Boosting Accuracy and Interpretability in Multilingual Hate Speech Detection Through Layer Freezing and Explainable AI

Layer freezing improves efficiency without sacrificing accuracy: The study found that freezing the first eight layers of transformer models while fine-tuning only the later layers maintained strong performance across all five languages (English, Korean, Japanese, Chinese, and French) while reducing computational costs—an important consideration for deploying responsible AI systems at scale.

Explainability is essential for hate speech detection: By integrating LIME (Local Interpretable Model-agnostic Explanations) to highlight which specific words influenced each prediction, the research demonstrates that transparency tools help verify the model isn't relying on spurious correlations or biases, making it safer for content moderation decisions that affect real users.

Multilingual systems require careful validation: Testing across five linguistically diverse languages revealed that model performance varies by language, suggesting that hate speech detection systems deployed globally must be individually validated rather than assuming a single model works equally well everywhere—critical for avoiding discriminatory outcomes in non-English communities.

XLM-RoBERTa-base showed strong cross-lingual performance: Among the three transformer models tested, XLM-RoBERTa-base demonstrated robust results across multiple languages, offering a practical option for organizations building multilingual content moderation systems that need to balance accuracy with interpretability.