Towards predicting reading comprehension from gaze behavior

As readers of a language, we all agree to move our eyes in roughly the same way. Yet might there be hidden within this self-similar behavior subtle clues as to how a reader is understanding the material being read? Here we attempt to decode a reader's eye movements to predict their level of text comprehension and related states. Eye movements were recorded from 95 people reading 4 published SAT passages, each followed by corresponding SAT questions and self-evaluation questionnaires. A sequence of 21 fixation-location (x,y), fixation-duration, and pupil-size features were extracted from the reading behavior and input to two deep networks (CNN/RNN), which were used to predict the reader's comprehension level and other comprehension-related variables. The best overall comprehension prediction accuracy was 65% (cf. null accuracy = 54%) obtained by CNN. This prediction generalized well to fixations on new passages (64%) from the same readers, but did not generalize to fixations from new readers (41%), implying substantial individual differences in reading behavior. Our work is the first attempt to predict comprehension from fixations using deep networks, where we hope that our large reading dataset and our protocol for evaluation will benefit the development of new methods for predicting reading comprehension by decoding gaze behavior.