APPROACH TO SOFTWARE ENGINEERING AND TESTING OF THE METHOD OF MULTI-LEVEL DETECTION OF CYBERBULLYING SUBJECTS BASED ON TRANSFORMERS IN A CLOUD ENVIRONMENT

Abstract

The article presents an engineering-based approach to multi-level detection of cyberbullying subjects with a combination of primary transformer detection of aggressive statements and subsequent role-based interpretation of discourse according to the initiator speech act addressee scheme in the Google Colab environment. The implementation is based on a pre-trained classifier cardiffnlp twitter roberta base offensive in the text classification configuration of the transformers library, which provides reproducible inference without additional training and allows you to focus on software engineering and testing. The proposed design includes interface modules, processing control, classification and syntactic semantic analysis with fixing package versions, random seeds and tokenization parameters, configuration logging and artifact storage on Google Drive. The methodological contribution consists of a notebook-oriented organization of the life cycle, which includes data format control, replicated microruns, a thresholding protocol for binary decisions, borderline case labeling for manual review, and built-in checks for consistency of role interpretation. Empirical feasibility is confirmed on a small expert-verified subsample, which covers typical situations of direct insult, ironic devaluation, and neutral citation of potentially toxic vocabulary. The evaluation includes fixing the median inference time per example and the interquartile range as indicators of operational stability in variable conditions of the collab environment, as well as analyzing the distribution of decision confidence taking into account the operating threshold. Qualitative validation demonstrates correct activations on explicit invectives, false positives in cases of metalinguistic citation, and false negatives for sarcastic statements, which justifies the need for role interpretation to remove contextual ambiguity. The results obtained indicate that the combination of a pre-trained classifier with a role level and procedurally designed testing provides reproducibility of experiments, interpretability of output, and controlled determinism of solutions within the computational constraints of Colab, creating a basis for scaling to larger corpora and multilingual domains.