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

Code

• github.com/google-research/multilingual-t5
  Officialtf★ 1,297
• github.com/huggingface/transformers
  pytorch★ 158,292
• github.com/google-research/byt5
  tf★ 539
• github.com/MorenoLaQuatra/bart-it
  pytorch★ 16
• github.com/manshri/tesum
  pytorch★ 1
• github.com/pwc-1/Paper-5/tree/main/mt5
  mindspore★ 0
• github.com/KoshiroSato/Simple_Transformers_mT5_finetuning
  none★ 0
• github.com/KoshiroSato/Flask_NLP_App
  none★ 0

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.

Tasks

Benchmark Results

Reproductions