Vertical Error Correction Using Classification of Transitions between Sequential Reading Segments

Abstract

In eye-tracking-based reading behavior research, gaze sampling errors often negatively affect gaze-to-word mapping. In this paper, we propose a method for more accurate mapping by first taking adjacent horizontally progressive fixations as segments, and then classifying the segments into six classes using a random forest classifier. The segments are then reconstructed based on the classification, and are associated with a document line using a dynamic programming algorithm. The combination of segment-to-line mapping and transition classification achieved 87% mapping accuracy. We also witnessed a reduction of manual annotation time when the mapping was used as an annotation guiding tool.