UCD-Goggle: A Hybrid System for Noun Compound Paraphrasing

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 230–233, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics UCD-Goggle: A Hybrid System for Noun Compound Paraphrasing Guofu Li School of Computer Science and Informatics University College Dublin guofu.li@ucd.ie Alejandra Lopez-Fernandez School of Computer Science and Informatics University College Dublin alejandra.lopez -fernandez@ucd.ie Tony Veale School of Computer Science and Informatics University College Dublin tony.veale@ucd.ie Abstract This paper addresses the problem of rank- ing a list of paraphrases associated with a noun-noun compound as closely as possi- ble to human raters (Butnariu et al., 2010). UCD-Goggle tackles this task using se- mantic knowledge learnt from the Google n-grams together with human-preferences for paraphrases mined from training data. Empirical evaluation shows that UCD- Goggle achieves 0.432 Spearman correla- tion with human judgments. 1 Introduction Noun compounds (NC) are sequences of nouns acting as a single noun (Downing, 1977). Re- search on noun compounds involves two main tasks: NC detection and NC interpretation. The latter has been studied in the context of many natural language applications, including question- answering, machine translation, information re- trieval, and information extraction. The use of multiple paraphrases as a semantic intepretation of noun compounds has recently be- come popular (Kim and Baldwin, 2006; Nakov and Hearst, 2006; Butnariu and Veale, 2008; Nakov, 2008). The best paraphrases are those which most aptly characterize the relationship be- tween the modifier noun and the head noun. The aim of this current work is to provide a ranking for a list of paraphrases that best approxi- mates human rankings for the same paraphrases. We have created a system called UCD-Goggle, which uses semantic knowledge acquired from Google n-grams together with human-preferences mined from training data. Three major com- ponents are involved in our system: B-score, produced by a Bayesian algorithm using seman- tic knowledge from the n-grams corpus with a smoothing layer of additional inference; Rt-score captures human preferences observed in the tail distribution of training data; and Rp-score cap- tures pairwise paraphrase preferences calculated from the training data. Our best system for SemEval-2 task 9 combines all three components and achieves a Spearman correlation of 0.432 with human rankings. This paper is organized as follows: the Bayesian B-score is introduced in section 2. In section 3 we describe two supervised approaches to mining the preferences of human raters from training data. Finally, section 4 presents the results of our empir- ical evaluation of the UCD-Goggle system. 2 Semantic Approach 2.1 Collecting Data Google have made their web n-grams, also known as Web-1T corpus, public via the Linguistic Data Consortium (Brants and Franz, 2006). This cor- pus contains sequences of n terms that occur more than 40 times on the web. We view the paraphrase task as that of suggest- ing the right verb phrase for two nouns (But- nariu and Veale, 2008). Previous work has shown the n-grams corpus to be a promising resource for retrieving semantic evidence for this approach. However, the corpus itself needs to …

Content truncated to 500 words