Data Driven Discovery of Attribute Dictionaries

Online product search engines such as Google and Yahoo shopping, rely on having extensive and complete product information to return accurate and timely search results. Given the expanding scope of products and updates to existing products, automated techniques are needed to ensure the underlying product dictionaries remain current and complete. Product search engines receive offers from merchants describing product specific attributes and characteristics. These offers normally contain structured attribute-value pairs, and unstructured (textual) descriptions describing product characteristics and features. For example, a laptop offer may contain attribute-value pairs such as “model-X42” and “RAM-8 GB”, and a text description of the software, accessories, battery features, warranty, etc. Updating the product dictionaries using the textual descriptions is a more challenging task than using the attribute-value pairs since the relevant attribute values must first be extracted. This task becomes difficult since the text descriptions often do not follow a predefined format, and the data in the descriptions vary across different merchants and products. However, this information needs to be captured to ensure a comprehensive and complete product listing. In this paper, we present techniques that extract attribute values from textual product descriptions. We introduce an end-to-end framework that takes an input string record, and parses the tokens in a record to identify candidate attribute values. We then map these values to attributes. We take an information theoretic approach to identify groups of tokens that represent an attribute value. We demonstrate the accuracy and relevance of our approach using a variety of real data sets.