TAN-IBE: Neural Machine Translation for the romance languages of the Iberian Peninsula

TAN-IBE: Neural Machine Translation for the romance languages of the Iberian Peninsula Antoni Oliver, Merc`e V`azquez, Marta Coll-Florit, Sergi ´Alvarez, V´ıctor Su´arez, Claudi Avent´ın-Boya Universitat Oberta de Catalunya {aoliverg,mvazquega,mcollfl,salvarezvid,vsuarezpi,caventinb}@uoc.edu Cristina Vald´es Universidad de Oviedo cris@uniovi.es Mar Font Universitat de Lleida mar.font@udl.cat Alejandro Pardos Universidad de Zaragoza apardoscalvo@gmail.com Abstract This paper describes the project TAN-IBE: Neural Machine Translation for the ro- mance languages of the Iberian Peninsula, a three year research project founded by the Spanish Ministry of Science and In- novation in the call Proyectos de gen- eraci´on de conocimiento 2021 (Reference: PID2021-124663OB-I00). This project has started in September 2022. 1 Introduction The main goal of this project is to explore the techniques for training NMT systems applied to Spanish, Portuguese, Catalan, Galician, Asturian, Aragonese and Aranese, a standardized subvari- ety of Gascon, which is a variety of Occitan. Aranese has the status of official language in the autonomous community of Catalonia. These languages belong to the same Romance family, but they are very different in terms of the lin- guistic resources available. Asturian, Aragonese and Aranese can be considered low-resource lan- guages. These characteristics make this setting an excellent place to explore training techniques for low-resource languages: transfer learning and multilingual systems, among others. The first months of the project have been dedi- cated to the compilation of monolingual and paral- lel corpora for Asturian, Aragonese and Aranese. 2 List of partners • Universitat Oberta de Catalunya1 (UOC), lead- ing the project and in charge of the technical © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. 1https://www.uoc.edu aspects regarding the training of the neural sys- tems. • Universidad de Oviedo2, working in the compi- lation of corpora for Asturian. • Universidad de Zaragoza3, in charge of the com- pilation of resources for Aragonese. • Universitat de Lleida4 (UdL), working in the compilation of texts for Aranese. 3 Project objectives The project’s main objective is to design, train and evaluate NMT systems between the Romance lan- guages of the Iberian Peninsula. This objective will be achieved through the following specific ob- jectives: • Compiling parallel and monolingual cor- pora for the languages included in the pro- posal, paying special attention to Asturian, Aragonese and Aranese. • Exploring new techniques for training neural machine translation engines. • Train neural machine translation systems be- tween Spanish and the rest of the languages of the project, in both directions. • Training neural multilingual systems capable to translate from and to all the languages of the project. • Evaluating all the trained systems using au- tomatic evaluation metrics and compare them with existing machine translation systems. • Performing manual evaluations of the ma- chine translation systems developed for Span- ish to Asturian, Aragonese and Aranese. • Creating guides and scripts that facilitate the training of neural machine translation en- 2https://www.uniovi.es/ 3https://www.unizar.es/ 4https://www.udl.cat/ca/ gines. • Publishing the results of TAN-IBE with open licences. 4 Summary of partial results The project started on September …

Content truncated to 500 words