ABSTRACT English Grammatical Error Correction (GEC) is an important component of Intelligent English Tutoring Systems (IETS), but edge deployment remains challenging because correction quality, latency, and pedagogical feedback must be balanced under tight hardware constraints. This paper presents TinyGEC, an edge‐oriented GEC framework based on multitask learning and dynamic sparse inference. TinyGEC jointly performs error detection, correction, and pedagogical rule attribution, while a lightweight router activates only relevant neural pathways to reduce computation. Evaluations on the BEA‐2019 benchmark and a representative Cortex‐M7 deployment target show that TinyGEC achieves an score of 61.5 with 8.4 ms latency. We further clarify threshold learning, PTQ settings, tokenizer configuration, and latency accounting to improve deployment transparency. The results indicate that TinyGEC provides a practical solution for privacy‐preserving, real‐time grammatical feedback in ubiquitous learning environments.
Huan He (Fri,) studied this question.