Research On Key Techniques Of Restaurant Comment Sentiment Analysis

With the development of Internet and electronic commerce,applications such as online shoping and online ordering which are fast and convenient have penetrated deeply into people’s daily lives.Accordingly,the comments which people delivered on these platforms are also increasing exponentially.These comments are in huge number and really have great research value.We analyze these comments to get the opinion polarity of every aspect term which not only can give consumers a lot of help on their consuming behavior but also can help businesses to understand consumer’s demand and thus to improve the quality of their product.In this thesis,we have a research on the method of aspect term extraction and aspect based sentiment analysis in the field of restaurant review.Then we apply the method which get the best result to the development of restaurant reviews sentiment analysis system.The study mainly includes the following aspects:Firstly,this thesis has a research on the task of aspect term extraction.We implement a model named output dependent bidirectional LSTM.The model use two independent hidden layer to analyse the comment so that it can make full use of the context of the comment.Besides,the model add self-connection to output layer to take full advantage of the relationship between output labels.Moreover,this thesis implement the method of CRF and improve the result by feature selection and combination.Then,we implement the BLSTM-CRF aspect term extraction method,which put the result of BLSTM to CRF to compute,and then get the best output labels.Secondly,this thesis focus on the method of aspect based sentiment analysis.We propose the model based on bidirectional LSTM which uses two BLSTM network that is BLSTML and BLSTMR to get the context semantic information of the aspect term.Meanwhile,we put the current word vector and the aspect term vector together to get a new vector and the put the new vector into the model,so that the model can capture the semantic relationship between every word and aspect term.This model get the best result among models in the same class.In addition,this thesis put forward the the method of model fusion based on boosting which fuse SVM with Random Forest.The method increase the weight of sample being incorrectly classified and decrease the sample being correctly classified after training the classifier.Finally,the final results are weighted according to the effect of each classification model.This method can combine the advantages of the linear classification model with the nonlinear classification model.Finally,this thesis implements a sentiment analysis system based on restaurant reviews.We apply the output dependent bidirectional LSTM method and the aspect based sentiment analysis method based on bidirectional LSTM into the system which improve the accuracy of aspect term extraction and aspect based sentiment analysis in the system.The system can visualy show people the proportion of the aspect term and sentiment polarity which consumer express to the restaurant by pie chart.