Estudo de caso do uso de um sistema de gerenciamento de banco de dados para detecção de cyberbullying

Abstract

The ease of access to social media has intensified episodes of cyberbullying among users, a practice that negatively affects victims’ lives. Given this problem, this work aims to evaluate the applicability of PGVector (an extension for creating VDBMSs) in processes for detecting messages potentially associated with cyberbullying, combining semantic similarity search with a supervised machine learning model. To this end, data from the X platform (formerly Twitter) were collected and submitted to a method structured in six steps: textual preprocessing, vector representation, sampling, storage in PostgreSQL with the PGVector extension, classification, and evaluation of results. In the classification step, three embedding generation models were compared — SBERT, BERT, and FastText —, each combined with the kNN algorithm executed directly in the VDBMS. The results showed that SBERT achieved the best performance (k = 9, f1-score of 0.77), outperforming BERT (k = 7, f1-score of 0.65) and FastText (k = 5, f1-score of 0.59), confirming that embeddings optimized for similarity comparison are more suitable for this type of task. It is concluded that PostgreSQL, through PGVector, is capable of acting as a complete similarity-based classification environment, validating its applicability for cyberbullying detection.