Aiming at the problems of semantic distortion and over-correction that often occur in the text error correction of English learners by existing generative models, this paper accordingly proposes a novel generative adversarial network method -integrating grammatical rule constraints (generative adversarial networks with grammatical rule constraints).By introducing formal grammatical knowledge as a flexible constraint into the network training process, the model is effectively guided to correct errors while better maintaining the original meaning and overall fluency of sentences.Experiments conducted on the public learner corpus show that this method significantly increases the error correction accuracy by 12% and effectively reduces the number of over-correction cases by 16%.The research thereby provides an effective way to solve the persistent balance problem between accuracy and naturalness in automatic grammar correction.
N.A. Liu (Thu,) studied this question.