Research On Fine-Grained Sentiment Analysis For Product Reviews

With the rapid development of the Internet and mobile Internet,online shopping sites have become the preferred platforms for people to shop.At the same time,Jingdong,Tmall,Taobao,Dangdang and other mainstream online shopping sites have accumulated a large number of product reviews.Analyzing these reviews can help consumers understand the real information of products and offer some suggestions for consumers when they select their favorite products.Based on user feedback in the reviews,merchants can make targeted improvements to the products.However,the number of product reviews are numerous,consumers and merchants need to spend a lot of time obtaining the product information they need,which is inconsistent with the characteristics of convenient and fast online purchase.Therefore,consumers and merchants urgently need an automated or semi-automated method to help them process coarse-grained product reviews and quickly obtain fine-grained product sentiment analysis.Fine-grained sentiment analysis aims to extract aspect information,and identify the sentiment views in terms of aspect.It has been a very hot research topic in recent years.Compared with general sentiment analysis,fine-grained sentiment analysis focuses on word or phrase,to extract the aspect and recognize the sentiment of aspect.Using fine-grained sentiment analysis to process product reviews can help consumers and businesses obtain effective information.Fine-grained sentiment analysis can be divided into two processes:attribute extraction and attribute sentiment recognition.Existing aspect extraction methods have achieved good results,but these methods often only consider the meaning of the word itself,ignoring the connection between words.In aspect sentiment recognition,there is also a problem that the recognition of sentiment words is incomplete,which affects the aspect sentiment polarity prediction.In order to solve the above problems,we propose a novel method based on dependency relationship embedding and conditional random field,to extract aspect terms from reviews.We also present a fine-grained sentiment analysis method based on dependency relationship embedding and attention mechanism.Specifically,the main contributions of this thesis are given as follows.Extracting aspect terms from product reviews is benefit for the sentiment analysis.It is capable of providing decision support for customers when they considering purchase goods,and providing feedback for merchants to improve the quality of goods and services.In this thesis,we propose a novel method based on dependency relationship embedding and conditional random field,to extract aspect terms from reviews.First,we use dependency relationship and dependency relationship embedding to construct three kinds of word feature representations.They are basic semantic information,structural semantic information and category semantic information.Then,the conditional random field model is applied to extract aspect terms with the three types of semantic information.The experimental results show that the performance is improved by 3.97%with DepREm-CRF.For the performance metric of F1,the proposed method outperforms the other competitive baselines by 7.65%.The work presented in this paper is able to effectively extract aspects terms from product reviews,laying a good research foundation for fine-grained opinion mining and sentiment analysis.Fine-grained sentiment analysis can help consumers or merchants understand product information and provide decision support for purchasing or improving products.In this thesis,we propose an LSTM model based on dependency relationship embedding and attention mechanism for fine-grained sentiment analysis of product reviews.First,the LSTM is applied to get the hidden state vector of the text.Second,the hidden state vector and the dependent word vector are combined as a new word vector.Then,the attention mechanism is designed to obtain the attention weight vector and weight representation of the text.and finally we use the softmax to obtain the predicted attribute emotional polarity.The experiments are conducted on three real data sets to demonstrate the performance of the proposed method.The experimental results show that DA-LSTM effectively improves the performance of attribute sentiment classification.