Low Resourced Machine Translation via Morpho-syntactic Modeling: The Case of Dialectal Arabic

Low Resourced Machine Translation via Morpho-syntactic Modeling: The Case of Dialectal Arabic Alexander Erdmann ae1541@nyu.edu Computational Approaches to Modeling Language (CAMeL) Lab, New York University Abu Dhabi Department of Linguistics, Ohio State University Nizar Habash nizar.habash@nyu.edu Computational Approaches to Modeling Language (CAMeL) Lab, New York University Abu Dhabi Dima Taji dima.taji@nyu.edu Computational Approaches to Modeling Language (CAMeL) Lab, New York University Abu Dhabi Houda Bouamor hbouamor@cmu.edu Department of Computer Science, Carnegie Mellon University Qatar Abstract We present the second ever evaluated Arabic dialect-to-dialect machine translation effort, and the first to leverage external resources beyond a small parallel corpus. The subject has not previously received serious attention due to lack of naturally occurring parallel data; yet its importance is evidenced by dialectal Arabic’s wide usage and breadth of inter-dialect variation, comparable to that of Romance languages. Our results suggest that modeling morphology and syntax significantly improves dialect-to-dialect translation, though optimizing such data-sparse models requires consideration of the linguistic differences between dialects and the nature of available data and resources. On a single-reference blind test set where untranslated input scores 6.5 BLEU and a model trained only on parallel data reaches 14.6, pivot techniques and morpho- syntactic modeling significantly improve performance to 17.5. 1 Introduction Arabic is widely spoken and highly diglossic, with Modern Standard Arabic (MSA) represent- ing the high register shared across the Arab World in educated circles. In contrast, the many spoken dialectal Arabic varieties (DA) are somewhat if not entirely mutually unintelligible, e.g., Moroccan and Kuwaiti. Chiang et al. (2006) compare the linguistic variation among Arabic di- alects to that among Romance languages, indicating the need for machine translation (MT) Proceedings of MT Summit XVI, vol.1: Research Track Nagoya, Sep. 18-22, 2017 | p. 185 between these dialects. However, while much MT research has been devoted to translating between Romance languages (Corbí Bellot et al., 2005; Armentano-Oller et al., 2006; Koehn et al., 2009), we are aware of only one work on Arabic DA-to-DA MT (Meftouh et al., 2015). It deals mainly with Maghrebi dialects and utilizes only a small parallel corpus.1 This work fo- cuses on the Egyptian and Levantine dialects, leveraging various available resources such as a morphological analyzer and additional monolingual and multilingual data.2 Compared to other dialects, Egyptian and Levantine’s wider range of available data/resources allows us to evaluate more MT approaches using different combinations of these data/resources. Thus, in future work on DA pairs which may not have the same data/resources, we can tailor MT systems based on this paper’s findings. The main challenge in developing DA-to-DA MT systems is the lack of data. While many Romance languages are official languages with written standards, naturally occurring in parallel corpora like the European Parliament (Koehn, 2005), DA has no official status and was rarely written until the advent of social media.3 The recent release of the first parallel multi-dialectal corpora (Bouamor et al., 2014; Meftouh et al., 2015) has enabled seminal, albeit low-resource MT experiments. We present some shortcomings of these corpora and introduce …

Content truncated to 500 words