Implicit Estimation of Paragraph Relevance From Eye Movements

Eye movements were shown to be an effective source of implicit relevance feedback in constrained search and decision-making tasks. Recent research suggests that gaze-based features, extracted from scanpaths over short news articles (g-REL), can reveal the perceived relevance of read text with respect to a previously shown trigger question. In this work, we aim to confirm this finding and we investigate whether it generalizes to multi-paragraph documents from Wikipedia (Google Natural Questions) that require readers to scroll down to read the whole text. We conduct a user study (n = 24) in which participants read single- and multi-paragraph articles and rate their relevance at the paragraph level with respect to a trigger question. We model the perceived document relevance using machine learning and features from the literature as input. Our results confirm that eye movements can be used to effectively model the relevance of short news articles, in particular if we exclude difficult cases: documents which are on topic of the trigger questions but irrelevant. However, our results do not clearly show that the modeling approach generalizes to multi-paragraph document settings. We publish our dataset and our code for feature extraction under an open source license to enable future research in the field of gaze-based implicit relevance feedback.

Abstract licence: CC BY

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.