Dual Input Stream Transformer for Vertical Drift Correction in Eye-Tracking Reading Data.

Area of Science:

• Cognitive Science
• Computer Science
• Human-Computer Interaction

Background:

• Eye-tracking during reading generates fixation point data.
• Accurate assignment of fixations to text lines is crucial for reading analysis.
• Vertical drift in eye-tracking data introduces noise, complicating fixation-to-line assignment.

Purpose of the Study:

• To introduce a novel Dual Input Stream Transformer (DIST) model for accurate fixation point assignment in reading data.
• To address the challenge of noise, specifically vertical drift, in eye-tracking data.
• To provide a robust and accurate solution for the post-processing bottleneck in reading research.

Main Methods:

• Development of the Dual Input Stream Transformer (DIST) architecture.
• Evaluation of DIST against eleven classical methods across nine diverse datasets.
• Ensemble methods combining multiple DIST instances and classical approaches were employed.

Main Results:

• DIST models, particularly when ensembled, demonstrated high accuracy across all tested datasets.
• An ensemble of DIST models combined with the best classical approach achieved an average accuracy of 98.17%.
• Ablation studies identified line overlap features and the dual input stream as key contributors to DIST's success.

Conclusions:

• The Dual Input Stream Transformer (DIST) offers a significant advancement in automating fixation-to-line assignment for eye-tracking data.
• DIST is robust across various experimental setups, making it a reliable tool for reading research.
• This approach effectively addresses the manual assignment bottleneck, accelerating reading data analysis.