Return to Past and Current Projects
Predictive scroller
The past decade has seen a dramatic shift in how people read news articles, technical manuals, and books. With the popularity of smartphones, tablets, and e-readers, individuals spend much time scrolling a touch screen to advance the visible window of text. Devices are often used in a manner that makes it cumbersome to keep one hand free to scroll, e.g., reading while lying on one's back. Attempts have been made to use front cameras to detect gaze and automatically scroll when the eyes reach the bottom of the screen (e.g., Biedert et al., 2012; Guo & Wang, 2017), though implementations in commercial offerings appear to have been unsuccessful (e.g., gaze-based scrolling was a short-lived and poorly-functioning feature of Samsung phones).
As an alternative, we explore an machine learning approach in which an individual reader's scrolling actions are used to train a personalized model to predict that individual's specific behavior conditioned on screen contents. We leverage a rich cognitive science literature on reading that has determined text features influencing reading time (Graesser, McNamara, & Kulikowich, 2011) to represent screen contents, and take advantage of the fact that users can provide abundant training data in the ordinary course of small-screen reading.
This domain seems particularly ripe for AI-based assistance, given that the cost of misprediction is low---the user can always pause the automatic scrolling or scroll backwards---and the predictions can be acted on in a graded manner to lessen the impact of incorrect predictions---for example, the strength of a prediction can determine scroll rate rather than distance. The AI assistant may also be useful for alerting users if mind wandering occurs: when reading times decouple from text complexity, comprehension drops significantly (Mills, Graesser, Risko, & D'Mello, 2017).
Students
Arman Aydemir (Computer Science, Boulder)
Michael Neuder (Computer Science, Boulder)