Community Detection and Hate Speech Analysis on Twitter using Transformer and Recurrent Neural Models

This paper addresses the problem of hate speech detection and community analysis on Twitter by combining deep learning models with network analysis. We propose a hybrid LSTM-GRU architecture for text classification and employ the Girvan-Newman algorithm for hate community detection. Experiments are conducted on a six-class dataset of ۲۷,۴۱۰ tweets. The results show that while BERT achieves the highest accuracy across all categories (۹۸.۹۵% on average), the LSTM-GRU model attains competitive performance (۹۸.۳۲%) with substantially lower computational cost. Although BERT remains superior in raw accuracy, its higher computational demands limit its practicality in real-time or resource-constrained environments. Hence, the main contribution of our hybrid LSTM-GRU lies not in surpassing BERT's accuracy but in achieving a favorable efficiency-performance trade-off. This efficiency makes the hybrid model a practical alternative for real-time moderation and resource-constrained environments. In parallel, community detection experiments reveal influential nodes and clusters that help characterize the spread of hate speech within social networks. Together, these findings highlight the trade-off between accuracy and efficiency in model selection, and demonstrate the value of combining text classification with graph-based community detection for comprehensive hate speech analysis.

کلیدواژه ها:

Deep Learning ، Social Network Analysis ، Community Identification ، Long-short-term Memory ، Bidirectional Encoder Representations from Transformers