A multi-source approach for Breton–French hybrid machine translation

A multi-source approach for Breton–French hybrid machine translation V´ıctor M. S´anchez-Cartagena, Mikel L. Forcada, Felipe S´anchez-Mart´ınez Dep. de Llenguatges i Sistemes Inform`atics, Universitat d’Alacant E-03690 Sant Vicent del Raspeig (Spain) {vmsanchez,mlf,fsanchez}@dlsi.ua.es Abstract Corpus-based approaches to machine trans- lation (MT) have difficulties when the amount of parallel corpora to use for train- ing is scarce, especially if the languages involved in the translation are highly in- flected. This problem can be addressed from different perspectives, including data augmentation, transfer learning, and the use of additional resources, such as those used in rule-based MT (RBMT). This paper fo- cuses on the hybridisation of RBMT and neural MT (NMT) for the Breton–French under-resourced language pair in an attempt to study to what extent the RBMT resources help improve the translation quality of the NMT system. We combine both translation approaches in a multi-source NMT archi- tecture and find out that, even though the RBMT system has a low performance ac- cording to automatic evaluation metrics, us- ing it leads to improved translation quality. 1 Introduction Corpus-based approaches to machine translation (MT), such as neural MT (NMT), struggle when the size of the available parallel corpora for a given language pair is scarce (Koehn and Knowles, 2017). Even though the problem can be partially mitigated with accurate hyper-parameter tuning (Sennrich and Zhang, 2019), taking advantage of additional re- sources can help to further improve the quality of the system. Monolingual texts in both languages can be lever- aged with the help of back-translation (Sennrich et al., 2016a; Hoang et al., 2018) to generate synthetic c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. parallel corpora. It is also possible to use only monolingual corpora and follow an unsupervised NMT approach (Artetxe et al., 2018). Parallel cor- pora from related language pairs can also be lever- aged thanks to multilingual NMT (Johnson et al., 2017) and other forms of transfer learning (Kocmi and Bojar, 2018). In addition to the use of corpora, linguistic re- sources can also be used to improve NMT. If mor- phological analysers or syntactic parsers are avail- able, they can be used to build a richer represen- tation of the words being translated (Sennrich and Haddow, 2016; Nadejde et al., 2017). Even full rule-based MT (RBMT) systems can be combined with NMT in order to build hybrid systems (Huang et al., 2020). In this work, we focus on an under-resourced lan- guage pair: Breton–French, and study mechanisms to build a hybrid system by combining NMT with the Breton–French system built with the Apertium RBMT platform (Forcada et al., 2011). We aim at producing sentences that combine knowledge extracted from the parallel corpus and from the RBMT system. Hence, we go beyond approaches that simply choose the best system (ei- ther RBMT or NMT) for each input sentence (see below). We use multi-source NMT and formalise the problem of combining both sources of knowl- edge as an automatic post-editing (Chatterjee et …

Content truncated to 500 words