Core Concepts
WebQAmGaze provides valuable insights into eye movement patterns during reading tasks, enhancing computational language processing models.
Abstract
WebQAmGaze is a multilingual webcam eye-tracking dataset designed to support the development of explainable computational language processing models. The dataset includes data from 600 participants reading texts in English, German, Spanish, and Turkish. Participants perform normal reading and information-seeking tasks followed by comprehension questions. The dataset aims to advance webcam-based reading studies and provide low-cost data collection options for diverse populations.
Structure:
-
Introduction to Eye Movement Data in NLP Models
- Eye-tracking recordings enhance NLP models by providing human inductive bias.
- Machine learning models benefit from eye movement data for various NLP tasks.
-
Leveraging Eye Movement Data for Language Understanding
- Eye movement signals modulate machine learning models towards cognitively plausible processing.
- Various datasets have been created to study different properties using eye-tracking data.
-
Webcam-Based Eye-Tracking Technology Advancements
- Low-cost video-based eye-tracking methods are explored with appearance-based gaze estimation models.
- Publicly available libraries like WebGazer show promise in improving eye-tracking accuracy.
-
WebQAmGaze Dataset Creation Process
- Experiment design includes normal reading and information-seeking tasks.
- Reading materials selected from XQuAD and MECO datasets for multiple languages.
-
Data Processing Steps for Fixation Detection and Area of Interest Identification
-
Comparison with High-Quality Eye-Tracking Data (MECO)
- Comparison of total reading time and number of fixations between WebQAmGaze and MECO datasets.
- Spearman correlation coefficients show strong correlations between relative fixation durations.
-
Towards Explainable AI with Eye Movement Rationales
- Hypothesis that eye movement information can be used to extract rationales without manual annotations.
- Investigation on whether fixation-based measures are indicative of correct answers in question-answering tasks.
Stats
WebQAmGazeは、600人の参加者から収集されたデータを含むマルチリンガルなウェブカムアイ・トラッキングデータセットです。
参加者は通常の読書と情報検索タスクを行い、理解度の質問に回答します。