Inferring Search User Language Proficiency from Eye Gaze Data

User language proficiency has been shown to significantly affect search result language preferences, search behaviors, and even interface preferences. Consequently, search systems may consider adapting to a user's language proficiency when retrieving, composing, and presenting search results. In order to perform such adaptation, it is necessary to first get an estimate of a user's proficiency, ideally through simply observing their behaviors while searching. To this end, this paper investigates the extent to which a user's language proficiency can be inferred from their eye movements while they are evaluating search results. Classification results involving data from English-, Spanish-, and Chinese-speaking study participants show that such an inference is indeed possible, and with relatively high accuracies for all languages. It is also shown that feature sets involving statistics from an entire search result page, combined with gaze data from individual results, have the highest classification accuracies. Moreover, a user's Average Fixation Durations, Refixations, and Pupil Dilations are found to be the most significant features.