Semantic Class Learning from the Web with Hyponym Pattern Linkage Graphs

Proceedings of ACL-08: HLT, pages 1048–1056, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Semantic Class Learning from the Web with Hyponym Pattern Linkage Graphs Zornitsa Kozareva DLSI, University of Alicante Campus de San Vicente Alicante, Spain 03080 zkozareva@dlsi.ua.es Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 riloff@cs.utah.edu Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 hovy@isi.edu Abstract We present a novel approach to weakly super- vised semantic class learning from the web, using a single powerful hyponym pattern com- bined with graph structures, which capture two properties associated with pattern-based extractions: popularity and productivity. In- tuitively, a candidate is popular if it was dis- covered many times by other instances in the hyponym pattern. A candidate is productive if it frequently leads to the discovery of other instances. Together, these two measures cap- ture not only frequency of occurrence, but also cross-checking that the candidate occurs both near the class name and near other class mem- bers. We developed two algorithms that begin with just a class name and one seed instance and then automatically generate a ranked list of new class instances. We conducted exper- iments on four semantic classes and consis- tently achieved high accuracies. 1 Introduction Knowing the semantic classes of words (e.g., “trout” is a kind of FISH) can be extremely valuable for many natural language processing tasks. Although some semantic dictionaries do exist (e.g., Word- Net (Miller, 1990)), they are rarely complete, espe- cially for large open classes (e.g., classes of people and objects) and rapidly changing categories (e.g., computer technology). (Roark and Charniak, 1998) reported that 3 of every 5 terms generated by their semantic lexicon learner were not present in Word- Net. Automatic semantic lexicon acquisition could be used to enhance existing resources such as Word- Net, or to produce semantic lexicons for specialized categories or domains. A variety of methods have been developed for automatic semantic class identification, under the rubrics of lexical acquisition, hyponym acquisition, semantic lexicon induction, semantic class learn- ing, and web-based information extraction. Many of these approaches employ surface-level patterns to identify words and their associated semantic classes. However, such patterns tend to overgenerate (i.e., deliver incorrect results) and hence require addi- tional filtering mechanisms. To overcome this problem, we employed one sin- gle powerful doubly-anchored hyponym pattern to query the web and extract semantic class instances: CLASS NAME such as CLASS MEMBER and *. We hypothesized that a doubly-anchored pattern, which includes both the class name and a class member, would achieve high accuracy because of its specificity. To address concerns about coverage, we embedded the search in a bootstrapping process. This method produced many correct instances, but despite the highly restrictive nature of the pattern, still produced many incorrect instances. This re- sult led us to explore new ways to improve the ac- curacy of hyponym patterns without requiring addi- tional training resources. The main contribution of this work is a novel …

Content truncated to 500 words