Research Of Cross-domain Sentiment Classification Methods Based On Domain Space Alignment

With the rapid development of social network,electronic commerce and Internet plus initiative technology,text sentiment classification technology becomes more and more important.Analyzing the sentiment polarity of texts,such as product evaluation,news and public opinion is helpful to understand users' attitude.In the commerce field,the text sentiment polarity can be used to analyze the approval degree of goods in the user's mind,which provides the basis for the commodity recommendation,information push service and so on.In the social news field,the text sentiment polarity can be used for public opinion analysis,which provides the basis for public opinion monitoring and information prediction.Therefore,the texs sentiment analysis becomes a hot research topic in the fields of natural language processing and artificial intelligence.Due to the wide range of reviews on the Internet,the difficulty in training a sentiment classifier for each domain is that we need to label enough samples for each domain by hand.And the differences between reviews in different domains and scenarios in practical applications can easily lead to the different distributions of samples in different domains.So,a classifier trained in a certain domain is difficult to directly apply to other domains.Therefore,the research of the cross-domain sentiment classification has become an important research topic in the field of natural language processing,especially using domain adaptive technology to achieve the knowledge transfer between domains and improve the accuracy and universality of the sentiment classifier.So,the research of the cross-domain sentiment classification has important value in application and research.In order to improve the accuracy of cross-domain sentiment classification,this paper uses domain adaption methods to eliminate the distribution divergence between samples in different domains by mapping the samples in different domains into the unified feature representation space for the feature alignment.The main research works and innovations are described as follows.1.This paper proposes a method for the cross-domain sentiment classification,which is named domain alignment based on multi-viewpoint domain-shared feature for crossdomain sentiment classification.The proposed method makes full use of the existing sentiment dictionaries and the mutual information algorithm to extract no sentiment polarity divergence domain-shared features from domans,and establishes direct mapping relationships between domain-shared and domain-specific features in the same domains based on feature word pairs with strong correlation relationships and the same sentiment polarity extracted by association rule algorithm and the syntax rule,respectively.By using the no sentiment polarity divergence domain-shared feature as the bridge,the proposed method establishes an indirect mapping relationship between domain-specific features in different domain,and constructs the unified feature space for different domains to eliminate the distribution divergence between different domains.The cross-domain experiments on the Amazon public product review dataset show that,compared with other cross-domain models,the proposed method can improve the accuracy of cross-domain sentiment classification and reduce the transfer loss.2.This paper also proposes a method for the cross-domain sentiment classification,which is named domain-invariant representation learning using an unsupervised domain adversarial adaptation deep neural network(DAA).The proposed method combines two modules: an additional task module and a specific task module into a unified deep network to extract domain-invariant,transferable and discriminative features,which can eliminate the distribution difference between samples from different domain and achieve the feature space alignment of different domains.In the additional task module,the domain classifier is designed by using the adversarial idea to update the values of parameters in the feature extractor and to ensure the features extracted by the feature extractor are domain-invariant.In the specific task module,the maximum mean discrepancy(MMD)constraints are added into domain feature alignment layers to map the features to reproducing kernel Hilbert spaces so as to increase the probability of feature transferability,feature matching in high dimensional space between different domains by using the kernel function.Aiming to the dataset shift problem existed in the image classification,an unsupervised domain adversarial adaptation deep neural network is established by using the convolution neural network as the feature extraction module.The cross-domain image classification experiments are conducted on the image datasets: Office-31,Office-Caltech and three digits image datasets: MNIST,SVHN and USPS,which show that the method is effective in solving the cross-domain image classification problem.The difference of image feature distribution between different domains is caused by external factors such as light,background and so on.And the experimental results also show that the generalization of the method proposed in this paper is improved.3.The paper designs cross-domain sentiment classification experiments to verify the usability of the domain-invariant representation learning using an unsupervised domain adversarial adaptation deep neural network.Because the marginalized denoising autoencoders(mSDA)can extract text robustness features,we use it to replace CNN as the feature extractor to establish a domain adaptation model named mSDA_DAA.The mSDA_DAA can reduce the time consumption for training a deep neural network by linear equation and improve the robustness of text features.Meanwhile,the mSDA_DAA combines the additional task module and the specific task module to learn a feature mapping function,which can map samples from the original space to a new feature space so as to match the feature distributions between different domains.This process makes full use of the adversarial idea,the domain adaptation theory and labeled samples in the source domain and unlabeled samples in the target domain to learn a feature mapping function.The cross-domain experimental results on the Amazon public product review dataset show that,compared with mSDA,mSDA_DAA can enhance the domain-invariant,transferable and label discrimination of the features learned by mSDA,and improve the accuracy of cross-domain sentiment classification.