Verbal Multiword Expression Identification: Do We Need a Sledgehammer to Crack a Nut?

Proceedings of the 28th International Conference on Computational Linguistics, pages 3333–3345 Barcelona, Spain (Online), December 8-13, 2020 3333 Verbal Multiword Expression Identification: Do We Need a Sledgehammer to Crack a Nut? Caroline Pasquer University of Tours, LIFAT France first.last@etu.univ-tours.fr Agata Savary University of Tours, LIFAT France first.last@univ-tours.fr Carlos Ramisch Aix Marseille Univ, Université de Toulon, CNRS, LIS, Marseille, France first.last@lis-lab.fr Jean-Yves Antoine University of Tours, LIFAT France first.last@univ-tours.fr Abstract Automatic identification of multiword expressions (MWEs), like to cut corners ‘to do an in- complete job’, is a pre-requisite for semantically-oriented downstream applications. This task is challenging because MWEs, especially verbal ones (VMWEs), exhibit surface variability. This paper deals with a subproblem of VMWE identification: the identification of occurrences of previously seen VMWEs. A simple language-independent system based on a combination of filters competes with the best systems from a recent shared task: it obtains the best averaged F-score over 11 languages (0.6653) and even the best score for both seen and unseen VMWEs due to the high proportion of seen VMWEs in texts. This highlights the fact that focusing on the identification of seen VMWEs could be a strategy to improve VMWE identification in general. 1 Introduction Multiword expressions (MWEs) are word combinations idiosyncratic with respect to e.g. syntax or semantics (Baldwin and Kim, 2010). One of their most emblematic properties is semantic non- compositionality: the meaning of the whole cannot be straightforwardly deduced from the meanings of its components, as in cut corners ‘do an incomplete job’.1 Due to this property and to their frequency (Jackendoff, 1997), MWEs are a major challenge for semantically-oriented downstream applications, such as machine translation. A prerequisite for an MWE processing is their automatic identification. MWE identification aims at locating MWE occurrences in running text. This task is very challenging, as signaled by Constant et al. (2017), and further confirmed by the PARSEME shared task on automatic identification of verbal MWEs (Ramisch et al., 2018). One of the main difficulties stems from the variability of MWEs, especially verbal ones (VMWEs). That is, even if a VMWE has previously been observed in a training corpus or in a lexicon, it can re-appear in morphosyntactically diverse forms. Examples (1–2) show two occurrences of a VMWE with variation in the components’ inflection (cutting vs. cut), word order, presence of discontinuities (were), and syntactic relations (obj vs. nsubj). (1) Some companies were cutting cornersobj to save costs. (2) The field would look uneven if cornersnsubj were cut. However, unrestricted variability is not a reasonable assumption either, since it may lead to literal or coincidental occurrences of VMWEs’ components (Savary et al., 2019b), as in (3) and (4), respectively.2 (3) Start with :::::: cutting one :::::: corner of the disinfectant bag. (4) If you ::: cut along these lines, you’ll get two acute ::::::: corners. This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/. 1Henceforth, the lexicalized components of a MWE, i.e. those always realized by the same lexemes, appear in bold. …

Content truncated to 500 words