Neural Machine Translation of Basque

Neural Machine Translation of Basque Thierry Etchegoyhen,1 Eva Mart´ınez Garcia,1 Andoni Azpeitia,1 Gorka Labaka,2 I˜naki Alegria,2 Itziar Cortes Etxabe,3 Amaia Jauregi Carrera,3 Igor Ellakuria Santos,4 Maite Martin,5 Eusebi Calonge5 1Vicomtech - {tetchegoyhen, emartinez, aazpeitia}@vicomtech.org 2IXA taldea, University of the Basque Country - {i.alegria, gorka.labaka}@ehu.eus 3Elhuyar - {i.cortes, a.jauregi}@elhuyar.eus 4ISEA - isantos@iseamcc.net 5Ametzagai˜na - maite@adur.com, ecalonge@ametza.com Abstract We describe the first experimental results in neural machine translation for Basque. As a synthetic language featuring aggluti- native morphology, an extended case sys- tem, complex verbal morphology and rel- atively free word order, Basque presents a large number of challenging characteristics for machine translation in general, and for data-driven approaches such as attention- based encoder-decoder models in partic- ular. We present our results on a large range of experiments in Basque-Spanish translation, comparing several neural ma- chine translation system variants with both rule-based and statistical machine transla- tion systems. We demonstrate that signif- icant gains can be obtained with a neu- ral network approach for this challeng- ing language pair, and describe optimal configurations in terms of word segmen- tation and decoding parameters, measured against test sets that feature multiple refer- ences to account for word order variability. 1 Introduction Neural machine translation (NMT) is fast becom- ing the dominant paradigm in both academic re- search and commercial exploitation, as evidenced in particular by large machine translation providers turning to NMT for their production engines (Crego et al., 2016; Wu et al., 2016) and NMT systems achieving the best results in most cases on stan- dard shared tasks datasets (Bojar, 2016). c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. Sequence-to-sequence neural networks have proved effective in modelling translation phenom- ena (Sutskever et al., 2014). In particular, atten- tional encoder-decoder models (Bahdanau et al., 2015) have become a default NMT architecture, with other architectural variants explored in recent work (Vaswani et al., 2017; Gehring et al., 2016). These models have already been applied to a wide range of languages, initially on the most studied European languages and recently to a larger set of cases that includes morphologically rich languages (Bojar, 2017). In this article we explore the applicability of neural machine translation to Basque, a language with noteworthy characteristics that may represent a challenge for encoder-decoder approaches with attention mechanisms. First, Basque is a synthetic language that fea- tures agglutinative morphology, i.e. where words can be formed via morphemic sequences, and a large number of case affixes that mark ergativity, datives, different types of locatives and genitives, instrumentality, comitativity or causality, among others. Verbal morphology is also relatively rare, displaying complex forms that include subject, di- rect object, indirect object and allocutive agree- ment markers, with number, tense and aspect be- ing marked as well. This kind of rich morphol- ogy raises difficulties in terms of word representa- tions and drastically increases data sparseness is- sues. A detailed description of Basque grammar can be found in (De Rijk and De Coene, 2008). …

Content truncated to 500 words