ScholarWorks@성균관대학교

초록

Existing artificial intelligence-powered agents in language education often lack principled feedback differentiation, limiting opportunities for targeted speaking development. This study proposes a structured framework for designing generative pre-training transformer (GPT)-based conversational agents to optimize the delivery of corrective feedback in second-language (L2) speaking tasks. The proposed framework addresses this gap by outlining the design principles for implementing three evidence-based feedback types (i.e., recasts, clarification requests, and metalinguistic feedback) in GPT-powered task environments. The framework specifies prompt engineering strategies, error detection mechanisms, and task sequencing protocols grounded in L2 acquisition theory while acknowledging the current limitations of GPT technology in educational contexts. Unlike existing generic chatbot approaches, the proposed framework (1) provides systematic feedback differentiation based on the error characteristics and learner proficiency levels, (2) outlines realistic implementation scenarios specifically tailored for language education, though it has yet to be tested with actual learners, and (3) offers concrete implementation guidelines, including robust prompt templates; evaluation metrics for complexity, accuracy, and fluency; and ethical protocols for learner data protection. This study may provide actionable guidance for developers and educators aiming to build pedagogically grounded, GPT-based speaking practice systems.

키워드