Abstract
Parallel corpora are indispensable for training neural machine translation(NMT) models, and parallel corpora for most language pairs do not exist or arescarce. In such cases, pivot language NMT can be helpful where a pivot languageis used such that there exist parallel corpora between the source and pivot andpivot and target languages. Naturally, the quality of pivot languagetranslation is more inferior to what could be achieved with a direct parallelcorpus of a reasonable size for that pair. In a real-time simultaneoustranslation setting, the quality of pivot language translation deteriorateseven further given that the model has to output translations the moment a fewsource words become available. To solve this issue, we propose multi-pivottranslation and apply it to a simultaneous translation setting involving pivotlanguages. Our approach involves simultaneously translating a source languageinto multiple pivots, which are then simultaneously translated together intothe target language by leveraging multi-source NMT. Our experiments in alow-resource setting using the N-way parallel UN corpus for Arabic to EnglishNMT via French and Spanish as pivots reveals that in a simultaneous pivot NMTsetting, using two pivot languages can lead to an improvement of up to 5.8BLEU.