A Simple Recipe for Multilingual Grammatical Error Correction
Rothe, Sascha, Mallinson, Jonathan, Malmi, Eric, Krause, Sebastian, Severyn, Aliaksei
–arXiv.org Artificial Intelligence
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.
arXiv.org Artificial Intelligence
Aug-9-2022
- Country:
- North America > United States
- California (0.28)
- Minnesota (0.28)
- North America > United States
- Genre:
- Research Report (0.70)
- Technology: