In this paper, we describe our approach at the PAN 2012 plagiarism detection competition. Our candidate retrieval system is based on extraction of three different types of Web queries with narrowing their execution by skipping certain passages of an input document. We have created queries based on keywords extraction, intrinsic plagiarism detection and headers extraction. We have also compared the performance of constructed queries used during the PAN 2012 test process. The proposed methodology was the best performing one in case of long term operation and also the most cost-effective one. Our detailed comparison system is based on detecting common features of several types (in the final submission, we have used two types of features: sorted word 5-grams and unsorted stop word 8-grams) in the input document pair. We propose a method of computing so called valid intervals from those features, represented by their offset and length attributes in both source and suspicious document. Previous works use the feature ordering as the measure of distance, which is not usable for multiple types of features, which do not have any natural ordering. From those valid intervals we compute final detections in the post-processing phase, where we merge neighbouring valid intervals and remove some types of overlapping detections. We further discuss other approaches which we explored, but which have not been used in our final submission. In the paper we also discuss the performance aspects of our program, parameter settings, and the relevance of current PAN 2012 rules (including the plagdet score) to the real-world plagiarism detection systems.
展开▼
