Papers
arxiv:2010.11934

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

Published on Oct 22, 2020
Authors:
,
,
,
,
,
,
,

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.

Community

Sign up or log in to comment

Models citing this paper 24

Browse 24 models citing this paper

Datasets citing this paper 4

Spaces citing this paper 38

Collections including this paper 3