A Multimodal Model for Detecting Vietnamese Toxic News Using Semi-supervised Learning

The proliferation of the Internet, particularly the rapid growth of social networks and online platforms, has facilitated the widespread dissemination of fake, toxic, and reactionary news content. In the context of Vietnam, reactionary news, such as online articles that disseminate false information or incite division within the national unity bloc, poses a significant societal threat due to its rapid propagation and diverse modalities of expression, including text, images, videos, or multimodal combinations thereof. Given the increasing prevalence and severity of such content in cyberspace, numerous studies have been conducted domestically and internationally to address its detection and mitigation. However, the majority of existing approaches primarily target English-language content. Moreover, the processing of news presented in non-textual formats, such as text embedded in images or videos, remains a major challenge, often resulting in reduced detection performance. To overcome these limitations, this paper introduces a multimodal detection model that integrates semi-supervised learning with the PhoBERT language model and Swin Transformer V2 for identifying toxic news represented in both textual and visual forms. The adoption of semi-supervised learning minimizes the dependence on large-scale labeled datasets while maintaining high detection performance. Experimental evaluations conducted on a dataset comprising 8,000 Vietnamese news articles in text and image formats demonstrate that the proposed multimodal model significantly outperforms both unimodal and previous state-of-the-art methods, achieving an F1-score of 90.99% using only 30% labeled data.