Citation-based plagiarism detection - idea, implementation and evalutation

Currently used Plagiarism Detection Systems solely rely on textbased comparisons. They only deliver satisfying results if the plagiarized text is copied literally (copy&paste), with minor alterations (e.g. shake&paste) or is machine translated. However, if the text is paraphrased or translated by a human, the currently used methods yield a very poor performance. Using the words of Weber Wulff, who organizes regular comparisons for Plagiarism Detection Systems (PDS), the current state of available systems can be summarized as follows: “[...] PDS find copies, not plagiarism.”. In contrast to the existing approaches for Plagiarism Detection, Citation-based Plagiarism Detection compares the occurrences of citations in order to identify similarities. The most basic form is to measure the bibliographic coupling strength (citation overlap). However, this alone would lead to numerous false-positives, thus it is advisable to include further factors such as the order of citations, their proximity to each other, their chance of cooccurrence, and other more sophisticated measures. If e.g. four papers are cited in a similar order in two documents, this can be interpreted as a subtle hint that both works may not have been created independently of one another. If none of these four papers have been co-cited before in another paper or the order of citations is identical, this might indicate plagiarism. The advantages and limitations of Citation-based Plagiarism Detection are very different from those of the currently used textbased methods. Text matching approaches continue to be suitable for detecting copy&paste plagiarism, even for short passages. They are also advantageous in that they do not require citation information; yet, they fail to identify e.g. paraphrased, translated and idea plagiarism. By applying the citation-based approach to the doctoral thesis of Guttenberg, which is a well-examined, real world plagiarism example tested by numerous conventional Plagiarism Detection Systems, we could show that the citationbased approach is able to identify 13 out of the 16 translated plagiarisms. Conventional methods failed to identify any of these sections. However, as expected, short passages of copy&paste plagiarism can usually only be identified by text-based approaches. Therefore, Citation-based Plagiarism Detection is by no means a replacement for the currently used text-based approaches, but should be considered as a complement for identifying currently hard to find well-disguised plagiarisms. Additionally, once signs of plagiarism have been found, neither the text-based approaches, nor the citation-based approaches eliminate the need for manual examination.