Abstract:Current English GEC methods tend to ignore the syntactic knowledge in texts, which plays an important role in grammatical error correction, and thus the error correction ability of English GEC models is affected. To address this problem, the study proposes a GEC method which is based on the differential fusion syntactic features. First, the proposed syntactic encoder can generate dependency graph and constituency syntactic tree information from raw data in an unsupervised way and conduct the feature fusion of these two heterogeneous syntactic structures to encode high-dimensional syntactic representation. Second, to utilize both semantic and syntactic information in the text, the differential fusion module first uses differential regularization to enhance the semantic encoder to capture the semantic features that the syntactic encoder fails to generate. Then the syntactic representation and semantic representation are further fused by cross attention as the output features of the Transformer encoder, which are finally input to the decoder to generate grammatically correct text. The comparison experiment on the CoNLL-2014 task dataset shows that the precision and F_0.5 value of this method are better than those of the GEC model based on the Copy-Augmented Transformer, and the F_0.5 value of this method is improved by 5.2 percentage points. The syntactic knowledge avoids the problem of lacking high-quality annotated training corpora and has a better performance in text error correction.