Search Engine Statistics Beyond the n-Gram: Application to Noun Compound Bracketing

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 17–24, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Search Engine Statistics Beyond the n-gram: Application to Noun Compound Bracketing Preslav Nakov EECS, Computer Science Division University of California, Berkeley Berkeley, CA 94720 nakov@cs.berkeley.edu Marti Hearst SIMS University of California, Berkeley Berkeley, CA 94720 hearst@sims.berkeley.edu Abstract In order to achieve the long-range goal of semantic interpretation of noun com- pounds, it is often necessary to first de- termine their syntactic structure. This pa- per describes an unsupervised method for noun compound bracketing which extracts statistics from Web search engines using a χ2 measure, a new set of surface features, and paraphrases. On a gold standard, the system achieves results of 89.34% (base- line 66.80%), which is a sizable improve- ment over the state of the art (80.70%). 1 Introduction An important but understudied language analy- sis problem is that of noun compound bracketing, which is generally viewed as a necessary step to- wards noun compound interpretation. Consider the following contrastive pair of noun compounds: (1) liver cell antibody (2) liver cell line In example (1) an antibody targets a liver cell, while (2) refers to a cell line which is derived from the liver. In order to make these semantic distinctions accurately, it can be useful to begin with the cor- rect grouping of terms, since choosing a particular syntactic structure limits the options left for seman- tics. Although equivalent at the part of speech (POS) level, these two noun compounds have different syn- tactic trees. The distinction can be represented as a binary tree or, equivalently, as a binary bracketing: (1b) [ [ liver cell ] antibody ] (left bracketing) (2b) [ liver [cell line] ] (right bracketing) In this paper, we describe a highly accurate un- supervised method for making bracketing decisions for noun compounds (NCs). We improve on the cur- rent standard approach of using bigram estimates to compute adjacency and dependency scores by intro- ducing the use of the χ2 measure for this problem. We also introduce a new set of surface features for querying Web search engines which prove highly ef- fective. Finally, we experiment with paraphrases for improving prediction statistics. We have evaluated the application of combinations of these features to predict NC bracketing on two distinct collections, one consisting of terms drawn from encyclopedia text, and another drawn from bioscience text. The remainder of this paper describes related work, the word association models, the surface fea- tures, the paraphrase features and the results. 2 Related Work The syntax and semantics of NCs is an active area of research; the Journal of Computer Speech and Lan- guage has an upcoming special issue on Multiword Expressions. The best known early work on automated un- supervised NC bracketing is that of Lauer (1995) who introduces the probabilistic dependency model for the syntactic disambiguation of NCs and argues against the adjacency model, proposed by Marcus (1980), Pustejovsky et al. (1993) and Resnik (1993). Lauer …

Content truncated to 500 words