Multiword Expression aware Neural Machine Translation

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3816–3825 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3816 Multiword Expression aware Neural Machine Translation Andrea Zaninello∗†, Alexandra Birch∗ ∗School of Informatics, University of Edinburgh, United Kingdom †Zanichelli editore, Bologna, Italy azaninello@zanichelli.it, a.birch@ed.ac.uk Abstract Multiword Expressions (MWEs) are a frequently occurring phenomenon found in all natural languages that is of great importance to linguistic theory, natural language processing applications, and machine translation systems. Neural Machine Translation (NMT) architectures do not handle these expressions well and previous studies have rarely addressed MWEs in this framework. In this work, we show that annotation and data augmentation, using external linguistic resources, can improve both translation of MWEs that occur in the source, and the generation of MWEs on the target, and increase performance by up to 5.09 BLEU points on MWE test sets. We also devise a MWE score to specifically assess the quality of MWE translation which agrees with human evaluation. We make available the MWE score implementation – along with MWE-annotated training sets and corpus-based lists of MWEs – for reproduction and extension. Keywords: multiword expressions, neural machine translation, evaluation 1. Introduction Multiword Expressions (MWEs) are a pervasive phe- nomenon in all natural languages to the point that, accord- ing to some studies, they represent approximately half of a language’s lexicon (Jackendoff, 1995). They also chal- lenge NLP applications because of their often unpredictable morpho-syntactic and lexico-semantic behaviour (Villavi- cencio et al., 2005). We call a MWE an expression that is composed of two or more words working as a unit with respect to some levels of linguistic analysis (Calzolari et al., 2002); a MWE displays idiosyncratic properties that cannot be explained solely on the basis of regular syntactic and semantic rules (Everaert et al., 2014) and is generally char- acterised by some degree of conventionality (Baldwin and Kim, 2010; Constant et al., 2017). In the last few years, Neural Machine Translation (NMT) has proved the best performing framework compared to previ- ous methodologies, with neural architectures producing ever more natural-sounding target language. Even so, NMT out- put is sometimes a poor translation of the source sentence (Nguyen and Chiang, 2018) and it is therefore important to investigate specific linguistic phenomena and improve trans- lation quality not only in terms of standard measurements. Previously dominant phrase-based and syntax-based Sta- tistical Machine Translation (SMT) techniques (Koehn et al., 2007; Junczys-Dowmunt et al., 2016) naturally take into account phrasal components, and there has been sig- nificant research on MWEs in these frameworks; however, for NMT, due to a lack of phrasal segmentation, it is less obvious how to address specific language phenomena such as MWEs. Moreover, while standard metrics are effective in terms of system comparison, their ability to account for more fine-grained improvements in MT is less straightfor- ward (Callison-Burch et al., 2006), and their effectiveness has been questioned. Therefore, evaluating the performance of NMT architectures in translating MWEs remains an open challenge. …

Content truncated to 500 words