OBJECT-ORIENTED APPROACH TO NEURAL NETWORK-BASED DETECTION OF CYBERBULLYING SUBJECTS FROM MESSAGES IN A MANAGED CLOUD ENVIRONMENT

Abstract

The aim of the work is to develop and substantiate an object-oriented approach to neural network detection of cyber- bullying subjects from messages with a combination of primary detection and subsequent syntactic-semantic interpretation in a managed cloud environment. A coherent architecture is proposed in which the neural network module filters messages at the “cyberbullying / non-cyberbullying” level, after which the results undergo dependency analysis with the reconstruction of role relationships of the “subject – action – object” type. An object model, which includes classes of messages, sentences, tokens, dependencies, predicates, role triples and summary structures, is defined as the basis for ensuring transparent traceability of decisions, while a managed cloud environment ensures the reproducibility of launches and scalability of experiments. The effectiveness of the initial detection was experimentally confirmed: the BERT-based module demonstrated a metric of F1 = 0.98 in a two-class setting (“cyberbullying” / “not cyberbullying”), which indicates a sufficient level of screening out irrel- evant messages before role analysis. Consistent indicators of the quality of role identification were established on the expert-ver- ified subset: for the subject, values of 0.88, 0.86, 0.87 were obtained for Precision, Recall and F1, respectively; for the object – 0.85, 0.83, 0.84; for the verb center – 0.91, 0.89, 0.90. The exact restoration of the role triple provided a value of F1 = 0.76. The inter-expert agreement was Cohen’s coefficient of 0.82 with 87 % complete agreement, which indicates the reliability of the reference labels and the correctness of the applied evaluation procedure. In controversial cases, a third auxiliary assessment by a language model with a fixed instruction was used; the final labels were determined by majority rule. The results obtained demonstrate that the proposed approach not only allows to determine the fact of aggressive commu- nication, but also provides a structured presentation of information about its addressability, remaining reproducible and audit- able in practical conditions. Conclusions are drawn that create a basis for further integration of the approach into moderation systems and possible expansion to corpora with more detailed role markup and multilingual support.