An Empirical Investigation of Multi-bridge Multilingual NMT models

TL;DR

This paper empirically investigates multi-bridge multilingual NMT models, demonstrating their ability to improve low-resource and zero-shot translation, and showing they can serve all translation directions with a single model.

Contribution

It provides an extensive empirical analysis of multi-bridge multilingual NMT models, highlighting their advantages over pivot and English-centric models in low-resource and zero-shot translation scenarios.

Findings

MB-M2M models can extract parallel data between non-English languages.

With limited data, MB-M2M models outperform pivot models.

MB-M2M models can serve all translation directions in a single system.

Abstract

In this paper, we present an extensive investigation of multi-bridge, many-to-many multilingual NMT models (MB-M2M) ie., models trained on non-English language pairs in addition to English-centric language pairs. In addition to validating previous work which shows that MB-M2M models can overcome zeroshot translation problems, our analysis reveals the following results about multibridge models: (1) it is possible to extract a reasonable amount of parallel corpora between non-English languages for low-resource languages (2) with limited non-English centric data, MB-M2M models are competitive with or outperform pivot models, (3) MB-M2M models can outperform English-Any models and perform at par with Any-English models, so a single multilingual NMT system can serve all translation directions.