cc 14(1): e7

Research Article

Harnessing Context for Vandalism Detection in Wikipedia

Download1084 downloads
  • @ARTICLE{10.4108/cc.1.1.e7,
        author={Lakshmish Ramaswamy and Raga Sowmya Tummalapenta and Deepika Sethi and Kang Li and Calton Pu},
        title={Harnessing Context for Vandalism Detection in Wikipedia},
        journal={EAI Endorsed Transactions on Collaborative Computing},
        keywords={Collaborative Social Media, Vandalism, Content-context, Contributor-context},
  • Lakshmish Ramaswamy
    Raga Sowmya Tummalapenta
    Deepika Sethi
    Kang Li
    Calton Pu
    Year: 2014
    Harnessing Context for Vandalism Detection in Wikipedia
    DOI: 10.4108/cc.1.1.e7
Lakshmish Ramaswamy1,*, Raga Sowmya Tummalapenta1, Deepika Sethi1, Kang Li1, Calton Pu2
  • 1: Computer Science Department, The University of Georgia, Athens, GA 30602, USA
  • 2: College of Computing, Georgia Institute of Technology, Atlanta, GA 30332, USA
*Contact email:


The importance of collaborative social media (CSM) applications such as Wikipedia to modern free societies can hardly be overemphasized. By allowing end users to freely create and edit content, Wikipedia has greatly facilitated democratization of information. However, over the past several years, Wikipedia has also become susceptible to vandalism, which has adversely affected its information quality. Traditional vandalism detection techniques that rely upon simple textual features such as spammy or abusive words have not been very effective in combating sophisticated vandal attacks that do not contain common vandalism markers. In this paper, we propose a context-based vandalism detection framework for Wikipedia. We first propose a contextenhanced finite state model for representing the context evolution ofWikipedia articles. This paper identifies two distinct types of context that are potentially valuable for vandalism detection, namely content-context and contributor-context. The distinguishing powers of these contexts are discussed by providing empirical results. We design two novel metrics for measuring how well the content-context of an incoming edit fits into the topic and the existing content of a Wikipedia article. We outline machine learning-based vandalism identification schemes that utilize these metrics. Our experiments indicate that utilizing context can substantially improve vandalism detection accuracy.