Weakly-Supervised Acquisition of Open-Domain Classes and Class Attributes from Web Documents and Query Logs

Proceedings of ACL-08: HLT, pages 19–27, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Weakly-Supervised Acquisition of Open-Domain Classes and Class Attributes from Web Documents and Query Logs Marius Pas¸ca Google Inc. Mountain View, California 94043 mars@google.com Benjamin Van Durme∗ University of Rochester Rochester, New York 14627 vandurme@cs.rochester.edu Abstract A new approach to large-scale information extraction exploits both Web documents and query logs to acquire thousands of open- domain classes of instances, along with rel- evant sets of open-domain class attributes at precision levels previously obtained only on small-scale, manually-assembled classes. 1 Introduction Current methods for large-scale information ex- traction take advantage of unstructured text avail- able from either Web documents (Banko et al., 2007; Snow et al., 2006) or, more recently, logs of Web search queries (Pas¸ca, 2007) to acquire use- ful knowledge with minimal supervision. Given a manually-specified target attribute (e.g., birth years for people) and starting from as few as 10 seed facts such as (e.g., John Lennon, 1941), as many as a million facts of the same type can be derived from unstructured text within Web documents (Pas¸ca et al., 2006). Similarly, given a manually-specified tar- get class (e.g., Drug) with its instances (e.g., Vi- codin and Xanax) and starting from as few as 5 seed attributes (e.g., side effects and maximum dose for Drug), other relevant attributes can be extracted for the same class from query logs (Pas¸ca, 2007). These and other previous methods require the manual spec- ification of the input classes of instances before any knowledge (e.g., facts or attributes) can be acquired for those classes. ∗Contributions made during an internship at Google. The extraction method introduced in this paper mines a collection of Web search queries and a col- lection of Web documents to acquire open-domain classes in the form of instance sets (e.g., {whales, seals, dolphins, sea lions,...}) associated with class labels (e.g., marine animals), as well as large sets of open-domain attributes for each class (e.g., circu- latory system, life cycle, evolution, food chain and scientific name for the class marine animals). In this light, the contributions of this paper are four- fold. First, instead of separately addressing the tasks of collecting unlabeled sets of instances (Lin, 1998), assigning appropriate class labels to a given set of instances (Pantel and Ravichandran, 2004), and identifying relevant attributes for a given set of classes (Pas¸ca, 2007), our integrated method from Section 2 enables the simultaneous extraction of class instances, associated labels and attributes. Sec- ond, by exploiting the contents of query logs during the extraction of labeled classes of instances from Web documents, we acquire thousands (4,583, to be exact) of open-domain classes covering a wide range of topics and domains. The accuracy reported in Section 3.2 exceeds 80% for both instance sets and class labels, although the extraction of classes requires a remarkably small amount of supervision, in the form of only a few commonly-used Is-A ex- traction patterns. Third, we conduct the first study in extracting attributes for …

Content truncated to 500 words