Systematic Drift Correction in Eye Tracking Reading Studies: Integrating Line Assignments with Implicit Recalibration

EasyChair Preprint no. 13293

Abstract

Eye tracking data is typically compromised by a systematic error that is commonly referred to as drift. In reading research, most manual and automated approaches to dealing with drift assign fixations to lines of text. However, correcting for the y dimension only means that horizontal misalignment of fixations is neglected. Available approaches for horizontal correction that involve inferring systematic error from probable fixation locations have not been used in conjunction with line assignment procedures. In this paper, we present a novel approach for extracting the systematic error across multiple reading trials. Starting with trial-by-trial line-to-word mapping for multi-line text, our approach uses a line assignment algorithm based on dynamic time warping. This initial step is followed by an extraction of systematic drift through spatial and temporal filtering to reduce artificial noise. We compare this approach with manual ground-truth line assignments and explicit validation grids. For a set of data from a reading study (N = 30), our method significantly reduced drift in both the horizontal and vertical dimensions. The agreement of a number of vertical drift correction algorithms with manual line assignment improved from 75 to 85\% to over 94\% by prior elimination of systematic drift with our method, even outperforming results by prior correction of drift derived from validation grids. This suggests that accounting for systematic drift over trials may lead to more accurate correction in reading studies.

Keyphrases: drift correction, eye tracking, implicit recalibration, line assignment, multiline reading

@Booklet{EasyChair:13293,
  author = {Wolf Culemann and Leana Neuber and Angela Heine},
  title = {Systematic Drift Correction in Eye Tracking Reading Studies: Integrating Line Assignments with Implicit Recalibration},
  howpublished = {EasyChair Preprint no. 13293},

  year = {EasyChair, 2024}}