“Detectors Lead, LLMs Follow”: Integrating LLMs and traditional models on implicit hate speech detection to generate faithful and plausible explanations

Social media platforms face a growing challenge in addressing abusive content and hate speech, particularly as traditional natural language processing methods often struggle with detecting nuanced and implicit instances. To tackle this issue, our study enhances Large Language Models (LLMs) in the detection and explanation of implicit hate speech, outperforming classical approaches. We focus on two key objectives: (1) determining whether jointly predicting and generating explanations for why a message is hateful improves LLMs’ accuracy, especially for implicit cases, and (2) evaluating whether incorporating information from BERT-based models can further boost detection and explanation performance. Our method evaluates and enhances LLMs’ ability to detect hate speech and explain their predictions. By combining binary classification (Hate Speech vs. Non-Hate Speech) with natural language explanations, our approach provides clearer insights into why a message is considered hateful, advancing the accuracy and interpretability of hate speech detection. • Performance Comparison : LLMs, such as GPT-4, outperform traditional models, like BERT and HateBERT, on explicit hate speech detection, but traditional models show stronger performance in non-hateful and implicit cases, highlighting complementary strengths. • Guided LLMs with External Information : Incorporating external data, such as detection labels, probabilities, and implied stereotypes from BERT-based models, into LLM prompts, improves detection accuracy, particularly for nuanced and implicit hate speech messages. • Impact on Explanation Quality : Incorporating external information from Hate Speech detectors, fine-tuned on Implicit Hate Speech data, can lead to more relevant, less repetitive explanations and with higher cogency.