Genetic Algorithm Versus Random Feedback Selection in Computer-Assisted Language Learning

Abstract

This study tests whether the rule used by a computer-assisted language learning (CALL) tool to select corrective-feedback types affects learners' English-tense accuracy. A browser-based tutor delivered one of six feedback types after each error. In one condition, a genetic algorithm assigned feedback using learner-specific value estimates, preference weights, and a repetition penalty; in the other, types were assigned randomly, while items, order, message count, and total feedback time were matched. Of 255 Moroccan university students allocated, 103 completed five sessions. The algorithm changed the feedback mix, but accuracy did not differ between conditions (odds ratio = 0.90, 95% confidence interval [0.67, 1.20]; 60.2% vs. 57.4%) and did not increase across sessions. Logs indicated time pressure for many messages, weak preference weights, and value estimates partly driven by item distribution. Adaptive feedback selection requires valid learner signals and adequate processing time, not merely a different assignment rule.