Machine Learning Approaches for Automated Vocabulary Acquisition in ESL Classrooms

Purpose: This study investigates the efficacy of machine learning (ML) approaches for automated vocabulary acquisition in English as a Second Language (ESL) classrooms. It focuses on transformer-based models (specifically BERT), comparing their performance to traditional supervised algorithms and examining effects on learner vocabulary gains. Methods: University-level ESL students in Azerbaijan (N = 60) participated in an experiment with an ML-driven vocabulary learning tool. A pre-trained BERT model was fine-tuned via TensorFlow for vocabulary prediction tasks and deployed to personalize practice for an experimental group, while a control group received conventional instruction. Support Vector Machine (SVM) and Random Forest models served as baseline algorithms for predictive performance benchmarking. Vocabulary knowledge was assessed pre- and post-intervention using standardized tests, and ML models were evaluated on accuracy, precision, and recall. Results: The fine-tuned BERT model achieved higher predictive accuracy (88%) than SVM (75%) or Random Forest (78%), with superior precision and recall. The experimental group outperformed the control on post-test vocabulary gains (mean improvement = 10.1 vs. 5.7 words, p < .01). Implications: Results indicate that transformer-based ML can enhance vocabulary learning outcomes, offering context-aware recommendations that surpass traditional models. We discuss how deep neural networks and reinforcement learning techniques can be integrated into ESL pedagogy to support adaptive vocabulary instruction. The study contributes a framework for applying state-of-the-art ML in language education and highlights implications for personalized learning and curriculum design.