Linguistic and Statistical Traits Characterising Plagiarism

This paper investigates the problem of distinguishing between original and rewritten text materials, with focus on the application of plagiarism detection. The hypothesis is that original texts and rewritten texts exhibit significant and measurable differences, and that these can be captured through statistical and linguistic indicators. We propose and analyse a number of these indicators (including language models, syntactic trees, etc.) using machine learning algorithms in two main settings: (i) the classification of individual text segments as original or rewritten, and (ii) the ranking of two or more versions of a text segment according to their “originality”, thus rendering the rewriting direction. Different from standard plagiarism detection approaches, our settings do not involve comparisons between supposedly rewritten text and (a large number of) original texts. Instead, our work focuses on the sub-problem of finding segments that exhibit rewriting traits. Identifying such segments has a number of potential applications, from a first-stage filtering for standard plagiarism detection approaches, to intrinsic plagiarism detection and authorship identification. The corpus used in the experiments was extracted from the PAN-PC-10 plagiarism detection task, with two subsets containing manually and artificially generated plagiarism cases. The accuracies achieved are well above a by chance baseline across datasets and settings, with the statistical indicators being particularly effective.