The Research And Application Of Sentiment Classification Based On Transfer Learning

With the development of Web 2.0,the sentiment classification task of User Generated Content(UGC)in the new domain gradually embodies great academic and commercial value.Aiming at the problem of poor sentiment classification performance caused by lack of large labeled corpus in new domain,this thesis explores and studies the task of cross-domain sentiment classification,and a Shared Knowledge Learning and Transfer Model(SKLT)is proposed.The experimental results show that SKLT model can effectively improve the performance of sentiment classification in new domain.In cross-domain sentiment classification tasks,based on the method of Transfer Learning,domain-independent sentiment knowledge can be extracted by using labeled data from several related fields,combined with improved Generative Adversarial Net(GAN),the accuracy of emotional classification tasks in new fields is improved after the knowledge is applied to the new field.The main work of this thesis is divided into three parts:In the shared k:nowledge learning part of SKLT,based on bi-GRU sentiment classification model,this thesis improves GAN by optimizing discriminator.This thesis extracts domain-independent shared knowledge,and realizes redundant feature penalty using vector orthogonality,and non-redundant shared sentiment knowledge and domain-specific sentiment knowledge can be successfully obtained.In the part of shared knowledge transfer of SKLT model,based on the method of transfer learning,this thesis applies the shared knowledge transfer extracted from source domain to new domain,and realizes new domain adaptation by using the idea of "partial weight transfer".The experimental results show that the shared knowledge extracted by SKLT model can effectively improve the accuracy of sentiment classification task in the new domain.In order to enhance the interpretability of sentiment knowledge,an"Attention Mechanism-based SKLT model" is proposed in this thesis,which realizes the visualization of shared knowledge and domain-specific knowledge,and the effectiveness of the model is verified.At the same time,this thesis also builds a prototype system of sentiment classification based on SKLT model,which can be applied to sentiment classification tasks in more new fields.