Paraphrase Plagiarism Detection in Student Assignments Using IndoBERT and Two-Level Cosine Similarity

Abstract

Plagiarism remains a serious threat to academic integrity, while most available detection tools operate through literal word matching and are therefore easily defeated by paraphrasing. This study implements the pre-trained IndoBERT model without fine-tuning as a semantic-similarity screening tool for Indonesian-language student assignment documents, using a two-level scheme: a chunk-level similarity score (200 words with a 50-word overlap) as a document-level indicator, and one-to-one sentence alignment as a locator of similar passages. A 768-dimensional embedding is formed through attention-masked mean pooling followed by L2 normalization, and vectors are compared using cosine similarity at a 0.92 threshold. Evaluation was conducted on 20 pairs drawn from 15 documents, of which only 11 were authentic student submissions while 4 were controls derived from a single source document through graded paraphrasing. Similarity scores decreased across the four control conditions ordered by intended paraphrase intensity, namely 100%, 66.67%, 42.34%, and 26.67%, whereas same-theme original pairs averaged 5.93%. Against a proxy sentence-level ground truth, the system obtained 84.79% accuracy, 96.82% precision, 61.13% recall, and a 74.94% F1-score with a 1.19% false positive rate. A User Acceptance Testing exercise with six respondents yielded an index of 82.56%, describing only the participating group rather than general acceptance. These results provide preliminary evidence that the two-level scheme is feasible as an initial semantic-similarity indicator, with the final judgment of plagiarism remaining with the examiner.