Predicting Vocabulary Acquisition in EFL Learners Using Machine Learning Models

<title>Abstract</title> <p>Artificial intelligence (AI) is increasingly being integrated into English as a Foreign Language (EFL) instruction. However, relatively little is known about how well learners' vocabulary development can be predicted by combining linguistic indicators with AI interaction data. This study investigated vocabulary growth among 180 undergraduate EFL learners over a 15-week intervention and evaluated the predictive performance of several supervised machine-learning models. Students were divided into two groups. The first is an AI-assisted group (n = 90), which used ChatGPT and Google Gemini to support writing and vocabulary learning. The second is a comparison (control) group (n = 90), which received conventional teacher feedback. Vocabulary knowledge was measured before and after the intervention. Changes in Type–Token Ratio (TTR), Measure of Textual Lexical Diversity (MTLD), and Academic Word List (AWL) coverage were extracted from learners' writing. Three behavioral indicators—AI usage time, engagement score, and revision frequency—were also recorded for the AI-assisted group. Inferential analyses showed significant vocabulary improvement in both groups. The AI-assisted group achieved substantially greater gains even after controlling for baseline vocabulary knowledge. Vocabulary gain was positively associated with improvements in lexical diversity and academic vocabulary use. In the final regression model, however, instructional group was the strongest unique predictor, while the behavioral indicators made little additional contribution within the AI-assisted group. Four supervised machine-learning models—Linear Regression (LR), Support Vector Regression (SVR), Random Forest (RF), and Gradient Boosting (GB)—were compared using five-fold cross-validation. Linear Regression achieved the highest predictive performance, followed by Support Vector Regression. Neither Random Forest nor Gradient Boosting improved predictive accuracy. Within the AI-assisted subgroup, behavioral indicators did not enhance prediction beyond the linguistic variables. The findings suggest that generative AI can strengthen vocabulary learning when integrated into teacher-guided writing instruction. They also demonstrate that inferential and machine-learning approaches provide complementary insights into vocabulary development and prediction. The study contributes an integrated analytical framework that combines educational data mining with language-learning research to better understand and predict vocabulary acquisition in AI-supported EFL contexts.</p>