SlideAudit: A Dataset and Taxonomy for Automated Evaluation of Presentation Slides
summary
Dataset of 2,400 slides and a 27 flaw types taxonomy across five categories, created with designers and annotated via trained Prolific workers (three annotators per slide, majority vote). Multiple LLMs and prompting strategies—plus visual cues like saliency and contrast—are evaluated; models score F1 0.331–0.655 and taxonomy-guided prompts boost performance up to +0.18. Existing critique pipelines (CrowdCrit, UI Crit) transfer poorly. Remediation experiments show 82.0% of slides improve and 87.8% improve more with taxonomy-based plans, with fixes judged more actionable. Good if you want concrete dataset, metrics, and failure cases.
AI-written summary · about
chapters (9)
- 0:00Introduction to SlideAudit
- 1:06AI and slide design flaws
- 2:10Building the taxonomy
- 3:53Creating the SlideAudit dataset
- 5:19AI flaw detection results
- 6:43Comparing design frameworks
- 8:11Remediation and fixes
- 9:11Future directions and conclusion
- 10:18Q&A on data collection
