Enhanced Hate Speech Detection through Mean-Pooling in Embedding Fusion

The rise of social media has transformed communication globally, but the anonymity it offers often leads to aggressive and offensive language, making it essential to address hate speech, cyberbullying, and harassment. While much research has focused on detecting hate speech, fewer studies have examined the intended audience and linguistic aggression. This paper investigates hate speech detection on Twitter, emphasizing the classification of both target individuals and aggressive behavior in hateful tweets. We propose a fusion-based deep learning model that combines Bidirectional Encoder Representations from Transformers (BERT) and fastText embeddings for enhanced hate speech classification. Our experiments utilize the benchmark dataset from SemEval 2019 Task 5, which is dedicated to the Shared Task for Multilingual Detection of Hate. Our model outperformed all methods proposed by participants in Subtask B–Aggressive Behavior and Target Classification –achieving a 67% F1-score. Additionally, it demonstrated competitive performance in Subtask A–Hate Speech Detection against immigrants and women –matching the top team’s 65% F1-score. Our approach contributes valuable insights into effective embedding fusion for improved hate speech detection on Twitter, along with advancements in classifying the intended audience and linguistic aggression.