Towards Interpretable Hate Speech Detection using Large Language Model-extracted Rationales

TL;DR

This work tackles hate speech detection with a need for interpretability by introducing SHIELD, a framework that leverages LLM-extracted rationales to augment a base detector. The architecture fuses two embeddings—the detector’s $h^{i}_{[CLS]}$ and an LLM-derived $h^{i}_{ft[CLS]}$—via a concatenated representation $h^{i}_{combined}$ fed into a two-layer MLP, while the LLM remains a rationale provider rather than the classifier itself. Empirical evaluation across explicit and implicit English datasets shows that the LLM-derived rationales align with human rationales (with substantial token and semantic overlap) and that SHIELD preserves competitive detection performance, with notable gains on Twitter. The approach demonstrates a path toward faithful interpretability in hate speech detection, balancing transparency with practical effectiveness, and highlights avenues for faithfulness evaluation and broader LLM integration.

Abstract

Although social media platforms are a prominent arena for users to engage in interpersonal discussions and express opinions, the facade and anonymity offered by social media may allow users to spew hate speech and offensive content. Given the massive scale of such platforms, there arises a need to automatically identify and flag instances of hate speech. Although several hate speech detection methods exist, most of these black-box methods are not interpretable or explainable by design. To address the lack of interpretability, in this paper, we propose to use state-of-the-art Large Language Models (LLMs) to extract features in the form of rationales from the input text, to train a base hate speech classifier, thereby enabling faithful interpretability by design. Our framework effectively combines the textual understanding capabilities of LLMs and the discriminative power of state-of-the-art hate speech classifiers to make these classifiers faithfully interpretable. Our comprehensive evaluation on a variety of English language social media hate speech datasets demonstrate: (1) the goodness of the LLM-extracted rationales, and (2) the surprising retention of detector performance even after training to ensure interpretability. All code and data will be made available at https://github.com/AmritaBh/shield.

Figures (2)

• Figure 1: Our proposed SHIELD framework.
• Figure 2: Examples with both LLM-annotated and human-annotated rationales. Overlap is in purple.