Using Distributional Similarity of Multi-way Translations to Predict Multiword Expression Compositionality

Abstract

We predict the compositionality of multiword expressions using distributional similarity between each component word and the overall expression, based on translations into multiple languages. We evaluate the method over English noun compounds, English verb particle constructions and German noun compounds. We show that the estimation of compositionality is improved when using translations into multiple languages, as compared to simply using distributional similarity in the source language. We further find that string similarity complements distributional similarity. 1 Compositionality of MWEs Multiword expressions (hereafter MWEs) are combinations of words which are lexically, syntactically, semantically or statistically idiosyncratic (Sag et al., 2002; Baldwin and Kim, 2009). Much research has been carried out on the extraction and identification of MWEs1 in English (Schone and Jurafsky, 2001; Pecina, 2008; Fazly et al., 2009) and other languages (Dias, 2003; Evert and Krenn, 2005; Salehi et al., 2012). However, considerably less work has addressed the task of predicting the meaning of MWEs, especially in non-English languages. As a step in this direction, the focus of this study is on predicting the compositionality of MWEs. An MWE is fully compositional if its meaning is predictable from its component words, and it is non-compositional (or idiomatic) if not. For example, stand up “rise to one’s feet” is composiIn this paper, we follow Baldwin and Kim (2009) in considering MWE “identification” to be a token-level disambiguation task, and MWE “extraction” to be a type-level lexicon induction task. tional, because its meaning is clear from the meaning of the components stand and up. However, the meaning of strike up “to start playing” is largely unpredictable from the component words strike and up. In this study, following McCarthy et al. (2003) and Reddy et al. (2011), we consider compositionality to be graded, and aim to predict the degree of compositionality. For example, in the dataset of Reddy et al. (2011), climate change is judged to be 99% compositional, while silver screen is 48% compositional and ivory tower is 9% compositional. Formally, we model compositionality prediction as a regression task. An explicit handling of MWEs has been shown to be useful in NLP applications (Ramisch, 2012). As an example, Carpuat and Diab (2010) proposed two strategies for integrating MWEs into statistical machine translation. They show that even a large scale bilingual corpus cannot capture all the necessary information to translate MWEs, and that in adding the facility to model the compositionality of MWEs into their system, they could improve translation quality. Acosta et al. (2011) showed that treating non-compositional MWEs as a single unit in information retrieval improves retrieval effectiveness. For example, while searching for documents related to ivory tower, we are almost certainly not interested in documents relating to elephant tusks. Our approach is to use a large-scale multi-way translation lexicon to source translations of MWEs and their component words, and then model the relative similarity between each of the component words and the MWE, using distributional similarity based on monolingual corpora for the source language and each of the target languages. Our hypothesis is that using distributional similarity in more than one language will improve the prediction of compositionality. Importantly, in order to make the method as language-independent and