<![CDATA[Academic plagiarism is becoming increasingly difficult to detect because it is not only carried out through direct copying but also through paraphrasing that changes sentence structure without altering meaning. This study develops a student assignment plagiarism detection system using Transformer-based Sentence Embeddings and Cosine Similarity. The dataset used consists of 600 Indonesian text pairs manually annotated as ground truth with plagiarism and non-plagiarism labels. The research stages include data collection, selective preprocessing, subword tokenization in the Transformer model, embedding vector formation, Cosine Similarity calculation, threshold determination, and performance evaluation using accuracy, precision, recall, and F1-score. This study compares the baseline TF-IDF method with two Transformer models, namely all-MiniLM-L6-v2 and paraphrase-multilingual-MiniLM-L12-v2. The test results show that the paraphrase-multilingual-MiniLM-L12-v2 model achieved the best performance, with an accuracy of 86.67 percent, precision of 0.83, recall of 0.95, and F1-score of 0.89. Meanwhile, all-MiniLM-L6-v2 achieved an accuracy of 74.67 percent, and TF-IDF achieved an accuracy of 62.00 percent. A threshold value of 0.70 was determined based on the analysis of changes in precision, recall, and F1-score, so that plagiarism decisions were not based merely on assumptions but on quantitative evaluation results. Therefore, the Transformer-based Sentence Embedding approach proved to be more effective than lexical methods in detecting paraphrase-based plagiarism in Informatics Engineering student assignments.]]>
