Scholarlink

<p>With the rapid proliferation of social media platforms, hate speech has become an increasingly prevalent social problem in online environments. This study proposes a hybrid model integrating DistilBERT, RoBERTa, DeBERTa, and ALBERT Transformer-based language models, the Cheetah Optimizer Algorithm (COA), and the BiLSTM-attention (att) network for the automatic detection of text-based hate speech. In the proposed three-stage framework, high-dimensional features are obtained from the Transformer models. Then, the most meaningful features are selected using the COA and transferred to the BiLSTM network, focusing on critical features using the attention mechanism. The effectiveness of the proposed model was evaluated on the CyberTrolls, Hatebase, and Supremacist datasets. Experimental results show that proposed DeBERTa + COA + BiLSTM-att framework outperforms all other models with accuracies of 92.50%, 96.57%, and 83.68%, respectively. These findings reveal that COA-based feature selection improves the performance of Transformer representations by reducing unnecessary features. The integration of BiLSTM and the attention mechanism further improves the model by learning complex interactions within Transformer-derived feature representations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid transformer–BiLSTM-attention framework enhanced by cheetah optimization algorithm for hate speech detection

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Fatih Demirci,
  • Zeynep Garip,
  • Ekin Ekinci

摘要

With the rapid proliferation of social media platforms, hate speech has become an increasingly prevalent social problem in online environments. This study proposes a hybrid model integrating DistilBERT, RoBERTa, DeBERTa, and ALBERT Transformer-based language models, the Cheetah Optimizer Algorithm (COA), and the BiLSTM-attention (att) network for the automatic detection of text-based hate speech. In the proposed three-stage framework, high-dimensional features are obtained from the Transformer models. Then, the most meaningful features are selected using the COA and transferred to the BiLSTM network, focusing on critical features using the attention mechanism. The effectiveness of the proposed model was evaluated on the CyberTrolls, Hatebase, and Supremacist datasets. Experimental results show that proposed DeBERTa + COA + BiLSTM-att framework outperforms all other models with accuracies of 92.50%, 96.57%, and 83.68%, respectively. These findings reveal that COA-based feature selection improves the performance of Transformer representations by reducing unnecessary features. The integration of BiLSTM and the attention mechanism further improves the model by learning complex interactions within Transformer-derived feature representations.