Fine-Grained Opinion Mining For User Reviews

The consumption pattern in the big data era is gradually developing from offline to online.The consumption platform has accumulated a large number of user reviews,which reflects the attitude and views of consumers,and also has a certain feedback effect on service and commodity providers.In view of the user reviews' large amount and high gathering speed,it takes a lot of time and energy to extract useful information manually.Therefore,a method of opinion mining and analysis for massive user reviews is urgently needed.A user review often contains several opinion words,which correspond to specific opinion aspect.For example,”hot pot”,”delicious” and other opinion words in the catering reviews belong to the aspect of ”dishes”.This thesis focuses on two tasks: the opinion aspect extraction and the opinion aspect sentiment analysis,to mine the fine-grained subjective emotions in user reviews,so as to help consumers get better consumer experience and make stores improve pertinently.In the work of aspect extraction,this thesis extracts the opinion words as specific entities,and divides them into specific opinion asepct.An improved ELMo pre-training model called fastELMo is proposed.Compared with other static word embedding methods,fastELMo integrates the syntactic features of reviews and solves the vector representation problem of polysemous words.In the task of opinion word extraction,the network structure of fastELMo combined with BiLSTM improved by more than 1% compared with other experimental models.Opinion words are divided into specific opinion aspects based on text similarity algorithm,In this thesis,the idea of Maximal Marginal Relevance algorithm is innovatively introduced into similarity calculation,which improves the effect of opinion aspect extraction by more than 3%.In the work of sentiment analysis based on the opinion aspect,this thesis focuses on the shortcomings of the existing models,innovatively proposes a gated convolutional network called IGCNN,and conducts experiments on Chinese and English datasets respectively.IGCNN extracts semantic features from the left context and the right context of opinion words respectively and then combines them together.The experimental results show that it has an accuracy of more than 70% in both data sets,which is about 1% higher than other excellent models.In terms of time efficiency,the model can train parallelly because of convolution and gating mechanism.Compared with other mainstream models such as IAN,which is based on LSTM and attention mechanism,the time spent on 100 iterations is reduced by nearly 1/2.On the basis of the good performance of the model,this thesis integrates it into a fine-grained opinion mining system for user reviews,realizes the real-time crawling and opinion analysis of user reviews,it extracts the hidden opinions effectively,and give a visual and clear result.The further work will focus on the research of end-to-end opinion mining methods and different domain transfer method of opinion mining algorithm.