Multilingual-E5 with RoBERTa-base for AI-Generated Vietnamese News Detection
Overview
This repository hosts the implementation of a hybrid model that combines Multilingual-E5 embeddings with a RoBERTa-base classification head to distinguish between human-authored and AI-generated Vietnamese news articles.
Developed as part of the research published in Computational Intelligence in Engineering Science (Springer CCIS, vol. 2587), the model achieves a classification accuracy of over 99%, offering a reliable tool for combating misinformation and enhancing journalistic integrity in the Vietnamese context.
By leveraging the semantic richness of Multilingual-E5 and the optimized pre-training of RoBERTa-base, the model effectively captures subtle linguistic and stylistic differences. Training was performed on a balanced dataset of 200,000 articles:
- 100,000 human-written texts sourced from reputable outlets (Thanh Niên, VnExpress)
- 100,000 AI-generated texts produced by advanced large language models (LLMs) such as GPT-4o Mini, Gemini Flash 1.5, Llama 3.3, and DeepSeek
Citation
If you use this model or dataset, please cite the following paper:
@InProceedings{10.1007/978-3-031-98170-8_11,
author = {Huynh, Minh-Phuc and Nguyen, Hoang-Anh and Le, Anh-Cuong and Truong, Dinh-Tu},
title = {Detecting AI-Generated Vietnamese News Articles with Multilingual-E5 and BERT},
booktitle = {Computational Intelligence in Engineering Science},
year = {2026},
publisher = {Springer Nature Switzerland},
address = {Cham},
pages = {130--144},
isbn = {978-3-031-98170-8}
}
Contact
For questions or clarifications regarding the dataset or evaluation procedure, please contact Lê Anh Cường at [email protected]