MELODI: A Supervised Distributional Approach for Free Paraphrasing of Noun Compounds

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 144–147, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics MELODI: A Supervised Distributional Approach for Free Paraphrasing of Noun Compounds Tim Van de Cruys IRIT, CNRS tim.vandecruys@irit.fr Stergos Afantenos IRIT, Toulouse University stergos.afantenos@irit.fr Philippe Muller IRIT, Toulouse University philippe.muller@irit.fr Abstract This paper describes the system submitted by the MELODI team for the SemEval-2013 Task 4: Free Paraphrases of Noun Compounds (Hendrickx et al., 2013). Our approach com- bines the strength of an unsupervised distri- butional word space model with a supervised maximum-entropy classification model; the distributional model yields a feature represen- tation for a particular compound noun, which is subsequently used by the classifier to induce a number of appropriate paraphrases. 1 Introduction Interpretation of noun compounds is making explicit the relation between the component nouns, for in- stance that running shoes are shoes used in running activities, while leather shoes are made from leather. The relations can have very different meanings, and existing work either postulates a fixed set of rela- tions (Tratz and Hovy, 2010) or relies on appropri- ate descriptions of the relations, through constrained verbal paraphrases (Butnariu et al., 2010) or uncon- strained paraphrases as in the present campaign. The latter is much simpler for annotation purposes, but raises difficult challenges involving not only com- pound interpretation but also paraphrase evaluation and ranking. In terms of constrained verbal paraphrases Wubben (2010), for example, uses a supervised memory-based ranker using features from the Google n-gram corpus as well as WordNet. Nulty and Costello (2010) rank paraphrases of compounds according to the number of times they co-occurred with other paraphrases for other compounds. They use these co-occurrences to compute conditional probabilities estimating is-a relations between para- phrases. Li et al. (2010) provide a hybrid sys- tem which combines a Bayesian algorithm exploit- ing Google n-grams, a score which captures human preferences at the tail distribution of the training data, as well as a metric that captures pairwise para- phrase preferences. Our methodology consists of two steps. First, an unsupervised distributional word space model is constructed, which yields a feature representation for a particular compound. The feature representa- tion is then used by a maximum entropy classifier to induce a number of appropriate paraphrases. 2 Methodology 2.1 Distributional word space model In order to induce appropriate feature representa- tions for the various noun compounds, we start by constructing a standard distributional word space model for nouns. We construct a co-occurrence matrix of the 5K most frequent nouns1 by the 2K most frequent context words2, which occur in a win- dow of 5 words to the left and right of the target word. The bare frequencies of the word-context ma- trix are weighted using pointwise mutual informa- tion (Church and Hanks, 1990). Next, we compute a joint, compositional repre- sentation of the noun compound, combining the se- 1making sure all nouns that appear in the training and test …

Content truncated to 500 words