mT5: A massively multilingual pre-trained text-to-text transformer

TL;DR

mT5 extends the T5 model to 101 languages, achieving state-of-the-art results on multilingual NLP tasks through a new training dataset and techniques to prevent accidental translation.

Contribution

This paper introduces mT5, a multilingual T5 model trained on a large dataset covering 101 languages, with improved multilingual performance and a method to avoid accidental translation.

Findings

Achieved state-of-the-art results on multiple multilingual benchmarks.

Demonstrated effective techniques to prevent accidental translation.

Provided publicly available code and model checkpoints.

Abstract

The recent "Text-to-Text Transfer Transformer" (T5) leveraged a unified text-to-text format and scale to attain state-of-the-art results on a wide variety of English-language NLP tasks. In this paper, we introduce mT5, a multilingual variant of T5 that was pre-trained on a new Common Crawl-based dataset covering 101 languages. We detail the design and modified training of mT5 and demonstrate its state-of-the-art performance on many multilingual benchmarks. We also describe a simple technique to prevent "accidental translation" in the zero-shot setting, where a generative model chooses to (partially) translate its prediction into the wrong language. All of the code and model checkpoints used in this work are publicly available.