Modeling Sub-Document Attention Using Viewport Time
summary
Compares three models of sub-document attention using 1.2M reading sessions from a news site and standard browser viewport data rather than eye tracking. A uniform time-on-page model overestimates reading rate (English ~600 wpm) while the Gaussian viewport attention model matches known empirical rates (English ≈ 220 wpm) across nine languages. Uses paragraph word counts to compute per-element reading rates and shows practical finds like image engagement varying by topic (sex > music > politics).
AI-written summary · about
chapters (8)
- 0:00Introduction to research
- 1:09Goal of sub-document attention
- 3:17Limitations of uniform attention model
- 4:42Gaussian viewport attention model
- 6:30Validating attention models with reading rates
- 8:13Applications to images and engagement
- 9:15Implications and future uses
- 9:36Conclusion and Q&A
related talks
19mHow to look at your data
38mWhere are we looking? Predicting human gaze using deep networks.
20mReport Design & Psychology (Quick Tips)
29mModel Evaluation for Custom Datasets With Open Models: Multi-Model Comparison With...
20mLearning Visual Importance for Graphic Designs and Data Visualizations
45mDon't trust the numbers!
19mAn Eye For Design: Gaze Visualizations for Remote Collaborative Work
53mDemocratising Attention Data at theguardian.com