Tracking Reading Progress Using an Auto-Encoding Neural Network

In the field of oculography tracking reading progress is challenging due to measurement errors in eye tracking systems. This paper presents a two-stage approach using an autoencoding neural network model. The model uses fixation duration and text structure to create a probability map representing the likelihood that each pixel is viewed and interprets this map using a specific criterion to determine whether a word has been read. The model was trained and evaluated using a synthetic dataset generated from the ZuCo 1.0 dataset. The model achieved an F-measure value of 0.9782 and an MIoU value of 0.9587 on a test set of 537 pages for the task of classifying words into read and unread.