Sentiment Analysis Unsupervised Domain Adaptation Problem Research

Aspect based sentiment analysis(ABSA)is one of the most important tasks in the field of sentiment analysis,which has been widely concerned by academia and industry.The objective of ABSA is to extract aspects and identify the emotional polarity of aspects,which can provide more valuable information in business intelligence,public opinion analysis and other scenarios.Compared with coarse-grained sentiment analysis,ABSA needs more expensive annotation cost,and it needs to label all aspects and corresponding sentiment in the text.The supervised model needs a large number of labeled data for parameter optimization,and the high labeling cost limits the practical application.In the unsupervised domain adaptation scenario,the target domain does not have any labeled data,and only relies on the unlabeled data and the training data of the source domain.The unsupervised domain adaptation method for ABSA task uses labeled data from other fields for model training,and then performs ABSA in the target domain,which can effectively alleviate the problem of labeled data dependence.This paper studies unsupervised domain adaptation for ABSA based on BERT.Specifically,this work can be divided into the following two aspects:1.Cross domain ABSA by combining feature and instance based methods.In previous work,the domain adaptation methods for ABSA are based on the traditional neural network model(such as LSTM),and they are all from the perspective of learning shared features to reduce domain differences.Based on Bert,the task of cross domain ABSA is carried out from two perspective of feature and instance.The experimental results on four standard datasets show that this work achieves the performance of cross domain ABSA.Experimental results on four standard data sets showed a 6.92 percentage points of improvement(Micro-F1).2.Cross domain ABSA based on style transformation.The essence of domain adaptation problem is the inconsistent distribution of training data and test data.Through the experimental observation,it is found that in the ABSA task,the difference of data distribution is mainly reflected in the lexicon difference,that is,the distribution of aspects and opinion words is inconsistent.This paper proposes the idea of domain style translation to reduce domain differences.The experimental results show that the pseudo samples generated by domain style transformation can effectively improve the ABSA task of the target domain.Moreover,the training method of multi-source fusion of original samples and pseudo samples further improves the effect of cross domain ABSA.