community detection and hate speech analysis on twitter using transformer and recurrent neural models

This paper addresses the problem of hate speechdetection and community analysis on twitter by combiningdeep learning models with network analysis. we propose ahybrid lstm-gru architecture for text classification andemploy the girvan–newman algorithm for hate communitydetection. experiments are conducted on a six-class dataset of27,410 tweets. the results show that while bert achieves thehighest accuracy across all categories (98.95% on average), thelstm-gru model attains competitive performance (98.32%)with substantially lower computational cost.although bert remains superior in raw accuracy, itshigher computational demands limit its practicality in realtime or resource-constrained environments. hence, the maincontribution of our hybrid lstm-gru lies not in surpassingbert’s accuracy but in achieving a favorable efficiency–performance trade-off.this efficiency makes the hybrid model a practicalalternative for real-time moderation and resource-constrainedenvironments. in parallel, community detection experimentsreveal influential nodes and clusters that help characterize thespread of hate speech within social networks. together, thesefindings highlight the trade-off between accuracy andefficiency in model selection, and demonstrate the value ofcombining text classification with graph-based communitydetection for comprehensive hate speech analysis.