OLD-TL: Offensive Language Detection in Gaming Live Stream Using Transfer Learning
Ferdousi Haque, Atanu Shome · 2025
Live -streaming gaming platforms are witnessing the prevalence of offensive language; these are harmful to both teenage and adult mental health. We propose an approach that leverages the BERT model to automatically detect offensive language in gaming live streams. Our work addresses the pressing need to create a safe and inclusive gaming environment. We perform two types of experiments: a) training and testing on in-case data (live-streaming audio to text) and b) training on out-case (existing similar cases in different platforms) dataset and testing on in-case data. We progressively fine-tuned the BERT model to achieve optimal performance. Our proposed approach obtains an F1-score of 96% on the in-case test data, demonstrating superior performance in this genre. Our approach presents a promising solution for detecting offensive language in gaming live streams, essential for creating a safe and inclusive gaming environment. By implementing our approach, platform administrators can proactively take measures to ensure a safer and more inclusive gaming environment for all users.