PAN-KK: A Language Resource for Plagiarism Detection in Low-Resource Kazakh

To the best of our knowledge, we introduce the first dataset, which was created with the intention of extrinsic plagiarism detection in Kazakh. The PAN-KK was developed through translation of over 20,000 pairs of plagiarised text in English that were in PAN 2010 and it includes a human-verified subsample of 2,000 samples, each being scrutinised by 10 native Kazakh speakers. We prepare and evaluate a range of different transformer-based models using PAN-KK and the most successful model is XLM-RoBERTa Large. As an example of practical use, we next use XLM-RoBERTa Large in a two-stage pipeline which integrates candidate retrieval with semantic verification and span localisation. In 1000 pairs that are manually annotated, the localisation of the tokens is P=0.929, R=0.843, F1=0.868, and the localisation of the spans is P=0.718, R=0.649, F1=0.667. In general, the aid of PAN–KK and its pipeline facilitates the development of the field of plagiarism detection in low-resource languages such as Kazakh.