Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation

TL;DR

This paper enhances massively multilingual neural machine translation by increasing model capacity, introducing language-specific components, and using random online backtranslation, significantly improving zero-shot translation quality.

Contribution

It proposes novel methods including language-specific modules and online backtranslation to boost multilingual NMT and zero-shot translation performance.

Findings

Narrowed the performance gap with bilingual models

Improved zero-shot translation by approximately 10 BLEU points

Demonstrated effectiveness on a 100-language dataset

Abstract

Massively multilingual models for neural machine translation (NMT) are theoretically attractive, but often underperform bilingual models and deliver poor zero-shot translations. In this paper, we explore ways to improve them. We argue that multilingual NMT requires stronger modeling capacity to support language pairs with varying typological characteristics, and overcome this bottleneck via language-specific components and deepening NMT architectures. We identify the off-target translation issue (i.e. translating into a wrong target language) as the major source of the inferior zero-shot performance, and propose random online backtranslation to enforce the translation of unseen training language pairs. Experiments on OPUS-100 (a novel multilingual dataset with 100 languages) show that our approach substantially narrows the performance gap with bilingual models in both one-to-many and many-to-many settings, and improves zero-shot performance by ~10 BLEU, approaching conventional pivot-based methods.