Abstract: Social media platforms are integral to shaping public discourse and facilitating the rapid dissemination of both positive and negative information within the contemporary digital landscape. However, the detection and management of abusive content on these platforms present significant challenges, as such content has the potential to incite social unrest and perpetuate hate. In response to this pressing issue, our study introduces an advanced DistilBERT model specifically designed for the efficient identification of abusive text. By leveraging the compact yet powerful BERT architecture, our model adeptly captures subtle contextual nuances, thereby enhancing predictive accuracy. A pivotal component of our methodology is the rigorous preprocessing of the dataset, which includes the removal of redundant or duplicate samples to ensure dataset integrity and mitigate biases that could compromise the model’s ability to generalize from unique instances. The DistilBERT model exhibits exceptional performance, achieving near-perfect scores in accuracy, precision, recall, and F1-score, significantly surpassing existing state-of-the-art and baseline methodologies. These findings underscore the model’s robustness and its potential as an effective instrument for monitoring and mitigating online abuse. By proposing a scalable solution that incorporates comprehensive data preprocessing, this research contributes to advancing the field of abusive content detection and aims to promote safer and more inclusive online communities.

Key words: abusive content detection, text classification, social network analysis, DistilBERT, multi-classification