Translating Similar Languages: Role of Mutual Intelligibility in Multilingual Transformers

Proceedings of the 5th Conference on Machine Translation (WMT), pages 381–386 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 381 Translating Similar Languages: Role of Mutual Intelligibility in Multilingual Transformers Ife Adebara El Moatez Billah Nagoudi Muhammad Abdul Mageed Natural Language Processing Lab University of British Columbia {ife.adebara,moatez.nagoudi,muhammad.mageed}@ubc.ca Abstract We investigate different approaches to trans- late between similar languages under low re- source conditions, as part of our contribution to the WMT 2020 Similar Languages Transla- tion Shared Task. We submitted Transformer- based bilingual and multilingual systems for all language pairs, in the two directions. We also leverage back-translation for one of the language pairs, acquiring an improvement of more than 3 BLEU points. We interpret our results in light of the degree of mutual intelli- gibility (based on Jaccard similarity) between each pair, finding a positive correlation be- tween mutual intelligibility and model perfor- mance. Our Spanish-Catalan model has the best performance of all the five language pairs. Except for the case of Hindi-Marathi, our bilin- gual models achieve better performance than the multilingual models on all pairs. 1 Introduction We present our findings from our participation in the WMT 2020 Similar Language Translation shared task, which focused on translation between similar language pairs in low-resource settings. Similar languages share a certain level of mutual in- telligibility that may aid the improvement of trans- lation quality. Depending on the level of closeness, certain languages may share similar orthography, lexical, syntactic, and or semantic structures which may make translation more accurate. The level of mutual intelligibility is such that speakers of one language can understand another language without prior instruction in that other lan- guage. They can also communicate without the use of a lingua franca which is a link or vehicular language used for communicating between speak- ers of different languages (Gooskens, 2007). It is important to mention that, sometimes, the level of intelligibility varies in both directions. For in- stance, Slovene - Croatian intelligibility is said to be asymmetric such that speakers or Slovene can understand spoken and written Croatian better than speakers of Croatian understand Slovene (Gol- ubovi´c and Gooskens, 2015). Machine translation of similar languages has been explored in a number of works (Hajic, 2000; Currey et al., 2016; Dabre et al., 2017). This can be seen as part of a growing need to develop mod- els that translate well in low resource scenarios. The goal of the current shared task is to encourage researchers to explore methods for translating be- tween similar languages. We also view the shared task as useful context for studying interaction be- tween degrees of similarity and mutual intelligibil- ity on the one hand, and model performance on the other hand. We explore the use of bilingual and multilingual models for all the 5 shared task language pairs. We also perform back-translation for one language pair. In the remainder of this paper, we discuss related literature in Section 2. We explain the methodology which includes a description of the Transformer …

Content truncated to 500 words