Detecting reading difficulties using eye-tracking metrics and machine learning models with the purpose of building adaptive tools for low-literacy support.

Abstract

In this study, we aimed to predict whether a word was too complex or not too complex for a low-literate individual to understand, with the purpose of creating a foundation for developing a text simplification tool tailored to a specific person. Eye-tracking and audio data are collected using an experiment. A baseline model is trained, which is a model that is not tailored to an individual but bases its prediction on the difficulty of the word. SVMs, AlexNet models, and ResNet50 models are trained using the collected data and the difficulty of the word. The models could not improve on the baseline model, probably due to too little data, but the ResNet50 model showed potential on audio, pupil, and eye movement data. Leading to suggestions for the next step in creating a simplification tool tailored to an individual.