Quality Control Based On User Behavior Analysis In Online Collaborative Editing

Online collaborative editing projects have already become an important kind of information resources as the tendency of referencing and citing contents from their websites spreads fast and widely. However, the reliability of these user generated contents is subject to the behaviors of online collaborative editors, since the openness and the natural tolerance of opinion diversity attribute to the dynamic and uncertain process of knowledge producing. Considering the connection between editor behaviors and the outcome quality of online collaborative editing projects, researchers are paying more attentions on how web users act when involved in online collaborative editing, trying to control and manage the information quality of collaborative editing outcomes through applying the patterns and features of those collaborative editing behaviors appropriately.In this thesis, we explored the human factors that decides the outcome quality of online collaborative projects and the quality control approaches based on them which includes:evaluation on the outcome quality of online collaborative editing projects, modeling the human factors in online collaborative editing systems, the relationship between human factors and the quality of collaborative editing outcomes, the process of eliminating the disagreement and approaching the consensus among different user opinions, the patterns and features of web users' collaborative editing behaviors, automatic vandalism detection in online collaborative projects, etc. Especially we proposed several innovations focusing on the following aspects:1. Evaluation and performance analysis of an online collaborative project are never easy tasks because the massive human involvement and other qualitative factors are hard to assess. To figure out the relationship between human related factors and quality of collaboration outcomes, we propose an effective formal approach to estimate the human involvement in collaboration process and testify our method on100articles extracted from Wikipedia and Scholarpedia, the qualities of whose historical contents have been evaluated manually by volunteers of specific background. Through comparison of the human involvement and the outcome quality in these two projects, we find that the quality of collaborative products is positively related with the number, dedication and experience of collaboration participants, among which experience decides the necessary amount of human resources for a high quality, and increasing number or dedication of participants can also make up to the lack of experiences.2. For online collaborative editing projects, their openness and tolerance of disagreements are obstacles for their expectation of an overall consensus. To figure out what the key is to make this paradoxical mechanism work and whether consensus is achievable or not for those collaborating groups, we explored202.472historical content states for73articles in Wikipedia, traced the "evolution" processes of article contents, and analyzed the collaborating behaviors of the contributors from the point of content editing tendency. Finally we found that Wikipedia users tend to generalize their own expression of consensus and avoid duplicating contents from outside resource, and during their editing process, the ubiquitous initiative to approach consensus, as well as the neutral deliberation on dissent are two essential factors for collaborative communities like Wikipedia to success.3. The open contents in online collaborative editing projects are vulnerable against vandalism. The current vandalism detection methods relying on basic statistic language features work well for explicitly offensive edits that perform massive changes. However, the recall of these anti-vandalism techniques remains relatively low as they are evadable for the elusive vandal edits which make only a few unproductive or dishonest modifications. In this paper, we proposed a contributing efficiency-based approach to detect the vandalism in Wikipedia. We testified the negative correlation between the probability of a revision being reverted as vandalism and the contributing efficiency of this revision, and evaluated the performance of an SVM classifier that incorporates the contributing efficiency along with other languages features. The results of extensional experiment on196,491tagged revisions show that the contributing efficiency can improve the recall of machine learning-based vandalism detection algorithms significantly.