Toward a Real-time (Day) Dreamcatcher: Detecting Mind Wandering Episodes During Online Reading

This paper reports the results from a sensor-free detector of mind wandering during an online reading task. Features consisted of reading behaviors (e.g., reading time) and textual features (e.g., level of difficulty) extracted from self-paced reading log files. Supervised machine learning was applied to two datasets in order to predict if participants were mind wandering as they navigated from one screen of text to the next. Mind wandering was detected with an accuracy of 20% above chance (Cohen’s kappa = .207; AUC = .609), which was obtained via leave-one-participant-out cross-validation. Similar to actual rates of mind wandering, predicted rates of mind wandering were negatively related to posttest performance, thus providing some evidence for the predictive validity of the detector. Applications of the detector to attention-aware educational interfaces are discussed.