Unsupervised Machine Translation in Real-World Scenarios

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3038–3047 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3038 Unsupervised Machine Translation in Real-World Scenarios Ona de Gibert1, Iakes Goenaga2, Jordi Armengol-Estap´e1, Olatz Perez-de-Vi˜naspre2 Carla Parra3, Marina S´anchez-Torr´on4, Marcis Pinnis5, Gorka Labaka2, Maite Melero1 Barcelona Supercomputing Center, Barcelona, Spain1 HiTZ zentroa - Ixa, Euskal Herriko Unibertsitatea UPV/EHU, Donostia, Spain2 RWS Language Weaver, Dublin, Ireland3 Unbabel, Lisbon, Portugal4 Tilde, Riga, Latvia5 {ona.degibert, jordi.armengol, maite.melero}@bsc.es1 {iakes.goenaga, olatz.perezdevinaspre, gorka.labaka}@ehu.eus2 CParraEscartin@rws.com3, marina.sanchez@unbabel.com4, marcis.pinnis@tilde.lv5 Abstract In this work, we present the work that has been carried on in the MT4All CEF project and the resources that it has generated by leveraging recent research carried out in the field of unsupervised learning. In the course of the project 18 monolingual corpora for specific domains and languages have been collected, and 12 bilingual dictionaries and translation models have been generated. As part of the research, the unsupervised MT methodology based only on monolingual corpora (Artetxe et al., 2017) has been tested on a variety of languages and domains. Results show that in specialised domains, when there is enough monolingual in-domain data, unsupervised results are comparable to those of general domain super- vised translation, and that, at any rate, unsupervised techniques can be used to boost results whenever very little data is available. Keywords: Machine Translation, Unsupervised, Neural Networks, Low-resource languages 1. Introduction The foundation of Machine Translation (MT) largely depends on the availability of large parallel corpus of the targeted language pair. Nevertheless, large amounts of bilingual corpora are not always available, specially for specific domains or low-resource languages. In re- cent years, to address this issue, unsupervised tech- niques have appeared which rely solely or partially on monolingual corpora to build MT systems (Artetxe et al., 2019; Lample et al., 2017; Artetxe et al., 2017). In this scenario, this work aims at generating bilingual resources for those language pairs and domains lacking sufficient parallel corpora by leveraging recent research carried out in the field of unsupervised learning; deriv- ing bilingual dictionaries and translation models from large amounts of monolingual corpora only. Partners include two research centers: the IXA group from the University of the Basque Country and the Barcelona Supercomputing Center; and three companies provid- ing translation services: Tilde, Iconic and Unbabel. The contributions of the project are publicly available1 and can be summarized as follows: • Release of 18 monolingual corpora for specific domains and a variety of languages in which both low-resource and high-resource languages are found. • Release of bilingual dictionaries, word embed- dings and translation models. • Expansion of unsupervised techniques (Artetxe et 1https://elrc-share.eu/repository/search/?q= MT4all al., 2019) to further domains and languages, in- cluding some low-resourced languages. The remaining of this article is organized as follows. In Section 2, we summarize the relevant related work. In Section 3, we describe the collected corpora and the experiments. Then, in Section 4, we go through the re- sults of the experiments. Finally, in Section 5, …

Content truncated to 500 words