Detecting Non-compositional MWE Components using Wiktionary

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1792–1797, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Detecting Non-compositional MWE Components using Wiktionary Bahar Salehi,♣♥ Paul Cook♥ and Timothy Baldwin♣♥ ♣ NICTA Victoria Research Laboratory ♥ Department of Computing and Information Systems The University of Melbourne Victoria 3010, Australia bsalehi@student.unimelb.edu.au, paulcook@unimelb.edu.au, tb@ldwin.net Abstract We propose a simple unsupervised ap- proach to detecting non-compositional components in multiword expressions based on Wiktionary. The approach makes use of the definitions, synonyms and trans- lations in Wiktionary, and is applicable to any type of MWE in any language, assum- ing the MWE is contained in Wiktionary. Our experiments show that the proposed approach achieves higher F-score than state-of-the-art methods. 1 Introduction A multiword expression (MWE) is a combina- tion of words with lexical, syntactic or seman- tic idiosyncrasy (Sag et al., 2002; Baldwin and Kim, 2009). An MWE is considered (semanti- cally) “non-compositional” when its meaning is not predictable from the meaning of its compo- nents. Conversely, compositional MWEs are those whose meaning is predictable from the meaning of the components. Based on this definition, a component is compositional within an MWE, if its meaning is reflected in the meaning of the MWE, and it is non-compositional otherwise. Understanding which components are non- compositional within an MWE is important in NLP applications in which semantic information is required. For example, when searching for spelling bee, we may also be interested in docu- ments about spelling, but not those which contain only bee. For research project, on the other hand, we are likely to be interested in documents which contain either research or project in isolation, and for swan song, we are only going to be interested in documents which contain the phrase swan song, and not just swan or song. In this paper, we propose an unsupervised ap- proach based on Wikitionary for predicting which components of a given MWE have a composi- tional usage. Experiments over two widely-used datasets show that our approach outperforms state- of-the-art methods. 2 Related Work Previous studies which have considered MWE compositionality have focused on either the iden- tification of non-compositional MWE token in- stances (Kim and Baldwin, 2007; Fazly et al., 2009; Forthergill and Baldwin, 2011; Muzny and Zettlemoyer, 2013), or the prediction of the com- positionality of MWE types (Reddy et al., 2011; Salehi and Cook, 2013; Salehi et al., 2014). The identification of non-compositional MWE tokens is an important task when a word combination such as kick the bucket or saw logs is ambiguous between a compositional (generally non-MWE) and non-compositional MWE usage. Approaches have ranged from the unsupervised learning of type-level preferences (Fazly et al., 2009) to su- pervised methods specific to particular MWE con- structions (Kim and Baldwin, 2007) or applica- ble across multiple constructions using features similar to those used in all-words word sense disambiguation (Forthergill and Baldwin, 2011; Muzny and Zettlemoyer, 2013). The prediction of the compositionality of MWE types has tradi- tionally …

Content truncated to 500 words