Learning Wrappers Efficiently for Web Information Extraction Using Unlabeled Examples

In this paper, we describe techniques for learning wrappers efficiently using very few user-supplied labels (typically, 1 or 2 labels, all within a single page). This is an improvement over previous work, which require multiple labeled examples on multiple pages. In effect, it brings the power of the wrapper down to the level of the end-user, who can teach, by only a few demonstrations, the labels that the wrapper should learn to extract. In contrast to other techniques, our approach also uses unlabeled web pages to guide the selection of appropriate features for the wrapper. We propose techniques to automatically acquire these unlabeled web pages, without the need for the user to supply them.