Automatically Assessing the Need of Additional Citations for Information Quality Verification in Wikipedia Articles

Abstract

Quality flaws prediction in Wikipedia is an ongoing research trend. In particular, in this work we tackle the problem of automatically assessing the need of including additional citations for contributing to verify the articles’ content; the so-called Refimprove quality flaw. This information quality flaw, ranks among the five most frequent flaws and represents 12.4% of the flawed articles in the English Wikipedia. Underbagged decision trees, biased-SVM, and centroid-based balanced SVM –three different state-of-the-art approaches– were evaluated, with the aim of handling the existing imbalances between the number of articles’ tagged as flawed content, and the remaining untagged documents that exist in Wikipedia, which can help in the learning stage of the algorithms. Also, a uniformly sampled balanced SVM classifier was evaluated as a baseline. The results showed that under-bagged decision trees with the min rule as aggregation method, perform best achieving an F1 score of 0.96 on the test corpus from the 1st International Competition on Quality Flaw Prediction in Wikipedia; a well-known uniform evaluation corpus from this research field. Likewise, biased-SVM also achieved an F1 score that outperform previously published results.