increasing web users like to express their opinions on the website,which makes the amount of information on the internet grows rapidly.Analyzing the emotional tendency of comment information and classifying them is the research content of text emotional analysis.As the research focus of natural language processing,sentiment analysis is widely used in policy feedback,public opinion analysis,product improvement and other fields.Compared with traditional sentiment analysis methods,sentiment analysis based on deep learning does not need higher labor cost and computational complexity to construct feature engineering.It only needs to design flexible network structure to deal with different length of text.Then it can learn the deep features of input data and get better effect of emotion classification by training model using text dataset.In this paper,convolutional neural network,long-short term memory network and attention mechanism are mainly used for text emotion analysis.The main work completed is as follows:(1)An emotion analysis algorithm based on multi-channel CNN and LSTM is studied.The algorithm performs convolution and pooling operations on multichannel word vector to form multichannel CNN.Then it innovative combines multichannel CNN and LSTM network to extract text features.The proposed method is compared with other models on four data sets of MR,SST-2,Subj and MPQA.The results show that the model in this paper improves the accuracy and AUC value of emotion recognition,which proves that the proposed method enhances the performance of emotion classification algorithm.(2)An emotion analysis algorithm based on multi-feature fusion CNN and LSTM network is studied in this paper.In order to fully extract the global features,local semantic features and long-distance dependencies of text,the algorithm uses a new global-local-global mode to combine CNN and LSTM network to form an LCL model.Then innovatively introduces multichannel word vectors into the BiLSTM network to extract the context semantic information,and finally fuses the above extracted multiple features in a parallel mode.The algorithm uses Chinese data set and English data set for experimental comparison.The results show that the model proposed in this paper can better integrate the global and local features of text and extract the context semantics and emotional information more accurately.(3)An emotion analysis algorithm based on cascade convolution and attention mechanism is studied.The ascade layer is added to convolution layer and attention mechanism.The algorithm cascades the deep features after multi-layer convolution with the attention weight of the local features to learn the effective feature representation of text.The algorithm uses multi-label datasets for experiments.The experiments analyze and compare effects of different activation functions and optimization algorithms on the proposed algorithm to select the most suitable activation function and optimization algorithm.It also compares the performance of algorithm in this paper with other models by several evaluation quotas.