Research On Corss-Domain Sentiment Analysis Based On Attention Mechanism

Sentiment analysis refers to the computational study of people's opinions,sentiments,emotions,attitudes,towards a certain entity.Cross-domain sentiment analysis aims to use labeled data from source domain as the training set to help train a sentiment analysis model about target domain and test on the data from target domain.When it comes to training a sentiment analysis model about new domain,annotating data is often the most expensive part,which limits the development process of the whole model.By using cross-domain sentiment analysis method,we can use the existing labeled data from other domain training the sentiment analysis classifier about the new domain,which greatly reduces the cost and development time.Attention mechanism can help the model get better semantic representation of text with less computing overhead and faster speed.Adding attention mechanism to the task of cross-domain sentiment analysis can promote the model to obtain more meaningful semantic representation of text,and get better alignment on the features from source domain to target domain.Firstly,we present Hierarchical Attention Transfer Network based on CrossDomain Attention(HATN_cda),which solves the shortcomings of existing methods that treat all data fairly,and cannot selectively highlight domain-specific data.Corssdomain Attention according to the characteristics of data,to give each of the data of different weight.For domain shared data,giving greater weight,for domain private data,giving less weight,thus to strengthen training about the private data from source domain,and focus on learning the influences of the private features on sentiment analysis.Experiments shows that the HATN_cda outperforms HATN in 17 of the 20 transfer tasks,and improve the accuracy by average of 0.53%.Secondly,we present Transfer Network based on Pre-training Attention Language Model(TN_palm),which solves the shortcomings of existing methods that words use unique embeding in different domain.BERT pre-training language model is adopted as the backbone network.Attention mechanism and context-based representation in BERT model are used to give different semantic representations of the same word in different contexts,so as to obtain more accurate semantic representations of texts.At the same time,the pseudo-label transfer method is combined to realize the cross-domain sentiment analysis.Experiments shows that TN_palm outperforms HATN and HATN_cda in all 20 transfer tasks,and the average accuracy improvement is 6.37% and 5.84%,respectively.In addition,we explore the influences of expanding number of source domain on cross-domain sentiment analysis based on TN_palm.We present Multi-Source Transfer Network based on Pre-training Attention Language Model(msTN_palm).msTN_palm expanded the source domain number from 1 to 4,try to fuse more source domain information.Experiments show that the performance of the msTN_palm is better than that TN_palm,and the average accuracy improvement is to 2.16%.