Grammar correction that tags instead of generating

What it does

This is the official implementation of GECToR, a grammatical error correction and text simplification system that assigns each input token a transformation label rather than drafting corrected text from scratch. It wraps pre-trained Transformer encoders inside an AllenNLP pipeline and ships with ready-made checkpoints for BERT, RoBERTa, and XLNet.

The interesting bit

Most grammar tools behave like ghostwriters that draft entire sentences; GECToR behaves like a copyeditor with a red pen, applying token-level transformation labels to the original text. Because the model is fully non-autoregressive, the authors report inference up to 10× faster than Transformer seq2seq systems for grammar correction, and over 11× faster for text simplification.

Key highlights

• Approaches GEC and text simplification as sequence-labeling tasks over pre-trained encoders rather than generation problems
• Pretrained models available for BERT, RoBERTa, and XLNet with published F0.5 and SARI scores on CoNLL-2014, BEA-2019, and ASSET benchmarks
• Two-stage fine-tuning: first on errorful corpora, then on a mix of errorful and error-free parallel data
• Inference tuned with confidence bias and minimum error probability thresholds that act as precision-recall dials
• Single codebase covers both grammar correction and text simplification via the same tagging interface

Caveats

• The reported benchmark scores differ slightly from the original papers because the codebase was upgraded to a newer transformers release; reproducing exact paper numbers requires an older commit
• The project was last validated on Python 3.7, so running it on modern environments may demand dependency adjustments

Verdict

Grab it if you need a fast, non-autoregressive baseline for grammar correction or simplification research. Pass if you want a polished product API or a seq2seq model that freely rewrites entire phrases.