INFORMATION TECHNOLOGY FOR IDENTIFYING THE UKRAINIAN-LANGUAGE TEXT CONTENT TOXICITY ON YOUTUBE AND GOOGLE PLAY BASED ON MACHINE LEARNING METHODS
- 1 April 2026
- journal article
- Published by Lviv Polytechnic National University in Information and communication technologies, electronic engineering
- Vol. 6 (1) , 64-74
Abstract
The article presents the results of the development and experimental study of specialised information technology for automatically identifying the toxicity of Ukrainian-language text posted on YouTube and Google Play. The relevance of the work stems from the rapid growth of user comments in the digital environment, which is accompanied by the spread of aggressive, discriminatory, and hostile statements that negatively affect the quality of online communication and users' psychological comfort. Particular attention is paid to the shortage of marked corpora in Ukrainian and the limited applicability of pre-trained transformer models without additional adaptation to the morphological and semantic specifics of the Ukrainian segment. The purpose of the study is to design and implement a universal machine learning pipeline for binary classification of text messages based on toxicity, accounting for the peculiarities of Ukrainian-language digital discourse. To form a training sample, an open dataset from the Kaggle platform and the author's collections of comments from YouTube and Google Play were combined, ensuring representativeness and diversity of stylistic constructions. The general corpus contains both overtly aggressive and contextually hidden manifestations of toxic behaviour. Within the proposed information technology framework, a complete data processing cycle has been implemented: preprocessing of text, normalisation, extraction of stop words, vectorisation using TF-IDF, class balancing using the SMOTE method, and hyperparameter optimisation using GridSearchCV. A comparative analysis of six machine learning models was conducted: Logistic Regression, SVM with RBF kernel, Linear SVC, SGD Classifier, Random Forest, and Gradient Boosting. Performance was evaluated using the Accuracy, Precision, Recall, F1-score, and AUC-ROC metrics. Experimental results showed that linear models perform competitively in classifying high-dimensional, sparse text vectors. In particular, Linear SVC and SGD Classifier provided the best balance between accuracy and completeness, while ensemble methods showed higher accuracy by reducing sensitivity to toxic examples. The results confirm the feasibility of using classical machine learning algorithms under limited-resource conditions for the Ukrainian language and serve as a baseline for further integration of deep and multimodal approaches. The proposed information technology can be used as a component of automated real-time content moderation systems, thereby increasing information security and creating a safer online environment for Ukrainian-speaking users.Keywords
This publication has 7 references indexed in Scilit:
- Toxicity Classification in Ukrainian2024
- System Project for Ukrainian-language Feedback Tonality Analysis in the Health Care Field Based on BERT Model2023
- Toxic Text Classification2020
- A Machine Learning Approach to Comment Toxicity Classification2019
- Vocabulary-based community detection and characterization2019
- Convolutional Neural Networks for Toxic Comment Classification2018
- A Review of Standard Text Classification Practices for Multi-label Toxicity Identification of Online Content2018