Cross-Domain Sentiment Classification Based On Deep Learning And Attention Mechanism

The continuous development of the Internet is exerting a profund and subtle influence on people's life.With the continuous emergence of social media,such as news commentary,Twitter,We Chat and forum and so on,more and more people tend to express their opinions and comments on products,services or hot topics through online platforms.These massive online reviews are usually presented in the form of text,which contains rich sentiment information of users.The technology of text sentiment classification can mine the sentiment polarity information(positive and negative)effectively,which can provide powerful decision support for product recommendation,service management and social public opinion analysis and other fields.In recent years,online comments have covered an increasingly wide range of fields.In terms of data processing,it is time-consuming and laborious to manually annotate the data in each field,and there may be difficulties in insufficient comment data in a certain domain.In addition,considering the data distribution and feature space of source domain and target domain are often different,which lead to the fact that the same word may express different sentiment features in different domains,and different domains contain their own domain-specific features,namely,the classifier trained in the source domain is hard to directly apply to classification task in the target domain.Therefore,crossdomain sentiment classification is a research topic with important theoretical value and practical significance.Starting from the perspectives of deep learning and attention mechanism,in this paper,we study the cross-domain sentiment classification under the conditions of single source domain and multi-source domain,respectively:(1)Aiming at the problem of cross-domain sentiment classification under the condition of a single source domain,this paper proposes a Capsule network method with Identifying Transferable Knowledge(CITK)as common knowledge for cross-domain sentiment classification.Firstly,a semantically complete word embedding representation is obtained by using BERT(Bidirectional Encoder Representation from Transformers)to convert sentences to equal length;Secondly,we use the chi-square test and cosine similarity to extract SCP(Significant Consistent Polarity)words as domain-share features;Finally,the capsule network encodes the intrinsic spatial part-whole relationship constituting domain invariant knowledge,which bridges the knowledge gap between the source and target domains.Extensive experimental results on the Amazon product review datasets show that the proposed method has better performance for the cross-domain sentiment classification task.(2)In order to solve the problem of cross-domain sentiment classification under the condition of multi-source domains,this paper proposes a Multi-source Cross-domain Sentiment Classification model(MCSC).The main innovation of the model is to enhance the domain-aware word embedding and domain-aware attention.The model mainly consists of two subnetworks,of which the left subnetwork is used for domain classification: we use a combination of Bi LSTM(Bi-directional Long Short-Term Memory)and Self-attention mechanism to extract domain-specific features in input sentences,and then combine them with the regular word embedding of input sentences in the right subnetwork to build domain-aware word embedding.The right subnetwork is used for sentiment classification: we construct domain-aware attention by using domain-aware sentence representation as query vector to screen significant features and further improve the accuracy of cross-domain sentiment classification.A large number of experiments have been performed on Amazon product review datasets demonstrate that the proposed method can effectively improve the accuracy of multisource cross-domain sentiment classification.