ML Model to Better Identify Instances of Bullying Faced by Members of the LGBTQ+ Community

Cyberbullying poses a significant threat to online communities, with the LGBTQ+ community facing disproportionately higher rates of harassment. While existing cyberbullying detection systems have made progress in identifying general instances of online harassment, they often fail to capture the nuanced and context-dependent nature of LGBTQ+-targeted bullying. This thesis presents a novel approach to this challenge by developing SpectrumNet, an LGBTQ+-centric transformer-based model for cyberbullying detection. Our research was conducted in two phases. In Phase 1, we evaluated the effectiveness of pre-trained transformer models (RoBERTa, BERT, and GPT-2) in identifying LGBTQ+-related cyberbullying. Building on these findings, Phase 2 introduced SpectrumNet which integrates identity-aware attention mechanisms with hierarchical attention networks to understand the contextual nuances of LGBTQ+-targeted harassment better. The model was trained and evaluated on Instagram comments, demonstrating SpectrumNet's superior performance in detecting LGBTQ+-specific harassment, with notable improvements in identifying subtle forms of bullying that traditional models often miss. This work contributes to the field of cyberbullying detection by introducing specialized architectural components designed specifically for identifying LGBTQ+-targeted harassment, potentially offering new directions for creating safer online spaces for marginalized communities.