Multilingual Neural Machine Translation: Case-study for Catalan, Spanish and Portuguese Romance Languages

Proceedings of the 5th Conference on Machine Translation (WMT), pages 447–450 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 447 Multilingual Neural Machine Translation: Case-study for Catalan, Spanish and Portuguese Romance Languages Pere Verg´es Boncompte and Marta R. Costa-juss`a TALP Research Center Universitat Polit`ecnica de Catalunya, Barcelona pere.verges@est.fib.upc.edu, marta.ruiz@upc.edu Abstract In this paper, we describe the TALP-UPC participation in the WMT Similar Language Translation task between Catalan, Spanish, and Portuguese, all of them, Romance lan- guages. We made use of different tech- niques to improve the translation between these languages. The multilingual shared en- coder/decoder has been used for all of them. Additionally, we applied back-translation to take advantage of the monolingual data. Fi- nally, we have applied fine-tuning to improve the in-domain data. Each of these techniques brings improvements over the previous one. In the official evaluation, our system was ranked 1st in the Portuguese-to-Spanish direc- tion, 2nd in the opposite direction, and 3rd in the Catalan-Spanish pair. 1 Introduction Research in the field of Machine Translation (MT) has been growing during these last years. From statistical approaches (Koehn et al., 2003) to neu- ral ones (Bahdanau et al., 2015), the progress has been impressive. Even after having achieved excep- tional results based only on attention mechanisms (Vaswani et al., 2017), there are still many chal- lenges and improvements remaining, for instance, multilingual translation from languages other than English, which have lower resources, and domain adaptation. In order to tackle these challenges, the Similar Language Task organized in the context of the Con- ference on Machine Translation (WMT 2020) has provided an appropriate setting for them. Within this task, the focus is the translation between lan- guages that are different from English, and more specifically, the focus consists of translating lan- guages that are from the same family. The families included are the following: South-Slavic, Indo- Aryan, and Romance. In our case, we have devoted the research to Romance languages, which include Spanish, Por- tuguese, and Catalan. The evaluation comprised all translation directions, but only provided parallel training data for Spanish-Portuguese and Spanish- Catalan. We approached the Portuguese-Catalan pair both from a pivot-based and zero-shot perspec- tive. In this paper, we make use of the well-known multilingual shared encoder/decoder and we show its effectiveness when applied to languages of the same linguistic family. Additionally, we benefited from back-translation and fine-tuning. 2 Background In this section, we show an overview of neural- based multilingual machine translation and domain adaptation using fine-tuning. 2.1 Multilingual translation When having multiple languages, there is the oppor- tunity to use several NMT architectures, based in the Transformer (Vaswani et al., 2017). Among the alternatives, we can share encoders and decoders (Johnson et al., 2017) or have specific encoders and decoders for each language (Escolano et al., 2020). In this paper, we are using the shared approach and we are leaving as further work to compare with other ones. Shared encoder-decoder One direct approach is using a single encoder/decoder shared for all languages (Johnson et …

Content truncated to 500 words