How (Much) to Trust Wikipedia

February 20, 2008: 12:00pm - 1:00pm
Location: 290 Hearst Memorial Mining Building, the Maria & Dado Banatao Conference Room, UC Berkeley

Luca de Alfaro [Associate Professor of Computer Engineering, UC Santa Cruz]

Part of the CITRIS Research Exchange at UC Berkeley. The complete schedule for the fall semester is online at RE-Spring2008. Sponsored by Infineon Technologies.

Abstract:

The Wikipedia is a collaborative encyclopedia: anyone can contribute to its articles simply by clicking on an "edit'' button. The open nature of the Wikipedia has been key to its success, but has a flip side: if anyone can edit, how can readers know whether to trust its content?

To help answer this question, we have developed a reputation system for Wikipedia authors, and a trust system for Wikipedia text. Authors gain reputation when their contributions are long-lived, and they lose reputation when their contributions are undone in short order. Each word in the Wikipedia is assigned a value of trust that depends on the reputation of its author, as well as on the reputation of the authors that subsequently revised the text where the word appears. To validate our algorithms, we show that reputation and trust have good predictive value: higher-reputation authors are more likely to give lasting contributions, and higher-trust text is less likely to be edited.

The trust can be visualized via an intuitive coloring of the text background. The coloring provides an effective way of spotting attempts to tamper with Wikipedia information. A trust-colored version of the entire English Wikipedia can be browsed at http://trust.cse.ucsc.edu/