Sascha Rothe Jonathan Mallinson Eric Malmi Sebastian Krause Aliaksei Severyn

Abstract

This paper presents a simple recipe to train state-of-the-art multilingual Grammatical Error Correction (GEC) models. We achieve this by first proposing a language-agnostic method to generate a large number of synthetic examples. The second ingredient is to use large-scale multilingual language models (up to 11B parameters). Once fine-tuned on language-specific supervised sets we surpass the previous state-of-the-art results on GEC benchmarks in four languages: English, Czech, German and Russian. Having established a new set of baselines for GEC, we make our results easily reproducible and accessible by releasing a cLang-8 dataset. It is produced by using our best model, which we call gT5, to clean the targets of a widely used yet noisy lang-8 dataset. cLang-8 greatly simplifies typical GEC training pipelines composed of multiple fine-tuning stages -- we demonstrate that performing a single fine-tuning step on cLang-8 with the off-the-shelf language models yields further accuracy improvements over an already top-performing gT5 model for English.