KazNU Scientists Develop AI System to Detect Disinformation on Social Media During Natural Disasters

In today's information society, the problem of disinformation is becoming increasingly acute, particularly on social media, where the speed of information spread is at its highest. During natural emergencies, rapidly spreading disinformation can negatively affect public perception of events, hinder coordination among rescue services, and heighten panic.

Disinformation not only undermines trust in information sources but also influences public opinion, which can have serious social, political, and economic consequences during emergencies. Research estimates suggest that the economic damage from disinformation can reach billions of dollars annually across the finance, healthcare, and public safety sectors. For Kazakhstan, where the risk of earthquakes, mudflows, and wildfires remains high, rapid and accurate information analysis is of particular importance.

To develop technology for detecting disinformation on social media during natural emergencies using machine learning methods. The research focuses on three types of natural disasters: mudflows, earthquakes, and wildfires.

The project is the first to propose creating and annotating a dataset related to natural emergencies in Almaty and the Almaty region, enabling models to be trained with regional specificity in mind. A new hybrid architecture will be developed, combining the strengths of convolutional and recurrent neural networks with an attention mechanism. Particular attention is given to Kazakh as a low-resource language and to the culturally specific aspects of texts, which is especially important for Kazakhstan's multilingual society.

The project comprises five tasks: developing and researching models and algorithms for disinformation detection; forming a dataset of social media posts related to natural emergencies, labeled for reliability; developing and training a hybrid deep learning model with an attention mechanism; evaluating the effectiveness of data labeling and disinformation detection methods; and developing an application that allows users to upload text data, run automated analysis, and view classification results.

The study employs natural language processing (NLP) methods — TF-IDF, word embeddings — as well as LSTM and transformer architectures. Data is collected through web scraping and social media APIs, followed by manual and semi-automated annotation. Model accuracy is assessed using F1-score, Precision, and Recall metrics, with cross-validation used to verify the reliability of results.

The project's results will help strengthen international cooperation in combating disinformation and ensuring media content transparency, reinforcing Kazakhstan's contribution to global information security. Research findings will be published in high-ranking scientific journals.

The developed technology will improve the efficiency of emergency response services by enabling timely detection of disinformation, enhance public information, strengthen trust in official sources, and reduce panic during emergencies. The project's results will also find application in media monitoring, information flow management, and combating fake news.

Registration of copyright certificates and patents for new algorithms and software for disinformation detection; training of specialists; publication of results in international peer-reviewed scientific journals.

Spanning 36 months (2025-2027), the project will generate new knowledge at the intersection of artificial intelligence, information security, and natural language processing. It will strengthen Kazakhstan's public safety during emergencies and enhance the international standing of domestic science in this field.