DDML: Multi-Student Knowledge Distillation for Hate Speech.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Social Science

Background:

• Online hate speech poses significant risks to user mental health and social stability.
• The increasing prevalence of online hate speech necessitates efficient detection methods.
• Transformer-based models show promise but are computationally intensive for deployment.

Purpose of the Study:

• To develop an efficient and effective method for hate speech detection.
• To address the computational challenges of large transformer-based models.
• To improve the performance of hate speech detection systems through a novel knowledge distillation approach.

Main Methods:

• Proposed Deep Distill-Mutual Learning (DDML), a knowledge distillation technique.
• DDML utilizes one teacher network and multiple student networks for mutual learning.
• Trained deep neural networks for hate speech detection using the DDML framework.

Main Results:

• DDML-based networks demonstrated strong performance across diverse datasets.
• The method was tested on ten languages and nine datasets.
• Achieved an average F1 score increase of 4.87% over baseline models.

Conclusions:

• DDML enhances the performance of deep neural networks for hate speech detection.
• The proposed method offers a computationally efficient solution for deploying advanced models.
• DDML shows promise for improving the accuracy and efficiency of online hate speech detection systems.