Neural Machine Translation into Language Varieties

Proceedings of the Third Conference on Machine Translation (WMT), Volume 1: Research Papers, pages 156–164 Belgium, Brussels, October 31 - Novermber 1, 2018. c⃝2018 Association for Computational Linguistics https://doi.org/10.18653/v1/W18-64016 Neural Machine Translation into Language Varieties Surafel M. Lakew†⋆, Aliia Erofeeva†, Marcello Federico⋆+ †University of Trento, ⋆Fondazione Bruno Kessler, +MMT Srl, Trento, Italy †name.surname@unitn.it, ⋆surname@fbk.eu Abstract Both research and commercial machine trans- lation have so far neglected the importance of properly handling the spelling, lexical and grammar divergences occurring among lan- guage varieties. Notable cases are standard national varieties such as Brazilian and Euro- pean Portuguese, and Canadian and European French, which popular online machine transla- tion services are not keeping distinct. We show that an evident side effect of modeling such va- rieties as unique classes is the generation of inconsistent translations. In this work, we in- vestigate the problem of training neural ma- chine translation from English to specific pairs of language varieties, assuming both labeled and unlabeled parallel texts, and low-resource conditions. We report experiments from En- glish to two pairs of dialects, European- Brazilian Portuguese and European-Canadian French, and two pairs of standardized vari- eties, Croatian-Serbian and Indonesian-Malay. We show significant BLEU score improve- ments over baseline systems when translation into similar languages is learned as a multilin- gual task with shared representations. 1 Introduction The field of machine translation (MT) is making amazing progress, thanks to the advent of neural models and deep learning. While just few years ago research in MT was struggling to achieve use- ful translations for the most requested and high- resourced languages, the level of translation qual- ity reached today has raised the demand and in- terest for less-resourced languages and the so- lution of more subtle and interesting translation tasks (Bentivogli et al., 2018). If the goal of machine translation is to help worldwide com- munication, then the time has come to also cope with dialects or more generally language vari- eties1. Remarkably, up to now, even standard na- tional language varieties, such as Brazilian and European Portuguese, or Canadian and European French, which are used by relatively large pop- ulations have been quite neglected both by re- search and industry. Prominent online commer- cial MT services, such as Google Translate and Bing, are currently not offering any variety of Por- tuguese and French. Even worse, systems offering such languages tend to produce inconsistent out- puts, like mixing lexical items from different Por- tuguese (see for instance the translations shown in Table 1). Clearly, in the perspective of delivering high-quality MT to professional post-editors and final users, this problem urges to be fixed. While machine translation from many to one varieties is intuitively simpler to approach2, it is the opposite direction that presents the most rel- evant problems. First, languages varieties such as dialects might significantly overlap thus mak- ing differences among their texts quite subtle (e.g., particular grammatical constructs or lexical diver- gences like the ones reported in the example). Sec- ond, parallel data are not always labeled at …

Content truncated to 500 words