Research On Sentiment Classification Towards Question-Answering Reviews

With the dramatical advancement of the Internet and amazing flourish of e-commerce,sentiment analysis has already attracted huge attention in both academic and industrial communities.On e-commerce platforms,it is meaningful to research sentiment analysis task towards user-oriented reviews due to its widespread applications,such as user feed-back and product recommendation.However,a large number of online merchants promote credit of their online shops and sales of their products through click farming,which brings numerous fake reviews.In addition,it also affects the users' shopping experience and the fairness on e-commerce platforms.To solve this problem,a novel reviewing form,namely user-oriented question-answering(QA)review,appears on many e-commence platforms.In this reviewing form,a potential customer asks questions about the target product,the platform picks some other customers with high credits who has already purchased this product,and these experienced customers provide answers.Comparing with traditional re-viewing form,this novel QA review can be more informative and convincing.In some de-gree,it can also avoids fake information.This thesis focuses on sentiment analysis towards question-answering reviews.Our exhaustive studies are as following:First,according to the characteristic of QA review,i.e.,both question text and answer text contain sentiment information,we propose a novel approach,namely bi-directional attention mechanism,to perform QA sentiment classification task.In detail,we first en-code both question text and answer text with bi-directional Long Short-Term Memory network respectively.Then,we calculate importance degrees of words in both question text and answer text through bi-directional attention mechanism simultaneously.Finally,we can get the sentiment information for QA review by these words' importance degrees.Em-pirical results,comparing with some baselines,demonstrate the impressive effectiveness of our bi-directional attention mechanism approach.Second,since traditional approaches can hardly deal with conflict instances and re-dundant information in QA review,we propose a novel neural network,namely hierar-chical matching network,to tackle these challenges in QA sentiment classification task.Specifically,we first segment both question text and answer text into sentences,and con-struct a number of<sentence,sentence>pairs where each pair contains one sentence from question text and the other sentence from answer text.Then,by leveraging a QA bi-directional matching mechanism,the proposed approach can learn the matching vectors for each<sentence,sentence>pair.Finally,we characterize the importance of the generat-ed matching vectors via a self-matching mechanism.Experimental results,comparing with a number of state-of-the-art baselines,demonstrate the impressive effectiveness of our proposed hierarchical matching network.Finally,due to the limited annotation corpus for QA sentiment classification,we pro-pose a joint learning approach to improve the performance for QA sentiment classification which treats QA sentiment classification as the main task while traditional review senti-ment classification as the auxiliary task.In detail,we first encode QA review into a senti-ment vector via main task model.Then,we employ an auxiliary task model,to learn auxil-iary sentiment information representation for QA review with the help of traditional review.Finally,we update the parameters in main task model and auxiliary task model simultane-ously with joint learning framework.Empirical results,comparing with a number of state-of-the-art baselines,demonstrate the impressive effectiveness of our proposed joint learning approach,which can integrate sentiment information between QA review and tra-ditional review well.