Gaze tracking and its utilization for readers’ behavioral analysis in wikipedia

Abstract

With the advent of research and technology, new techniques have been developed to capture human interest in a document. Eye gaze tracking is one of the well-researched and reliable mediums to predict human cognition. Eye gaze is a prominent nonverbal signal of human cognition compared to pointing, body posture, and other behaviors. In the context of document analysis, eye gaze distribution patterns can reveal crucial information regarding the sections of a document in which a reader is interested and where the reader faced difficulty in understanding the text. To conduct the relevant research, we select Wikipedia as the target platform where users perform reading operations. Wikipedia is an open-content encyclopedia that receives billions of visits per month. Wikipedia provides huge and publicly available data dumps. Wikipedia articles contain text as the primary medium, which makes it suitable to perform reading-based behavioral analysis. In this context, first, we present an eye gaze tracking approach to predict the line of sight of a human subject using an unsupervised machine learning technique and a laptop/desktop based camera. To train the proposed network in self-supervised manner, we collect a large ‘in the wild’ dataset containing 154,251 images from the web. In sequence, we present a number of research tracks that utilize eye gaze data to analyze the reading patterns of Wikipedia users. We propose an implicit attention feedback system for Wikipedia users. The feedback system is embedded in a website that interfaces Wikipedia content and camera-based eye-tracking technology. We also exploit the collected attention data for the evaluation of the significance of images in an article. Further, we propose a method for generating personalized summaries for Wikipedia articles based on user reading habits. We track eye gaze during the article reading session to perform reading pattern analysis. Eye gaze analysis aids in determining a reader’s attention distribution over an article. We extend the proposed method to generate a summary for multiple articles visited over the course of a single reading session. Finally, we also propose a summary recommendation technique. It helps users save time and energy to read lengthy Wikipedia articles. We present them with a customized and concise summary of the article which they intend to read. We personalize the summaries for each user based on their previous reading patterns onWikipedia. For the experimentation purposes, we gather volunteers to use our tools and read Wikipedia articles. Prospective participant-observers were required to have normal visual acuity and age in between 19 to 50 years, in-order for the eye tracker to perform accurately.