Main Page

This demo features the Wikipedia, with the text colored according to trust. Text on white background is fully-trusted text; the more orange the background, the lower the value of trust. Type a page title in the search box on the left to see the page trust-colored. You can also try your luck clicking on random page, or look at the following examples:

• A comparison of the Italian and French Cuisine
• The Reaganomics page after an opinionated edit
• Tampering with the Danish Prime Minister's last name (scroll down to see)
• Bogus information added on Cheerleading
• Some are naturals on the beach, others in outer space. Notice how the text gains trust once it is revised (click on "Newer revision" several times to see it)
• An example of how untrusted text can quickly gain trust as it is revised (click on "Newer revision" several times to see it)
• When vandalism is undone, the text regains its original trust (click on "Newer revision" to see it)
• Who would have thought that the sleep paralysis article would have been subject to so much vandalism?
• It's not only anonymous authors who can do vandalism
• Some vandals delete all article content. Once the content is reinserted, it has the trust value prior to deletion: we track the trust both text currently in the article, and of text that used to be in the article, but has been deleted. Click on "Newer revision" to see it.

Once on a page, click on history, click on some past revision, and go back and forth in time, watching how the text trust changes.

How is the trust computed?

The trust of text is computed in two steps:

• First, we compute the reputation of each author by analyzing the author's contributions. When an author makes a contribution that is preserved in subsequent edits, the author gains reputation. When an author makes a contribution that is undone or reverted quickly, the author loses reputation.

• The trust value of a new word is proportional to the reputation of its author. When subsequent authors edit the page, words that are left unchanged gain trust: by leaving them there, the authors implicitly agree with them. Words closer to the edit gain more trust, as the author of the edit is likely to have paid more attention to them. In contrast, text that has been rearranged (new text, text at the border of cut-and-paste, etc) has again a reputation proportional to the author of the edit.

The details of this procedure are described in two papers on reputation and trust computation. You can find more details, as well as the source code for wikitrust, on the UCSC Wiki Lab page. The real algorithms are somewhat more involved than described above, and keep track correctly of reversions, deleted text that is reinserted, and much more. Read the papers for detailed information.

What is in the demo? Is it up to date? Is it complete?

The demo contains all the articles as of the February 6, 2007 snapshot of the Wikipedia, the last for which we have a complete dump. Our code in fact works on Wikipedia dumps: we are working towards adapting it to process a live edit feed.

For each article, we compute the text trust of all the revisions, but to save space, we have loaded in this demo at most 100 revisions for each article.

Support

This work is supported by:

• CITRIS: Center for Information Technology Research in the Interest of Society, a University of California research center.
• [http://institutes.lanl.gov/isti/issdm/ Institute for Scalable

Scientific Data Management], an educational collaboration between LANL and the [http://www.ucsc.edu/ University of California Santa Cruz (UCSC).]

The project is part of the Wikipedia Quality Initiative supported by the Wikimedia Foundation.

Feedback and additional information

• For additional information, source code, and more, go to the UCSC WikiLab page.

• For general topics and discussion of Wikipedia Quality, please join the WikiQuality mailing list. Many Wikipedia people subscribe to this.

• For discussion of specific points of the demo (pages that seem to be improperly colored, code issues, etc) you can join and post to the Wiki Trust mailing list. The list is read mostly by WikiTrust developers.

• For code bug reports, use the wikitrust issue tracker.