Year: 2025
Type: article
Abstract: Method of detecting cyberbullying in textual content by neural networks means means is proposed, which allows analyzing text message and determining the generalized level of cyberbullying in messages and carrying out multiclass classification, providing separate indicators of the manifestation of individual types of cyberbullying. The method allows to separately determine the manifestation levels of age, religious, ethnic, gender and other cyberbullying types. The effectiveness of the proposed method of detecting cyberbullying in textual content by neural networks means was investigated by developing the appropriate software. Various machine learning models were used for software implementation. For binary classification – BiLSTM and RNN with LSTM layer were used. In the task of binary classification, the BiLSTM model showed the best results with Accuracy 96%, Precision 96%, Recall 95.9%, F1 95.7%. For the task of multiclass classification, Naïve Bayes, SVM Binary, FastForest, BERT, L-BFGS Maximum-EntropyMulti models were investigated. According to the obtained results of macro and micrometrics, BERT turned out to be the best of the models with macrometric indicators Accuracy 94%, Precision 93%, Recall 93%, F1 93%. High results were also shown by the L-BFGS Maximum-EntropyMulti-based model with macrometric indicators Accuracy 89.8%, Precision 92.1%, Recall 95.6%, F1 93.8%. However, after examining the values of micrometrics for each type of cyberbullying in more detail, it was established that the BERT model shows more stable indicators. The BERT model correctly classified 1525 text samples according to age cyberbullying, incorrectly – 27, according to religion correctly classified 1534, incorrectly – 9, according to ethnicity correctly classified – 1463, incorrectly – 9, according to gender correctly classified – 1357, incorrectly 90. Also the preference was given this model, because it supports the concept of zero-shot learning, which allows the model to make predictions for new or poorly represented categories, which is especially relevant for categories of cyberbullying types. Method of detecting cyberbullying in textual content by neural networks means is intended for implementation in social networking platforms, educational platforms, as well as in moderation systems of social platforms to provide psychological assistance to victims or witnesses of cyberbullying. Keywords: cyberbullying detection, cyberbullying, neural networks, BERT, BiLSTM, transformer neural networks.
Authors: Olena Sobko, Oleksandr Barmak
Institution: Khmelnytskyi National University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4410898269