Joyce Yue Chai, IBM T. J. Watson Research Center; Alan W. Biermann, Duke University; Curry I. Guinn, Research Triangle Institute
In a user-trained information extraction system, the cost of creating the rules for information extraction can be greatly reduced by maximizing the effectiveness of user inputs. If the user specifies one example of a desired extraction, our system automatically tries a variety of generalizations of this rule including generalizations of the terms and permutations of the ordering of significant words. Where modifications of the rules are successful, those rules are incorporated into the extraction set. The theory of such generalizations and a measure of their usefulness is described.