INFORMATION TECHNOLOGY FOR DETECTION OF DISINFORMATION SOURCES AND INAUTHENTICAL BEHAVIOR OF CHAT USERS BASED ON NLP AND MACHINE LEARNING METHODS

Context. In the modern digital environment, the spread of disinformation and inauthentic behaviour of users in chat rooms poses a serious threat to society. Natural language processing and machine learning methods offer effective approaches to detecting and countering such threats.Objective of the study is to develop information technology for automatically detecting the spread of sources of Ukrainian-language fake news and inauthentic behaviour of chat users, which is built using natural language processing methods and implemented, based on machine learning technologies.Method. To implement the project, such feature construction methods as the TF-IDF statistical indicator, the Bag of Words vectorization model, and part-of-speech mark-up were used. For other experiments, the FastText, W2V, and Glove word2vecvectorization models were used to obtain vector representations of words, as well as to recognize trigger words (reinforcing words, absolute pronouns, and “shiny” words). The idea is to find similar messages in terms of text/meaning (lexical/semantical), as well as analyse the results of the distribution of similar messages in time and space. Complement Naïve Bayes, Gaussian Naïve Bayes, HistGradientBoostingClassifier, MultinomialNB and Random Forest were used as the main modelling algorithms to identify sources of disinformation and inauthentic chat behavior.Results. This article discusses the development of software for detecting propaganda messages in social networks based on the analysis of Twitter text data. The main attention is paid to the methods of text pre-processing, data vectorization and machine learning for message classification. The process of collecting, preparing and cleaning data is described, and various approaches to training the model and evaluating its effectiveness are considered. 9 experiments were conducted for the selected methods of postprocessing data, vectorization models and modelling algorithms.Conclusions. The created models show excellent results in recognizing sources of propaganda, fakes and disinformation in social networks and online media. The best results so far are shown by experiment 5 on the main TF-IDF + Complement Naïve Bayes. The high recall value for class 1 (0.8) means that the model finds positive samples well, but for class 0 it is less effective (0.56). The correspondingly high precision value for class 1 (0.89) means that most of the samples predicted as class 1 are correct. The low precision for class 0 (0.38) indicates a large number of false predictions. At the same time, certain anomalies are observed in the series of experiments (in particular, in experiment 7 based on Glove + Random Forest), which require further research. The results obtained can be used to further improve the algorithms for detecting sources of disinformation, inauthentic chat behaviour and malicious content to increase the country’s transparency.