Predicting Attention Shifts in Second Language Learners Using LSTM and Attention Mechanism: A Multi-Dataset Approach

This paper presents a novel approach for predicting attention shifts in second language learners using a Long Short-Term Memory (LSTM) neural network integrated with an attention mechanism. Eye-tracking data from multiple datasets, including Dundee, ZuCo, Tobii, and GazePoint, were used to evaluate the performance of the proposed model. The LSTM model, enhanced by attention, effectively captures temporal dependencies in learners’ behavior and outperforms traditional models such as Naive Bayes, SVM, and RNN in terms of accuracy, F1-score, and Mean Absolute Error (MAE). The LSTM model achieved the highest accuracy of 86.0% and the lowest MAE of 0.10 on the GazePoint dataset. An ablation study further demonstrates the critical role of the attention mechanism in improving prediction accuracy. These results show that the proposed method provides a robust and scalable solution for modeling attention shifts in second language acquisition.