Sequence-to-Action: Grammatical Error Correction with Action Guided Sequence Generation

The task of Grammatical Error Correction (GEC) has received remarkable attention with wide applications in Natural Language Processing (NLP) recent years. While one the key principles GEC is to keep correct parts unchanged and avoid over-correction, previous sequence-to-sequence (seq2seq) models generate results from scratch, which are not guaranteed follow original sentence structure may suffer over-correction problem. In meantime, recently proposed sequence tagging can overcome problem by only generating edit operations, but conditioned on human designed language-specific labels. this paper, we combine pros alleviate cons both proposing a novel Sequence-to-Action (S2A) module. S2A module jointly takes source target sentences as input, able automatically token-level action before predicting each token, where generated three choices named SKIP, COPY GENerate. Then actions fused basic seq2seq framework provide final predictions. We conduct experiments benchmark datasets English Chinese tasks. Our model consistently outperforms baselines, while being significantly well holding better generality diversity generation compared models.