Research On Text Sentiment Classfication Based On Generative Adversarial Network And Heterogenous Ensemble Learning

With the rapid development of the WEB2.0 era,the way of people get information has gradually changed from the original traditional one-way propaganda media such as newspapers,periodicals,and broadcast to a new type of two-way Internet communication media.A large number of short text comments appeared on Internet platforms such as e-commerce,news,and social intercourse,and showed an exponential growth trend.How to analyze and mine the sentiment tendency of such data is one of the research hotspots in the field of Natural Language Processing(NLP).This research not only create huge business value,but also provide a very important reference basis for the public opinion supervision of enterprises and government agencies.In response to this research challenge,the main research work and innovations of this paper are as follows:1.Through relevant research on text sentiment classification tasks,it is found that a reasonable combination of Recurrent Neural Network(RNN)and Convolutional Neural Network(CNN)make the model fully synthesize the superior performance of the two during the learning process.Compared with using only its single model,the effect has been significantly improved,but the standard RNN and CNN have many problems such as gradient anomalies and importance features fuzzy.Therefore,this paper first combines the Bidirectional Gated Recurrent Unit(Bi GRU)and Deep Recurrent Neural Network(DRNN)into a Deep Bidirectional Gated Recurrent Unit(DBGRU).It is improved by designing an alternating direction propagation iteration mode between its network layers.The model not only avoid the problem of unreasonable update of information parameters caused by a long time span during the learning process,but also retain the semantic information in each direction of the text as comprehensively as possible,it fully considers the context information of the input words,and improves the effect of text semantic information processing.Then,in order to solve the problem that the important features in the CNN training process are not obvious,the single feature convolution of the standard CNN is expanded to multi-feature convolution,at the same time,the output vectors of the DBGRU and the Term Frequency-Inverse Document Frequency(TF-IDF)and sentiment attention of the fusion feature vectors are used to multiply calculations to give extra weights,and making it improve to a Multi-Feature Convolutional Neural Network(MFCNN).The output states of the hidden layer at each moment obtained by DBGRU training that using word vector and character vector of text distributed representation are input into MFCNN for feature extraction,and finally the pooled two feature vectors are fused for sentiment classification.Experiments show that DBGRU-MFCNN is able to further improve the accuracy of text sentiment classification relative to other related sentiment classification models.2.In view of the excellent performance of Generative Adversarial Network(GAN)and Variational Auto Encoder(VAE)in the field of image generation,this paper attempts to combine the two and apply it to the field of text processing,and then proposes a Variational Auto Encoder Ensemble Classifier Generative Adversarial Network(VAE-ECGAN).It use the DBGRU-MFCNN model as the discriminator(D)in GAN for text authenticity discrimination,while the encoder and decoder in VAE use a Bidirectional Long Short-Term Memory Network(Bi LSTM),and join The attention mechanism makes it possible to generate text and extract features of the text with attention.In addition,in order to be able to generate sentiment text of a specified category,this paper adds sentiment tags for instructional training which is based on the Auxiliary Classifier GAN(ACGAN),and make DBGRU-MFCNN,Naive Bayesian(NB)model,Decision Tree(DT)model,and Support Vector Machine(SVM)model combined through Stacking heterogeneous ensemble learning to serve as an auxiliary sentiment classifier for ACGAN.The final experiment shows that VAE-ECGAN not only makes its ensemble sentiment classification model through adversarial training better than DBGRU-MFCNN and other existing sentiment classification models in terms of classification performance,but also generate emotional text with better expression effects through VAE.Through the above two parts of the research results on the text sentiment classification problem,we know that the DBGRU-MFCNN based on RNN and CNN is better than the single model.At the same time,adding extra weights during the text feature learning process is able to make the model more convenient pay attention to the mapping relationship between features and categories.Integrating the idea of ensemble learning into VAE and GAN is able to make the model further improve the performance of text sentiment classification compared to other related models,and provide a better optimization strategy for the research of text sentiment classification based on deep learning.