Fine-grain Product Image Classification And Sentiment Prediction In Product Reviews

Recently,with the rapid development of the Internet and e-commerce,online shopping has become a part of our life.How to use the massive images,texts and video data generated by online shopping platform to provide customers with a better online shopping experience has become a hot research topic.Online shopping platform set the categories followed with the attribute of the goods,in order to obtain the preferable selecting of the goods for the customers.It will spend most of the time and energy to decide the attribute of the goods.The image classification is one of the important and base research topics in the fields of artificial intelligence and computer vision.It is a very valuable research issue about how to use the image classification technology to automatically determine the attribute of the goods.Therefore,we utilize the framework of the deep convolutional neural network to explore the women's clothes of the image fine-grained classification problem.We also explore the imbalanced distribution of the attribute and the correlation between the different attribute of women's clothes goods.According to the online-shopping survey,we found most customers are willing to share their opinions by uploading visual and textual information on online shops.These opinions will not only offer certain references for the next consumer but also provide a research basis for us to explore the hot commodities and personal interests of customers.After analyzing this visual and textual information,we discovered that they have obvious sentiment.So the sentiments of product reviews are very basic for other applications.In this paper,we use deep learning to address the visual-textual sentiment analysis in product reviews.First,we introduce a new dataset for visual-textual sentiment analysis,termed as Product Reviews-150K(PR-150K),which is collected from the product reviews of online shopping websites.PR-150K consists of 150k manually-labeled image-text pairs with three sentiment categories(i.e.positive,neutral,and negative)and ninety product categories.Second,we propose a deep Tucker fusion method for visual-textual sentiment analysis,which efficiently combines visual and textual deep representations based on the Tucker decomposition and a bilinear pooling operation.Extensive experiments on our PR-150K,MVSO,and VSO datasets show that our method outperforms several state-of-the-art methods.