Low-Cost Eye Tracking Corpus for Explainable Natural Language Processing

This infrastructure project enables the creation of a multilingual low-cost eye-tracking dataset, designed to make artificial intelligence-based language technologies fair and transparent. Deep learning models are ubiquitous in the present-day landscape of artificial intelligence. In order to understand what these models learn despite their generally opaque nature, and whether their rationales align with those of humans, we collect low-cost eye movement data through a crowd-sourcing platform and learn human rationales from gaze patterns. We collect gaze patterns from both task-specific and task-agnostic natural reading, providing a platform for evaluating the transparency and soundness of modern language technologies at scale. PIs: Anders Søgaard and Nora Hollenstein.