Predicting the Compositionality of Multiword Expressions Using Translations in Multiple Languages

In this paper, we propose a simple, languageindependent and highly effective method for predicting the degree of compositionality of multiword expressions (MWEs). We compare the translations of an MWE with the translations of its components, using a range of different languages and string similarity measures. We demonstrate the effectiveness of the method on two types of English MWEs: noun compounds and verb particle constructions. The results show that our approach is competitive with or superior to state-of-the-art methods over standard datasets. 1 Compositionality of MWEs A multiword expression (MWE) is any combination of words with lexical, syntactic or semantic idiosyncrasy (Sag et al., 2002; Baldwin and Kim, 2009), in that the properties of the MWE are not predictable from the component words. For example, with ad hoc, the fact that neither ad nor hoc are standalone English words, makes ad hoc a lexicallyidiosyncratic MWE; with shoot the breeze, on the other hand, we have semantic idiosyncrasy, as the meaning of “to chat” in usages such as It was good to shoot the breeze with you cannot be predicted from the meanings of the component words shoot and breeze. Semantic idiosyncrasy has been of particular interest to NLP researchers, with research on binary compositional/non-compositional MWE clasThe example is taken from http://www. thefreedictionary.com sification (Lin, 1999; Baldwin et al., 2003), or a three-way compositional/semi-compositional/noncompositional distinction (Fazly and Stevenson, 2007). There has also been research to suggest that MWEs span the entire continuum from full compositionality to full non-compositionality (McCarthy et al., 2003; Reddy et al., 2011). Investigating the degree of MWE compositionality has been shown to have applications in information retrieval and machine translation (Acosta et al., 2011; Venkatapathy and Joshi, 2006). As an example of an information retrieval system, if we were looking for documents relating to rat race (meaning “an exhausting routine that leaves no time for relaxation”), we would not be interested in documents on rodents. These results underline the need for methods for broad-coverage MWE compositionality prediction. In this research, we investigate the possibility of using an MWE’s translations in multiple languages to measure the degree of the MWE’s compositionality, and investigate how literal the semantics of each component is within the MWE. We use Panlex to translate the MWE and its components, and compare the translations of the MWE with the translations of its components using string similarity measures. The greater the string similarity, the more compositional the MWE is. Whereas past research on MWE compositionality has tended to be tailored to a specific MWE type (McCarthy et al., 2007; Kim and Baldwin, 2007; Fazly et al., 2009), our method is applicable to any MWE type in any language. Our experiments This definition is from WordNet 3.1.