Towards generalisable hate speech detection: a review on obstacles and solutions

Area of Science:

• Natural Language Processing
• Computational Social Science
• Artificial Intelligence

Background:

• Online hate speech is a growing societal concern, targeting individuals based on identity attributes.
• Automatic detection of hate speech is crucial for content moderation and online safety.
• Existing hate speech detection models exhibit poor generalization to unseen data, limiting their real-world applicability.

Purpose of the Study:

• To survey the generalizability of current hate speech detection models.
• To identify the underlying reasons for poor generalization in hate speech detection.
• To propose future research avenues for enhancing model generalization.

Main Methods:

• Literature review of existing hate speech detection models and generalization studies.
• Analysis of common challenges and limitations in current hate speech detection approaches.
• Synthesis of research findings to identify key obstacles and propose solutions.

Main Results:

• Current hate speech detection models often fail to perform well on diverse or novel datasets.
• Factors contributing to poor generalization include data bias, domain shift, and adversarial attacks.
• Existing methods to improve generalization show limited success, highlighting the need for novel approaches.

Conclusions:

• Improving the generalizability of hate speech detection models is essential for effective online safety.
• Future research should focus on robust feature engineering, diverse training data, and domain adaptation techniques.
• Developing more resilient and adaptable hate speech detection systems is a critical next step.