Research On The Application Of Cross-domain Sentiment Analysis Based On Domain Adaptation

As a research direction of Natural Language Processing,text sentiment analysis technology has made great progress in recent years,and has a large number of mature theoretical basis and available algorithms.Cross-domain text sentiment analysis is a higher-level task based on traditional text sentiment analysis technology.When we find a large number of text on the Internet,these data are often unmarked.It is difficult for us to train classifiers to conduct sentiment analysis system in an unsupervised environment.Therefore,we hope to transfer knowledge from another well-labeled dataset to help improving the model performance in the target domain.Different from the traditional iid assumption of semi-supervised tasks,there are often distribution shift between our two domains,which we call domain shift.The detailed research contents of this paper can be summarized as follows:Firstly,for narrowing domain shift,a parallel Ensemble Adaption Network is proposed in this paper.Firstly,we merge the two upper bounds of errors based on the Domain Divergence and Maximum Mean Discrepancy,and construct a larger upper bound.We hope to get better adaptive performance by enlarging the upper bound of error.Then,the network measures domain shift.For the hyper-parameters involved in the loss function in the network,it is difficult to use the hyper-parameter search method in unsupervised environment.Therefore,we propose a new metric,which we call "Average Uncertainty".Average Uncertainty is defined by conditional entropy,which can be used to measure the stability of the current system in an unsupervised training process.We find that the average uncertainty is inversely proportional to the accuracy.That is to say,when the model has high accuracy,it often has high stability.Secondly,we propose a Bi-Directional LSTM Parallel Network for cross-domain sentiment analysis tasks.After reading a lot of papers,we find Attention Mechanism,Bidirectional LSTM and Ensemble Adaptation Network can work conflict-free in an end-to-end deep network,and make up for the shortcomings of other models.The parallel ensemble adaptation network provides cross-domain capability of the model,and alleviates the problem of model precision decline caused by domain shift.The above framework implements the final cross-domain sentiment analysis model of this paper.Finally,we use Python 3.5.2+tensorflow-GPU 1.12.0+C# to design and implement a prototype system of cross-domain sentiment analysis,and report the system's requirements analysis,design details,system test and discuss the functions of the system.