EDCEW-LLM: Error detection and correction in English writing: A large language model-based approach

Abstract

This study investigates the use of large language models (LLMs) for error detection and correction in English writing (EDCEW), focusing on the challenges of overcorrection and pedagogical feedback. We explore the relationship between targeted prompting strategies, transformer model architectures, and the error profiles of English language learners across CEFR proficiency levels (A, B, and C). We evaluated LLM through zero-shot, few-shot and fine-tuning approaches, showing that overcorrection is more common in texts from advanced learners, while lower proficiency levels show unique patterns of errors. To address this issue, we introduce a two-stage pipeline: first, a fine-tuned LLM leverages sequence-to-sequence modeling to perform precise token-level error detection and correction, achieving competitive F1 scores across multilingual datasets (English, German, Chinese); second, a prompted LLM (GPT 3.5) generates rule-based, natural language explanations for each identified error, enhancing the educational values. We present comprehensive benchmark results on standard WEDC datasets (BEA-2019, JFLEG), supported by human evaluations, that show the effectiveness of the proposed approach in producing accurate corrections and linguistically informative error explanations. The proposed model improves correction accuracy and facilitates deeper linguistic comprehension through explicit, pedagogically sound feedback, addressing a significant limitation of existing automated GEC tools. Moreover, this research advances LLM-driven WEDC by providing a robust framework that balances correction precision with pedagogical utility. Qualitative results further highlight the model’s ability to address diverse error types, such as Subject+Verb agreement, redundant comparisons, and incorrect preposition usage, with contextually appropriate corrections.