The object of this study is the processes of identifying sources and networks of disinformation dissemination in the cyberspace of the world. With the growing influence of social networks on public opinion, the issue of identifying and neutralizing propaganda messages is becoming particularly relevant. Conventional methods of combating propaganda such as manual content moderation have proven to be insufficiently effective due to the large amount of information generated daily. It is important to use natural language processing and machine learning methods to analyze text, identify sources of disinformation dissemination and inauthentic behavior of bots. Based on the analysis of existing methods of intelligent disinformation search, methods have been devised to identify sources and ways of disinformation dissemination in cyberspace by searching for similar text chains and analyzing the similarity of writing style. Hybrid vector representation makes it possible to capture surface frequency characteristics of the text and semantic features, which has a positive effect on the quality of classification. Cosine similarity, Jacquard, Levenstein and Word2Vec are used to measure similarity. Clustering (DBSCAN, K-Means) helps group fake messages. Graph analysis detects central accounts and bot networks. Evaluation of the model’s performance by key metrics showed reliable results for identifying sources of disinformation distribution: accuracy – 0.82, F1.3 – 0.8, ROC-AUC – 0.86. The identified differences in lexical patterns for the “fake” and “true” classes confirm the model’s ability to capture the content features of texts. The proposed method for detecting disinformation distribution paths serves as the basis for building scalable systems for monitoring the information space and adapting to other text classification tasks
Vysotska et al. (Fri,) studied this question.