Semi-Supervised Learning to Extract Attribute-Value Pairs from Product Descriptions on the Web

We describe an approach to extract attribute-value pairs from product descriptions on the Web. The goal is to augment product databases by representing each product as a set of such attribute-value pairs. Such a representation is useful for a variety of tasks where treating the product as a set of attribute-value pairs is more useful than as an atomic entity. Examples include product recommendations, comparison of single products or complete offerings, and demand forecasting. We formulate the extraction as a classification problem and use Naı̈ve Bayes combined with a multi-view semi-supervised algorithm (co-EM). The extraction system requires very little initial user supervision: using unlabeled data, it automatically extracts an initial seed list that serves as training data for the classification algorithm. In addition to the automatically extracted training data, the co-EM algorithm uses the unlabeled data to extract product attributes and values. Finally, the extracted attributes and values are linked to form pairs using dependency information and co-location scores. We present promising results on Web product descriptions in two categories of sporting goods products.