Machine Translation Evaluation Metric for Text Alignment

As plagiarisers become cleverer, plagiarism detection becomes harder. Plagiarisers will find new ways to obfuscate the plagiarized passages so that humans and automatic plagiarism detectors are not able to point them out. So, a plagiarism detection system needs to be robust enough to detect plagiarism, no matter what obfuscation techniques have been applied. Our system attempts to do the same by combining two different methods, one stricter to catch mildly obfuscated passages and one more lenient to catch difficult ones. We use a machine translation evaluation metric and n-gram matching to detect overlaps between the source and suspicious documents. On the PAN’14 corpus, which contains data with various types of obfuscation to mimic human plagiarisers, we obtained plagdet scores of 0.84404 and 0.86806 on the two datasets.