Determining the Multiword Expression Inventory of a Surprise Language

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 471–481, Osaka, Japan, December 11-17 2016. Determining the Multiword Expression Inventory of a Surprise Language Bahar Salehi,1 Paul Cook2 and Timothy Baldwin1 1 NICTA Victoria Research Laboratory Department of Computing and Information Systems The University of Melbourne, Victoria 3010, Australia 2 Faculty of Computer Science University of New Brunswick Fredericton, NB E3B 5A3, Canada bsalehi@student.unimelb.edu.au, paul.cook@unb.ca, tb@ldwin.net Abstract Much previous research on multiword expressions (MWEs) has focused on the token- and type- level tasks of MWE identification and extraction, respectively. Such studies typically target known prevalent MWE types in a given language. This paper describes the first attempt to learn the MWE inventory of a “surprise” language for which we have no explicit prior knowledge of MWE patterns, certainly no annotated MWE data, and not even a parallel corpus. Our proposed model is trained on a treebank with MWE relations of a source language, and can be applied to the monolingual corpus of the surprise language to identify its MWE construction types. 1 Introduction Multiword expressions (“MWEs”) are word combinations which have idiosyncratic properties relative to their component words (Sag et al., 2002; Baldwin and Kim, 2010), such as taken aback or red tape. The need for an explicit model of MWEs has been shown to be important in NLP tasks including ma- chine translation (Venkatapathy and Joshi, 2006), parsing (Constant et al., 2012), and keyphrase/index term extraction (Newman et al., 2012). However, existing approaches to MWE identification/extraction typically target specific MWE types that are known to be prevalent in a given language, such as: (a) com- pound nouns in languages such as English (Copestake, 2003; ´O S´eaghdha, 2008), German (Schulte im Walde et al., 2013) and Japanese (Tanaka and Baldwin, 2003); (b) light verb constructions (LVCs) in languages such as English (Butt, 2003), Persian (Karimi-Doostan, 1997) and Italian (Alba-Salas, 2002); and (c) compound verbs in languages such as Japanese (Uchiyama et al., 2005). Note here that the combination of highly-productive MWE types can vary greatly across languages: English is rich with compound nouns and LVCs are also common, but lacks compound verbs; Persian is rich with LVCs and adjective–noun compounds, but has very few compound nouns and compound verbs; and Japanese is rich with LVCs and compound nouns and verbs, but adjective–noun MWEs are rarer. Even for collocation extraction, this knowledge is generally assumed for a given language, in targeting only highly productive constructions such as adjective–noun or verb–noun collocations (Krenn and Evert, 2001; Pecina, 2008). But what if the language of interest is one where no such prior knowledge exists, e.g. because it is a “surprise” language where rapid deployment of language technologies is required and there is no access to an informant with sufficient linguistic training to be able to inventorise the MWE types in the language (Oard, 2003; Maynard et al., 2003)? Here, there is little expectation of success without an automatic method for determining the inventory and relative frequency of MWEs …

Content truncated to 500 words