Automatically Learned vs. Hand-crafted Text Analysis Rules 1 Domain-speciic Text Analysis Figure 1: Output from a \management Succession" Text

As vast quantities of on-line text become available, there is an increasing need for systems that automatically analyze the conceptual content of natural language text. Systems that operate on narrowly de ned domains show promise, but require a di erent set of domainspeci c rules for each application. This paper describes CRYSTAL, a system that learns text analysis rules automatically from examples. Rules induced by CRYSTAL achieve performance approaching that of hand-crafted rules. CRYSTAL has a particularly e cient learning algorithm that is not improved by more extensive search. This o ers a practical alternative to time-consuming manual knowledge engineering for each new domain. 1 Domain-speci c Text Analysis With the increasing amounts of on-line text available, the need is growing for automated text analysis systems that go beyond keywords to extract the conceptual content of the text. This requires a system that can reliably extract both explicitly stated information and information that can be reasonably inferred. General purpose text understanding is still beyond the reach of current technology, but considerable progress has been made by restricting the problem to a prede ned set of concepts in a narrowly de ned domain. A text analysis system with the appropriate domainspeci c knowledge sources can identify references to information that is of interest to a particular domain, which consists of a corpus of texts together with a set of concepts to be identi ed in those texts. The target concepts in a medical domainmight be references to symptoms and diagnoses in patient records. In a collection of Wall Street Journal articles, the target This material is based on work supported in part by the National Science Foundation, Library of Congress and Department of Commerce under cooperative agreement number EEC-9209623 and in part by NRaD Contract Number N66001-94-D-6054. concept might be management succession events: persons moving into top management positions in corporations and persons moving out of those positions. The ARPA-sponsored Sixth Message Understanding Conference [MUC-6 1995] used such a \Management Succession" domain. This domain is illustrated by Figure 1.