Reading ability detection using eye-tracking data with LSTM-based few-shot learning

Previous works demonstrated that eye-tracking data supplied meaningful information for reading ability detection, and have gained promising results by employing machine learning methods [1-18]. The eye-tracking based methods of reading ability detection fell into two main categories: the one estimated reading ability with finite number of classes [1-14], providing qualitative evaluation of subjects' reading ability. The other predicted reading ability scores with regression models [15-18], rendering quantitative evaluation of subjects' reading ability. Although the former exhibited satisfactory accuracy in detecting certain classes of abnormalities in reading, it lacked the capability of predicting exact scores of reading ability, which was emphasized in highly interactive educational environments (such as online learning) to make personal and intelligent reactions to subjects. However, precise score prediction of reading ability using eye-tracking data is not easy [15-18], especially when the sample data of subjects are few. In this paper, with few-shot learning strategy, a regression model for score prediction is proposed by combining Long Short Time Memory (LSTM) [19] and light-weighted neural networks. The proposed model exhibits higher accuracy than previous score prediction models tested on the same dataset.