Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Meysam Shirdel Bilehsavar, Negin Mahmoudi, Mohammad Jalili Torkamani, Kiana Kiashemshaki

TL;DR

The study tackles multilingual hate speech detection and sentiment analysis by freezing the first eight layers of three transformer models and applying LIME for explainability across five languages. XLM-RoBERTa-base with layer freezing achieves the highest accuracy (about 92.3%) and F1 (≈0.91), underscoring the value of retaining pretraining representations while adapting task-specific layers. LIME provides concrete, word-level explanations that enhance transparency in predictions across languages, including Japanese, French, and Chinese. The findings support the viability of accurate, interpretable multilingual moderation tools, with future work aimed at improving explanations in mixed-language text and integrating additional explainability methods.

Abstract

Sentiment analysis focuses on identifying the emotional polarity expressed in textual data, typically categorized as positive, negative, or neutral. Hate speech detection, on the other hand, aims to recognize content that incites violence, discrimination, or hostility toward individuals or groups based on attributes such as race, gender, sexual orientation, or religion. Both tasks play a critical role in online content moderation by enabling the detection and mitigation of harmful or offensive material, thereby contributing to safer digital environments. In this study, we examine the performance of three transformer-based models: BERT-base-multilingual-cased, RoBERTa-base, and XLM-RoBERTa-base with the first eight layers frozen, for multilingual sentiment analysis and hate speech detection. The evaluation is conducted across five languages: English, Korean, Japanese, Chinese, and French. The models are compared using standard performance metrics, including accuracy, precision, recall, and F1-score. To enhance model interpretability and provide deeper insight into prediction behavior, we integrate the Local Interpretable Model-agnostic Explanations (LIME) framework, which highlights the contribution of individual words to the models decisions. By combining state-of-the-art transformer architectures with explainability techniques, this work aims to improve both the effectiveness and transparency of multilingual sentiment analysis and hate speech detection systems.

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

TL;DR

The study tackles multilingual hate speech detection and sentiment analysis by freezing the first eight layers of three transformer models and applying LIME for explainability across five languages. XLM-RoBERTa-base with layer freezing achieves the highest accuracy (about 92.3%) and F1 (≈0.91), underscoring the value of retaining pretraining representations while adapting task-specific layers. LIME provides concrete, word-level explanations that enhance transparency in predictions across languages, including Japanese, French, and Chinese. The findings support the viability of accurate, interpretable multilingual moderation tools, with future work aimed at improving explanations in mixed-language text and integrating additional explainability methods.

Abstract

Sentiment analysis focuses on identifying the emotional polarity expressed in textual data, typically categorized as positive, negative, or neutral. Hate speech detection, on the other hand, aims to recognize content that incites violence, discrimination, or hostility toward individuals or groups based on attributes such as race, gender, sexual orientation, or religion. Both tasks play a critical role in online content moderation by enabling the detection and mitigation of harmful or offensive material, thereby contributing to safer digital environments. In this study, we examine the performance of three transformer-based models: BERT-base-multilingual-cased, RoBERTa-base, and XLM-RoBERTa-base with the first eight layers frozen, for multilingual sentiment analysis and hate speech detection. The evaluation is conducted across five languages: English, Korean, Japanese, Chinese, and French. The models are compared using standard performance metrics, including accuracy, precision, recall, and F1-score. To enhance model interpretability and provide deeper insight into prediction behavior, we integrate the Local Interpretable Model-agnostic Explanations (LIME) framework, which highlights the contribution of individual words to the models decisions. By combining state-of-the-art transformer architectures with explainability techniques, this work aims to improve both the effectiveness and transparency of multilingual sentiment analysis and hate speech detection systems.
Paper Structure (13 sections, 7 figures, 2 tables)

This paper contains 13 sections, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Literature review
  3. Methodology
  4. Overview of Dataset
  5. Model Selection: Fine-tuning
  6. Tokenization and Data Preprocessing
  7. Model Fine-tuning
  8. Training Configuration and Hyperparameters
  9. Model Training and Evaluation
  10. Lime explanation
  11. Results and Discussion
  12. Role of LIME in Model Interpretability
  13. Conclusion

Figures (7)

  • Figure 1: Language distribution in dataset
  • Figure 2: Label distribution in dataset
  • Figure 3: Overview of the proposed method
  • Figure 4: Explanation by Lime for Japanese language
  • Figure 5: Explanation by Lime for French language
  • ...and 2 more figures