Leveraging LLMs for Context-Aware Implicit Textual and Multimodal Hate Speech Detection

TL;DR

This work investigates using Large Language Models as dynamic knowledge bases to generate background context for textual and multimodal hate speech detection. It compares two context-generation strategies (named entities vs full-text prompts) and four context-incorporation methods (text append, embedding fusion, hierarchical fusion, and LLM-based text enhancement) on Latent Hatred and MAMI. Embedding-level fusion (Embed & Concat) with full-text context consistently yields the strongest gains, up to 3 F1 points for textual HSD and 6 F1 points for multimodal HSD, while pure LLM predictions work well for binary tasks but falter on fine-grained classes. The findings highlight the importance of context quality and separation between post and context, offering practical guidance for deploying context-aware HSD systems and suggesting avenues for more robust prompting and dynamic weighting in future work.

Abstract

This research introduces a novel approach to textual and multimodal Hate Speech Detection (HSD), using Large Language Models (LLMs) as dynamic knowledge bases to generate background context and incorporate it into the input of HSD classifiers. Two context generation strategies are examined: one focused on named entities and the other on full-text prompting. Four methods of incorporating context into the classifier input are compared: text concatenation, embedding concatenation, a hierarchical transformer-based fusion, and LLM-driven text enhancement. Experiments are conducted on the textual Latent Hatred dataset of implicit hate speech and applied in a multimodal setting on the MAMI dataset of misogynous memes. Results suggest that both the contextual information and the method by which it is incorporated are key, with gains of up to 3 and 6 F1 points on textual and multimodal setups respectively, from a zero-context baseline to the highest-performing system, based on embedding concatenation.

Figures (8)

• Figure 1: Example false negative meme alongside its LLM-generated context.
• Figure 2: Example false positive meme alongside its LLM-generated context.
• Figure 3: Examples of memes and their labels from the MAMI dataset.
• Figure 4: Examples of memes with potentially irrelevant named entities present in the generated image description.
• Figure 5: Confusion matrices for Zero-Context (left) and Full-Text Embed & Concat (right) binary HSD predictions on the Latent Hatred dataset.