Cross-Lingual Feedback in Adaptive Learning Platforms: A Mixed-Methods Evaluation in Undergraduate Courses

Abstract

Adaptive learning platforms increasingly employ multilingual natural language processing (NLP) to support diverse student cohorts, yet empirical evidence on pedagogically effective, equitable cross-lingual feedback remains limited. We present and evaluate \emph{CLAF} (Cross-Lingual Adaptive Feedback), a system that combines multilingual sentence representations with a pedagogy-aligned feedback engine and an uncertainty-aware controller to regulate automated suggestions. In an eight-week quasi-experimental study across two universities (UK and Italy) involving 412 undergraduates writing in English, Italian, and Spanish, we compare CLAF-assisted courses with business-as-usual instruction. Using mixed-effects models, we observe medium learning gains on rubric-aligned writing outcomes (overall $d=0.38$), with larger effects for students below the median baseline ($d=0.52$). Instructor grading time decreases by $27\%$ without significant reduction in feedback quality. Bias and parity audits suggest no statistically significant performance gaps across target languages after covariate adjustment. Qualitative analysis of 36 interviews indicates perceived transparency and utility when feedback includes metacognitive prompts and source-linked exemplars. We discuss design implications for cross-lingual educational NLP, including model documentation, participatory evaluation, and safeguards against over-reliance on automated feedback.