Instance Hardness Threshold for Effective Sampling in Bangla Cyberbullying Detection Using Transformer Models
Md. Nihal Morshed, Md. Shahid Uz Zaman, Mohiuddin Ahmed · 2024
In recent years, social networking sites have offered users several possibilities to exchange information, interact, and engage through positive interactions. However, these same platforms can also be used maliciously, fostering an environment prone to online abuse and cyberbullying. Cyberstalking, in which people use the internet to mock, abuse, defame, or humiliate victims without physically interfering, has increased in popularity with the growth of online platforms. These kinds of inappropriate activities have become common on Facebook in particular. Given the potential for widespread distress, it is vital to develop automated methods for recognizing and eliminating harassment via web content. This study proposes a transformer-based deep neural network framework to identify online harassment in Bengali on social networking websites. Bengali text data is transformed into an organized format by the model utilizing efficient text preprocessing approaches, and TfidfVectorizer (TFID) is used to extract features in order to detect important textual aspects. The Instance Hardness Threshold (IHT) approach is used to balance the dataset, mitigating potential risks of overfitting and underfitting. A public dataset of 44,001 Bengali comments from Facebook is used in our study. We use deep learning models based on transformers to categorize abusive comments, specifically BERT and ELECTRA. Our studies generated remarkable outcomes for multi-label classification, with F1 scores of 93% and 91% for BERT and ELECTRA architectures, respectively. These findings represent a substantial improvement over previous efforts, confirming the framework’s capacity to detect and categorize cyberbullying content in Bengali effectively. Therefore, this technique may assist in safeguarding people from online harassment and misconduct.