Research On Aspect-based Sentiment Analysis Method For User Reviews

With the flourishing development of development social networks,more and more people like to publish posts on the Internet to express their sentiment.For example,users can comment on purchased products,comment on takeaway orders and post their opinions about social hot spots on Weibo or other websites,etc.Studying the emotional tendencies in texts contains a lot of user's emotions,attitudes and opinions has become a very popular research direction in the field of natural language processing,which has practical commercial value and academic research significance.In the past,sentiment analysis research was mostly at the text level or sentence level and the text is classified as a sentiment.This kind of research has been unable to meet the needs of the“Internet+”economy.In order to get closer to practical applications,the Aspect-Based Sentiment Analysis(ABSA)came into being.The ABSA task is a finer-grained sentiment classification designed to identify the emotional polarity of a given object,such as the comment text“This restaurant tastes good but the price is slightly more expensive”,the sentiment polarity of“taste”is positive and the“price”is“negative”.The ABSA task consists of 4 subtasks:Aspect term extraction(AE),Aspect-term sentiment classification(ASC),Aspect category classification(ACC),and Aspect category sentiment classification(ACSC),the research of this paper is carried out around the ASC task.The traditional solutions of ASC are mostly based on sentiment dictionaries and machine learning methods,but these methods all rely on huge feature engineering and have performance bottlenecks.Deep learning can automatically learn and extract features by building a neural network,and it has become a mainstream method in the field of sentiment analysis.Therefore,this thesis also proposes different models based on deep learning to solve the ASC problem.In a word,the main work is as follows:1.In order to solve the traditional aspect-level sentiment classification task,we proposed a hybrid neural network HCL-MHA,which extracts text features by mixing the LSTM-MHA and the CNN feature extraction module.The bidirectional LSTM in LSTM-MHA can effectively learn contextual semantics from the forward and backward aspects of the target,and the MHA can capture long-distance dependence.On the other hand,the CNN network can extract local features,both complement each other,so that the model has both the global semantics of the text and the local sentiment semantics.The model has an experimental accuracy of 79.35%on the three-classification task of the Restaurant dataset,and 73.39%of the experimental accuracy on the the Laptop dataset.The F₁value is also higher than other baseline models.2.In order to solve the the problem that the small dataset in the aspect-level sentiment classification task limits the performance of the model,a transfer capsule network model MATC based on multi-head attention is proposed.The model uses the idea of transfer learning to apply document-level attention knowledge to aspect-level sentiment analysis tasks,and uses multi-head self-attention to extract global features to make up for the defect that CNN can only capture local features.The MATC model uses the capsule network as the classification layer to classify the sentiment polarity according to the length of the neuron vector to better improve the classification performance.Experiments show that the accuracy and F₁value of MATC are indeed better than the HCL-MHA model.In addition,we applied the document-level module of the MATC to the task of identifying zombie fans on Weibo.3.In order to solve the cross-domain problem in aspect-level sentiment classification task,a model DA-BERT was proposed which based on BERT and domain adversarial mechanism.The model is composed of three parts:BERT-SAN feature extraction module,sentiment classifier and domain classifier.The BERT-SAN module is composed of BERT stacked self attention network,the sentiment classifier only classifies the source domain dataset,the domain classifier realizes the feature mixing of the source domain and the target domain through domain confrontation,so as to realize the generalization of the model.Experiments on six cross-domain pairs composed of Restaurant,Laptop and Twitter show the effectiveness of the model.