A Study on Automated Recognition of Grammatical Errors in English Language Acquisition Based on BERT Modeling

Abstract

Due to the internationalization of English, many non-native English learners have developed incorrect pragmatic habits. Traditional manual correction of English pragmatic habits is both inefficient and difficult to adapt to large-scale teaching. This paper proposes a multi-task learning grammar error recognition framework based on the combination of BERT, BiLSTM, and CRF structures. In multi-task learning, the BERT language model is used for feature extraction of grammar errors, combined with BiLSTM for sequence learning and CRF for global optimal feature extraction. The three-layer structure model trained through multi-task learning can effectively identify various types of grammatical errors, achieving an average F1 score of 88.0% in evaluation metrics. The model demonstrates excellent performance across different types of grammatical error recognition, with high accuracy and robustness. Through multi-task learning, the grammatical error recognition model exhibits strong generalization capabilities, providing intelligent and personalized technical support for English education and marking a new step toward the intelligentization of future educational practices.