Intelligent Understanding Of Text For Product Reviews

With the rapid popularization of smart mobile devices(e.g.mobile phones),users can quickly generate information anytime and anywhere,which changes the information generation mode of the Internet from Web2.0 era,in which website employees are the main information producer,to now,in which users are the main information producer.Benefited from the convenience of information dissemination,online shopping has become a daily routine for many people.The online comment system and question answering system of the e-commerce Websites provide the interactive platforms for the customers: they can describe their shopping experiences on the comment system and they can also consult merchants or other customers about a specific product through the online question answering system.Customers generate large amount of text data on these platforms.Mining above texts and then using the hidden information can improve the efficiency of the online shopping and user experiences.Different from the formal texts such as official statements,the review texts include users' sentiment orientation.The sentiment analysis on the customer reviews is the base of the review text understanding.In recent years,deep learning methods achieved good performances on sentiment classification,but the lack of large-scale and high-quality labeled data is the bottleneck.To solve this problem,we proposed a weakly-supervised deep learning method.The ratings of the reviews are treated as the weakly-supervised signals to pre-train the model and then we fine-tune the whole model on the small amount of labeled data.This method achieves a good performance on the customer reviews.Based on the sentiment analysis,we proposed a quick response algorithm for the submitted questions to improve the efficiency of the online shopping.Traditional online Q&A system of the e-commerce Website has a drawback: users need to wait patiently for others' reply.But the instant demands of the questioners may disappear after they do not receive other's reply in a short time.However,customer reviews are an invaluable source of information where the user question may happen to be addressed therein.Based on this,we propose a deep multi-task neural network for product-related question answering.We use the existing question answer data to assist the learning of the mapping relationship between questions and relevant review sentences.Moreover,the submitted questions convey strong instant demands of the users on a specific product.Recommendation based on user questions can entice more consumption and improve user experiences.Hence,we propose a new recommendation problem,i.e.question-driven recommendation problem.To deal this problem,we design a question-driven attentive neural network(QDANN)to predict users' purchasing propensity.This method uses the questioners' instant demands and the general preferences to perform the purchasing behavior analysis.To sum up,the innovations of this paper are:(1)We propose a weakly-supervised deep learning method with the ratings as the weak labels.We use the ratings of the reviews as the weakly-supervised signals to pre-train the model,and a small labeled dataset is then used to fine-tune the parameters of the whole pre-trained model.A triplet-based training method is proposed to reduce the influence of the noises.(2)For product-related question answering,we propose a multi-task attentive model.The large amount of existing question answer data in the online Q&A system are used to assist the learning of the mapping relationship between questions and relevant review sentences.A partial sharing transfer training strategy is proposed for this multi-task model.(3)For the question-driven product recommendation problem,a question-driven attentive neural network is proposed to predict questioners' purchasing propensity.The instant demands and the personal preferences of the questioners affect their purchasing behavior.Relevance and positive sentiment orientation are the two important supporting evidences for the purchasing behavior.The attention mechanisms are used to provide explanations for recommendations.We evaluated the above methods on the public Amazon dataset and Taobao dataset.The results show the efficacy of our method and its superiority over baseline methods.