HATE SPEECH DETECTION FROM SOCIAL MEDIA USING WORD EMBEDDINGS AND GRAPH NEURAL NETWORKS

Abstract

The speedy expansion of online social networks has transformed the medium of communication and information promulgation, facilitating unprecedented levels of connectivity worldwide. OSN has increased social outreach as well and it also highlights any issues faster than conventional systems. However, this digital revolution has also provided a fascinating field for the procreation of hate speech, posing significant threats to both individual well-being and societal cohesion. In response to this pressing issue, researchers have vigorously pursued various methodologies aimed at identifying and detecting hate speech in OSN. Among these methodologies, deep learning techniques have emerged as particularly promising solutions as it provides more accurate results, leveraging their capacity to analyze vast amounts of textual data and extract meaningful patterns. This major report undertakes a complete comparative analysis of different deep learning approaches for HSD, focusing intently on evaluating their performance using performance metrics. By analyzing each method on diverse datasets including Davidson-ICWSM, Waseem EMNLP, Waseem-NAACL, and VLSP, we systematically evaluate the efficiency of multiple deep learning methods. Convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), transformers, graph convolutional networks (GCNs), and ensemble learning methodologies undergo stringent scrutiny in this study. Our analysis uncovers refinement insights into the strengths and limitations of each approach in the context of HSD. Among these, LR shows good results, but LSTM and bi-LSTM modeling have demonstrated exceptional performances, even though facing challenges such as handling multilingual datasets and classification issues. Additionally, BERT-based models show outstanding results in detecting derogatory language and slang across diverse linguistic landscapes. Moreover, the introduction of graph convolutional network (GCN) models presents a promising approach for enhancing HSD capabilities. By capitalizing on the inherent structural relationships within online social networks, GCNs display notable potential in capturing complex patterns of HS propagation. v In conclusion, this major report serves as a valuable compass for advancing the field of HSD, tracing a course toward the development of more robust and effective strategies to uphold a safer and more inclusive digital environment.