Adaptive AI-Based Plagiarism Detection System

Traditional plagiarism detection systems normally compute a separate similarity score applying a corrected combination of detection techniques, regardless of the type of plagiarism presence. This approach performs poorly when the nature of text-based overlap varies; for instance, verbatim copying produces high scores across all word-related measures, whereas rephrasing produces low word-related similarity but high meaning-related similarity. This paper presents an adaptive plagiarism detection system that first computes four unaffiliated similarity measures; exact sequence correspondence, TF-IDF cosine similarity, n-gram overlap, and Sentence-BERT meaning-related similarity; then classifies the anticipated plagiarism pattern utilizing a customizable, rule-based classifier, and ultimately re-weights the four measures according to the detected pattern before computing a last and ultimate adaptive score. The system moreover offers sentence-level evidence and a traceable, non-templated explanation.