The Research And Implementation Of Transfer Lrarning Based Sentiment Analysis

With the advent of the Web2.0 era,the rapid development of the Internet has greatly changed the way people express their opinions and feelings.Social platforms such as Weibo and forums are constantly emerging,and people are gradually accustomed to commenting on these platforms.The comments contain a lot of important information,such as sentiment orientation.Through the sentiment analysis of comments,it can help people to carry out product recommendation,public opinion analysis,etc.Therefore,it is of great practical significance to explore sentiment analysis methods with good performance.Existed sentiment analysis methods are mainly divided into methods based on sentiment dictionaries,machine learning,and deep learning.The traditional deep learning method relies on unsupervised training word vectors to represent the text,but this method does not fully express the contextual context of the text,and the recurrent neural network often used to process text is more complicated and difficult to train.In addition,with the emergence of new products in various fields,there isn't a large amount of tagged data to train models in new fields.Therefore,it is important to study how to use the existed tagged training data to conduct sentiment analysis in new fields.This paper studies the existed problem of sentiment analysis methods and explores the application of transfer learning techniques in sentiment analysis area.The main work includes the following three aspects:(1)The unsupervised training word vectors cannot represent the contextual context.This paper proposes a model-transfer based hierarchical attention network for sentiment analysis.It uses machine translation task to train an encoder and transfers the encoder structure to sentiment analysis for generating the distributed representation of the text.Since the translation model needs to fully extract the key information in the context to achieve the conversion from one language to another as accurately as possible.The word vectors obtained in this way cover the context and improve the performance of sentiment analysis method.(2)In this paper,the hierarchical attention mechanism neural network is used to deal with text sentiment analysis tasks.The network is mainly divided into word level and sentence level.A simplified neural network structure called minimum gate unit is used in each level to reduce the model parameters and simplify the training of the model,and attention mechanisms are introduced at each level to extract important information.(3)In order to solve the problem that the sentiment analysis method trained with the tagged data from one field cannot be applied to other fields,this paper proposes a cross-domain transferred sentiment analysis method based on feature,which uses an encoder to extract domain-invariant representations and domain-specific representations of the target domain,and then combine these two representations with the tagged data of the source domain and the few tagged data of the target domain to train the classifier to achieve cross-domain sentiment analysis.