A Systematic Comparison of Computational Models for Detecting Adolescent Vulnerability in Bengali Text

The internet has replaced the life of teenagers and the software installed to protect the young users on the internet do not understand any other language other than English. This becomes a huge population of teenagers who are native speakers of the Bengali language and who lack the appropriate protection against the risks of cyberbullying and harassment. This digital safety gap is explicitly filled in the study as it forms and critically evaluates computational models that detect signs of vulnerabilities in Bengali text. The current paper compares three levels of techno-based classification using a custom-built dataset of 5,209 manually annotated social media comments of traditional machine learning classifiers, deep neural networks (including CNNs and LSTMs) and state-of-the-art Transformer-based models. We observe a clear hierarchical improvement in performance across model categories. The deep learning models offer a meaningful addition to the more traditional methods, though the latter serves as a decent baseline. Transformer architectures, however, proved to be in a league of their own. The most efficient was the XLM-RoBERTa with fine-tuning and the remarkable F1-score of 98.63%. Such level of accuracy demonstrates that it is not only possible but quite possible to create highly consistent, automated systems of monitoring and flagging dangerous material in Bengali. This research establishes a significant precedent and gives a clear path of how to transform the Internet into a safer zone to the millions of young people who use it.