Investigating Advanced Techniques for Document Content Similarity Applied to External Plagiarism Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 240–246, Hissar, Bulgaria, 12-14 September 2011. Investigating Advanced Techniques for Document Content Similarity Applied to External Plagiarism Analysis Daniel Micol and Rafael Mu˜noz Dept. of Software and Computing Systems University of Alicante San Vicente del Raspeig, Alicante, Spain {dmicol,rafael}@dlsi.ua.es ´Oscar Ferr´andez Dept. of Biomedical Informatics University of Utah Salt Lake City, Utah, USA oscar.ferrandez@utah.edu Abstract We present an approach to perform exter- nal plagiarism analysis by applying sev- eral similarity detection techniques, such as lexical measures and a textual entail- ment recognition system developed by our research group. Some of the least expen- sive features of this system are applied to all corpus documents to detect those that are likely to be plagiarized. After this is done, the whole system is applied over this subset of documents to extract the ex- act n-grams that have been plagiarized, given that we now have less data to pro- cess and therefore can use a more complex and costly function. Apart from the ap- plication of strictly lexical measures, we also experiment with a textual entailment recognition system to detect plagiarisms with a high level of obfuscation. In addi- tion, we experiment with the application of a spell corrector and a machine translation system to handle misspellings and plagia- risms translated into different languages, respectively. 1 Introduction We believe there are two main user scenarios where external plagiarism detection tools are ap- plied, sharing both of them the fact that they have a large source documents corpus. The difference, however, is that the first scenario is based on a large number of suspicious documents being pro- cessed at the same time, so the detection approach needs to be highly efficient and scalable. An ex- ample of this scenario would be the 1st and 2nd International Competitions on Plagiarism Detec- tion (Potthast et al., 2009; Potthast et al., 2010), where the corpora contain multiple source and suspicious documents. For this first use case we have developed a system to detect external docu- ment plagiarism that is highly efficient and scal- able. It contains a first phase where a small sub- set of source documents are selected as possible candidates to be the origin of the plagiarism for a given suspicious document. Given that this phase processes the whole corpora, it uses a simple and lightweight function to select the subset of can- didate source documents. After this is done, a more complex function is applied over this sub- set to extract which documents contain the pla- giarism, and the exact position within these docu- ments. This two-step approach is common among research systems, as described in (Potthast et al., 2009). The second use case assumes that we only have to process one suspicious document at a time. Therefore, we can apply more complex techniques that are less efficient but highly accurate, as there is less data to process. An example of this use case could be an online system to detect if a sci- entific …

Content truncated to 500 words