Improved grammatical error correction by ranking elementary edits
We offer a rescoring method for grammatical error correction which is based on two-stage procedure: the first stage model extracts local edits and the second classiifies them as correct or false. We show how to use an encoder-decoder or sequence labeling approach as the first stage of our model. We achieve state-of-the-art quality on BEA 2019 English dataset even with a weak BERT-GEC basic model. When using a state-of-the-art GECToR edit generator and the combined scorer, our model beats GECToR on BEA 2019 by $2-3\%$. Our model also beats previous state-of-the-art on Russian, despite using smaller models and less data than the previous approaches.
PDF AbstractDatasets
Results from the Paper
Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
---|---|---|---|---|---|---|
Grammatical Error Correction | BEA-2019 (test) | clang_large_ft2-gector | F0.5 | 77.1 | # 4 |