Neural Machine Translation between similar South-Slavic languages

Proceedings of the 5th Conference on Machine Translation (WMT), pages 430–436 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 430 Neural Machine Translation between similar South-Slavic languages Maja Popovi´c, Alberto Poncelas ADAPT Centre, School of Computing Dublin City University, Ireland name.surname@adaptcentre.ie Abstract This paper describes the ADAPT-DCU ma- chine translation systems built for the WMT 2020 shared task on Similar Language Trans- lation. We explored several set-ups for NMT for Croatian–Slovenian and Serbian– Slovenian language pairs in both translation directions. Our experiments focus on differ- ent amounts and types of training data: we first apply basic filtering on the OpenSubti- tles training corpora, then we perform addi- tional cleaning of remaining misaligned seg- ments based on character n-gram matching. Finally, we make use of additional monolin- gual data by creating synthetic parallel data through back-translation. Automatic evalua- tion shows that multilingual systems with joint Serbian and Croatian data are better than bilin- gual, as well as that character-based cleaning leads to improved scores while using less data. The results also confirm once more that adding back-translated data further improves the per- formance, especially when the synthetic data is similar to the desired domain of the devel- opment and test set. This, however, might come at a price of prolonged training time, es- pecially for multitarget systems. 1 Introduction Machine translation (MT) between closely related languages is, in principle, less challenging than translation between distantly related languages, but it is still far from being solved. While MT be- tween closely related South-Western Slavic lan- guages, Croatian, Slovenian and Serbian based on the rule-based (RBMT) and the phrase-based (PB- SMT) approaches has been investigated in the last years (Etchegoyhen et al., 2014; Petkovski et al., 2014; Klubiˇcka et al., 2016; Arˇcan et al., 2016; Popovi´c et al., 2016a), to the best of our knowledge, the new state-of-the-art neural machine translation (NMT) has not been investigated yet for these lan- guages. In this work, we first compare bilingual and mul- tilingual systems in order to determine whether joining Serbian and Croatian data is useful. After- wards, we investigate additional cleaning of remain- ing misaligned segments by using character n-gram matching scores (Popovi´c, 2015). The beauty of the method for similar languages is that it can be ap- plied directly to the given training corpus providing matching scores for each pair of the source-target segments. For distant languages, translation of one side of the training corpus would be required. Fi- nally, we make use of monolingual data in each of the three languages by creating additional synthetic parallel training sets via back-translation (Sennrich et al., 2016a; Poncelas et al., 2018; Burlot and Yvon, 2018). 2 Language properties Common properties All three languages, Croa- tian, Serbian and Slovenian, belong to the South- Western Slavic branch. As Slavic languages, they have a very rich inflectional morphology for all word classes: six cases and three genders for all nouns, pronouns, adjectives and determiners. For verbs, person and many tenses are expressed by the suffix so that …

Content truncated to 500 words