Cross Domain Text Sentiment Analysis Based On Domain Adversarial Network

With the continuous development of web2.0,the Internet has penetrated into every aspect of people's lives,meanwhile,it provides a more convenient and broad information exchange platform for us.More and more users share their ideas on social platforms,express their views on the shopping platform about the products,and share their personal opinions on the news platform.As a result,a large amount of text data containing information about users' views are generated.Therefore,it is of great value to analyze the sentiment polarity of these textual data.Sentiment analysis,also known as opinion mining,is the process of analyzing,processing,summarizing and reasoning various online news resources,social media comments and other user-generated content.Text sentiment analysis is a branch of the sentiment analysis algorithm,so the typical supervised classification algorithms are suitable for textual sentiment polarity analysis.However,when the training data and the test data are not from the same domain,the prediction of the traditional classification method will be worse.Since the strong sentiment features from source domain that may no longer have these features or show other sentiment polarity in the target domain.In order to generalize the features of the source domain into the features of the target domain and conduct sentiment analysis in the target domain,cross-domain sentiment analysis becomes one of the solutions.Generally,the method of cross-domain sentiment analysis mainly focuses on extracting shared sentiment features from multiple domains using feature extractors.Recently,using the excellent feature extraction ability of feature extraction network in deep learning to extract shared sentiment features from different domains,and then using shared sentiment features for sentiment analysis is one of the main research directions of cross-domain text sentiment analysis at present.The deep learning scheme,which make use of the autoencoder and the domain adversarial mechanism to realize the extraction of shared sentiment feature.So this paper mainly optimizes the domain adversarial ability and extracts the semantic information of sentences by using network structure in deep learning,and then the extracted shared sentiment features from different domains are used for text sentiment analysis.The main research contents and innovations of this paper are as follows:(1)Because of the different features of different domains,the previous domain adversarial method was prone to gradient disappearance and gradient explosion,and the features had poor generalization ability.In this paper,we propose a domain adversarial approach based on Wasserstein distance,and use orthogonal constraints combine to extract the better deep shared domain features.At the same time,we use a stacked denoising autoencoder on the overall network structure,so that the feature extractor can extract the domain shared sentiment features with stronger robustness.(2)In view of the lack of sentence semantic information in shared sentiment features,we use BERT model to obtain the sentence semantic information,then use convolution neural network to help feature selection and feature dimension reduction.And then we use domain adversarial confusion the features between the source domain and target domain,at the same time use the data of source domain training sentiment classifiers.In the end,the affective polarity of the target field is predicted on the data set of amazon,and a good prediction result is obtained.