Research And Application Of Chinese Text Sentiment Analysis Method Based On Transfer Learning

With the development of information technology,more and more users join the Internet and publish a large amount of self created data.Most of these self created data contain users' personal emotional information.How to extract emotional information from massive data and make use of it is the main research content of the subtask sentiment analysis in the field of natural language processing.Sentiment analysis task plays an important role in public opinion monitoring,commodity recommendation,sales decisionmaking and other fields.However,a good deep sentiment classification model often needs a large number of labeled samples.And new fields emerge one after another.If we only rely on manual annotation,the cost is too high.If we directly apply the existing classification model to new fields,it is often difficult to achieve ideal results.Based on the transfer learning method,this thesis proposes a relationship based emotional knowledge learning and transfer model(R-EKLT).Firstly,through joint training on multiple source domains with labeled data and target domains without labels,the model obtains a separate feature extraction network for each source domain and a cross domain shared feature extraction network,domain discriminator network and sentiment classification network.Through the shared feature extractor,the model learns shared emotional knowledge.Using the idea of knowledge distillation,the feature extraction network of several source domains is used as the teacher network and the feature extraction network of the target domain is used as the student network to transmit private emotional knowledge according to a certain weight.In the process of knowledge transfer,first accumulate a certain amount of pseudo label samples,and then directly transfer the shared feature extraction network and the private feature extraction network of the target domain to the target domain by means of model transfer.The pseudo tag data is used to fine tune the network parameters,which further improves the accuracy of sentiment classification in the target domain.The main innovations and contributions of this thesis are as follows:1.A new vectorization representation of Chinese text is designed.Taking the common word vector + part of speech tagged word vector as the final text representation of a Chinese word can alleviate the polysemy of a word in Chinese words to a certain extent.2.A relationship based emotional knowledge learning and transfer model R-EKLT is proposed.Compared with the traditional method,R-EKLT model not only learns the shared emotional knowledge of the source domain,but also weighted learn the private emotional knowledge of the source domain based on the relationship between the target domain samples and the source domain.3.Combined with self-attention mechanism,the visualization of emotional knowledge is realized,and a cross domain sentiment classification prototype system based on R-EKLT model is built.It is more convenient to train sentiment classifiers in new fields without rich labels.