Efficient Grammatical Error Correction with Hierarchical Error Detections and Correction

Noisy text is common in semantic services and can have bad effects. Grammatical Error Correction (GEC) can be used to improve text quality, but traditional neural machine translation approaches need hundreds of milliseconds to correct a single text, which is unacceptable to time-sensitive services. To improve the efficiency of GEC, we choose to detect errors first and then make corrections. We present an intuitive multitask learning approach by checking if the text contains errors, finding errors' positions, and finally generating corrections. Two classifiers are introduced to serially detect sentence-level and token-level errors as errors only take a few parts in common corpora. Different from traditional approaches, we adapt a non-autoregressive decoder and only generate needed words to correct those detected errors, making the correction stage efficient. Experiments show that our approach can be ten times faster than the traditional approach in inference, and can achieve a comparable GEC performance.