Extracting diverse attribute-value information from product catalog text via transfer learning

Abstract

E-commerce sites are increasingly becoming the norm for how consumers search, purchase, and review products. Such sites internally list millions of products, creating a torrent of product options that can overwhelm a browsing consumer. To facilitate their search, it helps to annotate each product with a table of attributes describing general features such as color, size, etc. However, the tables must be provided by the merchant, so there is a business incentive to automate this task by extracting attribute-value information directly from product titles and descriptions. However, while past methods have done extraction for only a handful of attributes, in practice their exists hundreds of diverse attributes. In this thesis, we present a single model for extracting information on all attributes. In addition, we show that incorporating extra information about intra-attribute similarity improves performance for data-poor attributes.