Specialists from the Moscow Technical University of Communications and Informatics (MTUCI) have developed a service that automatically identifies potentially dangerous, unreliable, and undesirable information in online publications. As RIA “Novosti” was informed by the university's press service, it is based on machine learning technologies and modern methods of text data processing.
The system architecture is three-tiered. The first level is dictionary-based search, which matches text with keywords and provides basic filtering. The second is morphological analysis with lemmatization, which allows for different word forms to be taken into account without having to include all possible variants in the dictionary. The third, intelligent contour, is implemented on a pre-trained transformer-based model – it provides high classification accuracy and is capable of capturing context, not just trigger words.
The advantage of the development is that it combines classical linguistic methods with a neural network approach. A simple dictionary filter will miss malicious content if the author rephrases a phrase, while a transformer model analyzes semantics and can recognize dangerous meaning even in a veiled form. Researchers have already implemented key system modules: user authentication and authorization, dictionary management, logging, and analytical components.
Plans include further improvement of machine learning models, expansion of the dictionary, and increased classification accuracy.
In fact, MTUCI is creating not a one-time moderation tool, but a self-learning platform that will adapt to new threats. Unlike static filters, which attackers sooner or later bypass, a system with a transformer model is capable of evolving with the threat landscape.
