The Research On Domain Adaptation For Sentiment Classification Of Product Reviews

Text sentiment classification technology has high commercial value and is one of the important issues in the field of natural language processing.At present,the supervised deep learning technology has made significant progress in solving the problem of sentiment classification.However,training deep models requires a large amount of labeled data,which limits the further promotion of this technology in the field of new products.In recent years,how to solve the problem of lack of training data has become a research hotspot in the field of natural language processing.For the sentiment classification in the new product field,we studied how to use the existing training data of other related source fields to improve the sentiment classification in the target field and reduce the labeling cost of the target field from the perspective of building cross-domain models and learning domain semantic representation.The main contributions and innovations of the text are:(1)A multi-source cross-domain sentiment classification method based on the Multi-Domain Attention Mechanism-Adversarial Training Bidirectional Gated Recurrent Unit(MDAM-ATBiGRU)model is proposed,which overcomes the limitations of the traditional single-source cross-domain method.In order to make full use of all source domain information in the presence of labeled data in multiple domains,the text adds a set of parameter matrices for learning the characteristics of each source domain domain in the attention layer.It enables the attention layer to further distinguish the input data according to the domain,and select the most significant part of the input data for the emotion classification of the specific domain.In order to suppress the influence of domain changes in text modeling and improve the robustness of the model,we conducted domain confrontation training on the BiGRU layer.Through experiments on Amazon multi-domain product review corpus,it is verified that the text method can effectively improve the performance of cross-domainsentiment classification.(2)A Domain Adaptation word embedding layer(DAL)based on dual-channel Convolutional Neural Network is proposed to enhance the domain semantics of universal text representation and improve the performance of existing sentiment classification models.The text first uses a keyword dictionary to map the domain-specific word embedding to the vector space in which the general word is embedded,and then combines the two through DAL learning weights to obtain a text representation suitable for the specific domain.DAL can be used as the input layer of the existing sentiment classification model.Experiments on Amazon product review corpus verify the effectiveness of the method proposed in this article.