Transforming online grammar education: Integrating GETN V2 and RoBERTa embeddings for effective English grammar correction

Abstract

The increasing demand for intelligent language learning systems, traditional grammar correction models often fail to capture the complex syntactic dependencies and contextual nuances present in learner-generated texts. These limitations hinder accurate detection and correction of grammatical errors particularly in academic writing. This paper proposes a novel grammar correction framework that integrates deep contextual embeddings from RoBERTa with syntactic graph structures using Graph Embedded Transformer Network (GETN V2). The input sentences are first pre-processed through normalization, tokenization, syntactic parsing and stop word removal using the ChatLang-8 dataset. RoBERTa is then applied to extract high-dimensional contextual embeddings for each token capturing semantic dependencies. These embeddings are fused with syntactic graphs derived from dependency parsing where grammatical relationships such as subject–verb and modifier-noun are represented as labeled edges. The GETN V2 encoder combines these inputs through multi-relational message passing and relation-aware attention mechanisms dynamically weighting syntactic dependencies using graph-augmented transformer layers. A dual-module architecture performs error detection and correction in such a way that the detection layer identifies grammatical inconsistencies via masked attention and classification blocks, while the correction module leverages the confidence estimator and correction generator to refine output. The system also incorporates a feedback loop with a confusion matrix to dynamically update error correction strategies. Experimental evaluation on benchmark datasets demonstrates that the proposed model achieves an accuracy of 88.26% with F0.5 scores exceeding 84% on syntactically grounded errors such as subject–verb agreement and verb tense. Furthermore, the model maintains low latency under high concurrency making it suitable for real-time educational deployment.