DESIGNING OF SOFTWARE ARCHITECTURE BASED ON OBJECT-ORIENTED APPROACH FOR IMPLEMENTATION OF NEURAL NETWORK ALGORITHM FOR DETECTING ETHNIC HATE IN MESSAGES

Abstract

The article considers a modern approach to neural network detection of ethnic hatred in digital text content in conditions of large, constantly growing amounts of data. It is proposed to use machine learning and natural language processing technologies for automatic detection of ethnic hatred in digital texts. The algorithm developed in the work automates the process of analyzing publications on the network to detect ethnic hatred and conflicting statements between ethnic groups. It consists of two stages: pre-processing of the text (removing stop characters and emoticons) and prediction using the FastForest model, which converts the text into a numerical sequence for further analysis. The algorithm generates a percentage estimate of the probability of the presence of ethnic hatred in the text, which facilitates the process of detecting and examining such manifestations. A pipeline for processing and classifying text data is proposed, which includes converting text messages into numerical features and combining them into a vector of characteristics for further analysis. The target variable is converted into numeric keys, after which the OneVersusAll method and the FastForest algorithm are used to train the model on each class separately, using an ensemble of four decision trees, which increases accuracy and reduces the risk of overtraining. After training, the model classifies new content for the presence or absence of ethnic hatred. The software consists of several subsystems: data processing, ethnic hatred detection, data preprocessing, as well as auxiliary subsystems for forming the training sample and training the model. The subsystem forms a data set for training the classifier, and the model training subsystem is responsible for its configuration and storage. The subsystems were integrated into a single system to automate the detection of ethnic hatred in digital content. The algorithm’s effectiveness was assessed using metrics such as MicroAccuracy, MacroAccuracy, LogLoss, ConfusionMatrix, f1-measure, and Recall. The results showed a high level of accuracy: MicroAccuracy 0.9890, MacroAccuracy 0.9889, LogLoss 0.0463, with high Precision and Recall values for both classes. The model successfully classified the majority of cases, predicting 4,854 out of 4,918 cases with ethnic hatred. The results open up new opportunities for creating safer online environments, which is an important step in combating discrimination and hatred in the digital space.