Similar Language Translation for Catalan, Portuguese and Spanish Using Marian NMT

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 292–298 November 10–11, 2021. ©2021 Association for Computational Linguistics 292 Abstract This paper describes the SEBAMAT contri- bution to the 2021 WMT Similar Language Translation shared task. Using the Marian neural machine translation toolkit, transla- tion systems based on Google’s trans- former architecture were built in both direc- tions of Catalan–Spanish and Portuguese– Spanish. The systems were trained in two contrastive parameter settings (different vo- cabulary sizes for byte pair encoding) using only the parallel but not the comparable corpora provided by the shared task organ- izers. According to their official evaluation results, the SEBAMAT system turned out to be competitive with rankings among the top teams and BLEU scores between 38 and 47 for the language pairs involving Portuguese and between 76 and 80 for the language pairs involving Catalan. 1 Introduction In recent years, neural machine translation (NMT) has become the state of the art in machine transla- tion (MT). Using toolkits such as Marian NMT (Junczys-Dowmunt et al., 2018), it is relatively straightforward to construct end-to-end NMT sys- tems which need only little pre-processing of the training corpora and post-processing of the system output. As NMT is a supervised approach to MT based on machine learning technology, training is usually conducted using sentence aligned human translations. Given a large number of source/target- language sentence pairs, the neural system fully au- tomatically learns how to translate. The SEBAMAT submission to the Similar Lan- guage Translation (SLT) task1 of the 6th Confer- ence on MT is based on work conducted as part of the SEBAMAT project2 (semantics-based machine 1 http://www.statmt. org/wmt21/similar.html 2 https://cordis.europa.eu/project/id/844951 translation; Rapp & Tambouratzis, 2020). This pro- ject has a focus on experiments introducing seman- tics into MT but, for comparative purposes, also deals with standard NMT systems. The latter were used as the basis for the current shared task. During the SEBAMAT project a number of MT systems had been developed for language pairs involving English, French, German, Greek and Spanish, but there had been no prior work on Catalan and Por- tuguese. The aims of the participation in the SLT  shared task were the following: See in how far the SEBAMAT-based MT sys- tems are competitive.  Extend the number of SEBAMAT languages by Catalan and Portuguese.  Find out whether systems for new language pairs can be developed in a very short time.  See whether reasonably well working systems can be developed without much proficiency of the respective languages on the developer side. 2 Resources 2.1 Corpora For the training of the NMT systems sentence- aligned parallel corpora are required. We used all parallel corpora suggested by the SLT shared task organizers who in their task description explicitly stated that no additional parallel corpora were al- lowed for training. For Catalan–Spanish the following parallel cor-  pora were used: Wiki Titles v3 (476,475 sentence pairs) (Bar- rault et al., 2020)3  ParaCrawl (6,870,183 sentence pairs) (Bañón et al., 2020)  DOGC …

Content truncated to 500 words