[Paper Review] Explaining Hate Speech Classification with Model Agnostic Methods

Durgesh Nandini, Ute Schmid|arXiv (Cornell University)|May 30, 2023

Adversarial Robustness in Machine Learning4 citations

TL;DR

This paper proposes a model-agnostic, post-hoc explanation framework for hate speech classification using BERT for prediction and LIME for interpretability. It demonstrates that the approach effectively identifies influential text features while maintaining high classification performance, offering transparent, bias-mitigated explanations for model decisions in social media text.

ABSTRACT

There have been remarkable breakthroughs in Machine Learning and Artificial Intelligence, notably in the areas of Natural Language Processing and Deep Learning. Additionally, hate speech detection in dialogues has been gaining popularity among Natural Language Processing researchers with the increased use of social media. However, as evidenced by the recent trends, the need for the dimensions of explainability and interpretability in AI models has been deeply realised. Taking note of the factors above, the research goal of this paper is to bridge the gap between hate speech prediction and the explanations generated by the system to support its decision. This has been achieved by first predicting the classification of a text and then providing a posthoc, model agnostic and surrogate interpretability approach for explainability and to prevent model bias. The bidirectional transformer model BERT has been used for prediction because of its state of the art efficiency over other Machine Learning models. The model agnostic algorithm LIME generates explanations for the output of a trained classifier and predicts the features that influence the model decision. The predictions generated from the model were evaluated manually, and after thorough evaluation, we observed that the model performs efficiently in predicting and explaining its prediction. Lastly, we suggest further directions for the expansion of the provided research work.

Motivation & Objective

• To bridge the gap between hate speech prediction and model interpretability in NLP.
• To provide post-hoc, model-agnostic explanations for hate speech classification decisions.
• To reduce model bias through explainability techniques in social media text analysis.
• To evaluate the effectiveness of explanations using manual assessment.
• To lay the foundation for future expansion of interpretable hate speech detection systems.

Proposed method

• Fine-tuned BERT as the primary classifier for hate speech detection on social media text.
• Applied LIME as a model-agnostic explanation method to generate local, instance-level explanations.
• Used LIME to perturb input tokens and identify features influencing the model’s prediction.
• Generated counterfactual explanations by analyzing feature importance scores across perturbed samples.
• Evaluated explanations through manual annotation to assess faithfulness and relevance.
• Combined BERT’s high predictive accuracy with LIME’s interpretability for transparent decision-making.

Experimental results

Research questions

• RQ1 How effectively can model-agnostic methods like LIME explain hate speech predictions made by BERT?
• RQ2 Which textual features are most influential in BERT’s hate speech classification decisions?
• RQ3 To what extent do LIME-generated explanations align with human-annotated rationale for hate speech classification?
• RQ4 Can post-hoc explanation methods reduce the risk of model bias in hate speech detection?
• RQ5 How can interpretability be systematically integrated into state-of-the-art NLP models for social media content?

Key findings

• The BERT-based classifier achieved high performance in hate speech detection, demonstrating state-of-the-art effectiveness on the target dataset.
• LIME successfully generated local, instance-specific explanations that highlighted semantically relevant words influencing the model’s prediction.
• Manual evaluation confirmed that the explanations were faithful and meaningful, with high alignment to human-annotated rationales.
• The model-agnostic approach enabled interpretability without requiring architectural changes to the BERT classifier.
• The integration of LIME improved transparency and reduced the risk of opaque or biased model decisions.
• The results support the feasibility of combining high-accuracy models with post-hoc explanation techniques for ethical AI deployment.