VietPlagNet-X: An Explainable Multimodal Architecture for Text-Reuse Detection and Academic Plagiarism-Risk Assessment in Vietnamese

Academic plagiarism detection is increasingly difficult amid digital publishing, the rapid advance of large language models (LLMs), and automated paraphrasing tools. Commercial tools such as Turnitin and iThenticate provide similarity indices useful for screening, but a similarity score is not equivalent to a plagiarism verdict, and effectiveness varies with text type, language, and concealment strategy (Milovanovic et al., 2025; Rahman et al., 2024). This paper presents VietPlagNet-X, an explainable, multimodal architecture for text-reuse detection and plagiarism-risk assessment in Vietnamese academic documents. The architecture comprises a structure-, OCR-, citation-, and entity-aware preprocessing pipeline; five core evidence branches (lexical, semantic, cross-lingual, citation/context, structural- metadata); and two auxiliary signals, namely AI-generated-text likelihood and internal stylistic deviation. The auxiliary signals are not fed directly into the plagiarism-risk score but instead support human review. Every architectural choice is grounded in a structured review of 26 sources (2011-2025), rather than treating a similarity score as an automatic verdict. The lexical and semantic branches have been implemented and piloted on two public Vietnamese datasets (ViSTS-100, 100 sentence pairs; ViSP-pairs, 84 sentence pairs). On ViSP-pairs, the semantic branch achieved ROC-AUC = 1.0 and F1 = 1.0 on a controlled paraphrase-discrimination task; this should not be read as plagiarism-detection performance. On ViSTS-100, semantic_score achieved Pearson r = 0.6852, lower than tfidf_score (0.7802), indicating that the value of semantic evidence depends on dataset characteristics and lexical overlap. The paper treats these findings as initial technical-feasibility evidence and identifies the steps required to validate the full system on a large-scale Vietnamese plagiarism corpus.