Solving Relational Similarity Problems Using the Web as a Corpus

Proceedings of ACL-08: HLT, pages 452–460, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Solving Relational Similarity Problems Using the Web as a Corpus Preslav Nakov∗ EECS, CS division University of California at Berkeley Berkeley, CA 94720, USA nakov@cs.berkeley.edu Marti A. Hearst School of Information University of California at Berkeley Berkeley, CA 94720, USA hearst@ischool.berkeley.edu Abstract We present a simple linguistically-motivated method for characterizing the semantic rela- tions that hold between two nouns. The ap- proach leverages the vast size of the Web in order to build lexically-specific features. The main idea is to look for verbs, preposi- tions, and coordinating conjunctions that can help make explicit the hidden relations be- tween the target nouns. Using these fea- tures in instance-based classifiers, we demon- strate state-of-the-art results on various rela- tional similarity problems, including mapping noun-modifier pairs to abstract relations like TIME, LOCATION and CONTAINER, charac- terizing noun-noun compounds in terms of ab- stract linguistic predicates like CAUSE, USE, and FROM, classifying the relations between nominals in context, and solving SAT verbal analogy problems. In essence, the approach puts together some existing ideas, showing that they apply generally to various semantic tasks, finding that verbs are especially useful features. 1 Introduction Despite the tremendous amount of work on word similarity (see (Budanitsky and Hirst, 2006) for an overview), there is surprisingly little research on the important related problem of relational similarity – semantic similarity between pairs of words. Stu- dents who took the SAT test before 2005 or who ∗After January 2008 at the Linguistic Modeling Depart- ment, Institute for Parallel Processing, Bulgarian Academy of Sciences, nakov@lml.bas.bg are taking the GRE test nowadays are familiar with an instance of this problem – verbal analogy ques- tions, which ask whether, e.g., the relationship be- tween ostrich and bird is more similar to that be- tween lion and cat, or rather between primate and monkey. These analogies are difficult, and the aver- age test taker gives a correct answer 57% of the time (Turney and Littman, 2005). Many NLP applications could benefit from solv- ing relational similarity problems, including but not limited to question answering, information re- trieval, machine translation, word sense disambigua- tion, and information extraction. For example, a relational search engine like TextRunner, which serves queries like “find all X such that X causes wrinkles”, asking for all entities that are in a par- ticular relation with a given entity (Cafarella et al., 2006), needs to recognize that laugh wrinkles is an instance of CAUSE-EFFECT. While there are not many success stories so far, measuring seman- tic similarity has proven its advantages for textual entailment (Tatu and Moldovan, 2005). In this paper, we introduce a novel linguistically- motivated Web-based approach to relational simi- larity, which, despite its simplicity, achieves state- of-the-art performance on a number of problems. Following Turney (2006b), we test our approach on SAT verbal analogy questions and on mapping noun-modifier pairs to abstract relations like TIME, LOCATION and CONTAINER. We further apply it to (1) …

Content truncated to 500 words