Using a custom GPT to scaffold corpus-based vocabulary learning for low-proficiency learners

Abstract

A corpus-based approach is often considered unsuitable for low-proficiency learners, who frequently experience cognitive overload in such learning contexts. To help these learners manage the cognitive demands of corpus-based vocabulary learning, this study designed a custom GPT configured with six scaffolding methods to address three key challenges commonly faced by low-proficiency learners and integrated it into the framework of corpus-based language pedagogy (CBLP). Unlike teacher-led scaffolding, which may not allow simultaneous support for multiple learners due to inevitable physical constraints, the custom GPT delivers on-demand, contingent support through learner-initiated requests. This mixed-methods study compared two within group conditions: CBLP with and without custom GPT scaffolding, involving twenty low-proficiency L2 English learners at a vocational college in China. Data sources included vocabulary tests, GPT chat logs, and semi-structured interviews. Results showed that GPT-scaffolded CBLP significantly improved vocabulary knowledge, with the most substantial gains in use. Content analysis of GPT chats revealed that the custom GPT delivered six scaffolding methods in ways that responded contingently to the specific challenges learners faced. Thematic analysis of interview transcripts further indicated that GPT scaffolding demonstrated the characteristic of contingency. Integrating quantitative and qualitative findings, we found that GPT scaffolding appeared to have enhanced the effectiveness of CBLP. The present study contributes to the limited but much-needed knowledge of how to integrate contingent scaffolding, enabled by generative AI chatbots, into CBLP lesson designs aimed at enhancing vocabulary learning for low-proficiency learners.