Controllable data synthesis method for grammatical error correction

Due to the lack of parallel data in current grammatical error correction (GEC) task, models based on sequence framework cannot be adequately trained obtain higher performance. We propose two synthesis methods which can control rate and ratio types synthetic data. The first approach is corrupt each word monolingual corpus with a fixed probability, including replacement, insertion deletion. Another train generation further filtering decoding results models. experiments different show that 40% same improve model performance better. Finally, we synthesize about 100 million achieve comparable as state art, uses twice much use.