[extensive reading of papers] multimodal joint attribute prediction and value extraction for e-commerce product

Abstract

   This paper presents a multimodal method to jointly predict product attributes , With the help of product image, the value is extracted from text product description . We believe that product attributes and values are highly correlated . Given product attributes , It is easier to extract values . therefore , We jointly model attribute prediction and value extraction tasks from multiple aspects , To realize the interaction between attributes and values . Besides , For different product attributes and values , Product image has different influence on our task . therefore , We selectively extract useful visual information from product images to enhance our models . We annotate a file that contains 87194 Multimodal product attribute value data set of instances , The experimental results on this dataset show that , Explicitly modeling the relationship between attributes and values helps our method to establish the correspondence between them , The selective use of visual product information is a necessary condition to complete this task . Our code and data sets are available in https://github. com/jd-aig/JAVE.