W4A 2010 - Web Not For All: A Large Scale Study of Web Accessibility

18 pages
4 views

Please download to get full document.

View again

of 18
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Share
Description
1. Web Not For All A Large Scale Study of Web Accessibility Rui Lopes1, Daniel Gomes2, Luís Carriço1 1 LaSIGE, University of Lisbon 2 FCCN 2. Context ã The Web is the…
Transcript
  • 1. Web Not For All A Large Scale Study of Web Accessibility Rui Lopes1, Daniel Gomes2, Luís Carriço1 1 LaSIGE, University of Lisbon 2 FCCN
  • 2. Context • The Web is the biggest information source for Mankind. Decentralised architecture made it blossom. • Humans (and computers!) contribute to information production and consumption, leading to ~45B Web pages.
  • 3. Context • Growth of users contributing and interacting with the Web leads to significant diversity of users, including people with disabilities. • The openness and decentralisation of the Web leads to an uncontrolled quality check of Websites’ usability (and accessibility).
  • 4. What is the state of accessibility on the Web?
  • 5. • It is known that Web accessibility adequacy is often far worse than desired. • Studies tend to focus on a restricted (small) set of Web sites. • Do macroscopic properties of Web accessibility emerge from analysing at a large scale?
  • 6. Experiment background • The Portuguese Web Archive initiative periodically crawls contents from the Portuguese Web (.pt and others) for future preservation. • Services are built on top of crawled collections: search (end users) & analysis framework (researchers).
  • 7. Methodology data acquisition - obtaining the document collection • Collect a sufficiently large portion of the Web, yet representative (e.g., national Webs) • Spider traps handled gracefully • Boostraped with 200,000 Website addresses from the .pt TLD • Collected March/May 2008
  • 8. Methodology data acquisition - evaluation process • Implementation of 39 WCAG 1.0 checkpoints yield pass, fail, warn. (collection previous to WCAG 2.0 TR) • Overcome computational effort with Hadoop cluster, streams, caching, etc.
  • 9. Methodology data analysis • Failure rate, 3 criteria:
  • 10. Results general • 28M Web pages were evaluated. (58%) • 21GB evaluation data collected for analysis. • 40B HTML elements evaluated. (~1500/page) • 1.5B elements passed. (56/page, 3.89%) • 2.9B elements failed. (103/page, 7.15%) • 36B elements warned. (1291/page, 89%)
  • 11. Results rates versus page count distribution conservative optimistic strict
  • 12. Results rates versus page complexity (# HTML elements) conservative optimistic strict
  • 13. Discussion on the results • Large scale confirms predictions of small scale studies - the Web is still not for all. • Smaller Web pages tend to have greater accessibility quality. • Nature of warnings is more striking than expected, completely different interpretations. • Automated evaluation is just the beginning.
  • 14. Discussion on the limitations of the experiment • HTML structure vs. content rhetorics. (CSS & Javascript can change it all) • Collecting the Web is hard. (deep Web - AJAX & forms -, infinite generation, robots.txt, etc.) • Scaling evaluation & analysis processes is hard. (evaluation streamability, resource inter-dependencies, billion node graphs, etc.)
  • 15. Conclusions • Large scale accessibility evaluation of the Portuguese Web. • Re-confirmed studies at the large. • Educating developers & designers about warnings is crucial for accessibility success! • Automated evaluation is just the start. Always need for expert & users evaluations.
  • 16. Ongoing Work we’re still at the tip of the iceberg • Linking properties (ranking vs. accessibility) • Evolution of accessibility compliance in time (different document collections) • Cross-cuts: gov, e-com, personalisation, etc. • Developing countries countries) (Portuguese speaking African
  • 17. Ongoing Work help wanted from community! • Making available evaluation datasets (e.g., Linked Data). Ours and yours! • Larger document collections. • Transforming warnings into failures with machine learning.
  • 18. Thank you! rlopes@di.fc.ul.pt
  • Related Search
    We Need Your Support
    Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

    Thanks to everyone for your continued support.

    No, Thanks