SpectrumNet: Detecting LGBTQ+ Cyberbullying with Dynamic Context-Aware Attention

Abstract

Cyberbullying remains a critical societal issue, with LGBTQ+ individuals disproportionately affected. Although previous work proposed general cyberbullying detection models, LGBTQ+-targeted cyberbullying detection remains relatively unexplored. SpectrumNet, a novel transformer-based model introduced in this paper, goes beyond conventional cyberbullying detection by adding conversational context and identity-aware modeling. SpectrumNet freezes the RoBERTa backbone and adds three key components: a hierarchical attention network to capture linguistic nuance, a GRU-based encoder to better capture comment history, and a dynamic fusion module to effectively weigh contextual signals. To address dataset imbalance, we apply focal loss and weighted sampling. Trained on a large, annotated Instagram dataset, SpectrumNet effectively differentiates between non-bullying, general bullying, and LGBTQ+-targeted bullying. In particular, it achieves strong recall on targeted content and excels at detecting subtle forms of discrimination often missed in isolation but evident within threaded interactions.