arxiv_cl 95% Match Research Paper NLP Researchers,Machine Learning Engineers,Content Moderation Specialists 2 weeks ago

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

large-language-models › multimodal-llms

📄 Abstract

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.

Authors (2)

Joshua Wolfe Brook

Ilia Markov

Submitted

October 17, 2025

arXiv Category

cs.CL

arXiv PDF

Key Contributions

This research introduces a novel approach to Hate Speech Detection (HSD) by leveraging LLMs to generate background context and incorporate it into HSD classifiers. It explores different context generation strategies and methods for incorporating context, demonstrating significant performance gains.

Business Value

Improved accuracy in content moderation systems for social media platforms and online forums, leading to safer online environments and reduced spread of harmful content.

Paper Metadata

Innovation Type

Methodological

Deployment Feasibility

Moderate. Requires integration with LLMs and potentially complex fusion mechanisms, but the core task is classification.

Limitations Addressed

Lack of contextual understanding in existing HSD models, especially for implicit hate speech and multimodal content.

Performance Gains

Up to 3 F1 points on textual setups and up to 6 F1 points on multimodal setups compared to a zero-context baseline.

Technical Tags

Hate Speech DetectionLLM Context GenerationMultimodal FusionTransformer NetworksImplicit Hate SpeechNamed Entity RecognitionText EnhancementEmbedding ConcatenationHierarchical Fusion

Research Topics

Natural Language ProcessingComputer VisionMachine LearningHate Speech AnalysisContextual Understanding

Methods & Architectures

LLM-based context generationText concatenationEmbedding concatenationHierarchical transformer-based fusionLLM-driven text enhancement TransformerLLM

Applications & Tasks

Social Media Online Content Moderation Cybersecurity ClassificationContent ModerationBias Detection Hate Speech DetectionImplicit Hate Speech DetectionMultimodal Hate Speech Detection

Datasets & Benchmarks

Datasets

Latent Hatred, MAMI

Benchmarks

F1 score gain of up to 3 points (textual) • F1 score gain of up to 6 points (multimodal)

F1 score

Related Fields

Computational LinguisticsSocial ComputingInformation Retrieval

Keywords

Hate Speech DetectionLLMContextual InformationMultimodal AnalysisImplicit Hate SpeechTransformer NetworksOnline SafetyContent ModerationText ClassificationMeme AnalysisLatent Hatred DatasetMAMI Dataset

Academic Context

#Natural Language Processing#Computer Vision#Machine Learning#Hate Speech Analysis#Contextual Understanding

Technology Stack

Frameworks & Libraries

Transformer

Commercial Potential

Potential Products

Automated content moderation toolsHate speech detection APIs

Target Industries

Social MediaTechnologyOnline Advertising

Use Case Examples

Filtering hateful comments on social mediaIdentifying misogynous content in memes

Competitive Edge

Offers a more nuanced and context-aware approach to HSD compared to traditional methods, particularly for implicit and multimodal cases.

Market Opportunity

Large market for content moderation and online safety solutions.

Revenue Models

SaaS for moderation servicesAPI access fees.

Resource Requirements

Compute Needs

High (for LLM inference and training)

Data Requirements

Labeled textual and multimodal datasets for hate speech.

Deployment Constraints

Latency for real-time moderation, computational cost of LLM inference.

Scalability

Scalability depends on the efficiency of LLM integration and the chosen fusion methods.

Regulatory Considerations

Data privacyethical use of AI in content moderation.

Production Readiness

Maturity Level

Research

Time to Market

1-3 years

View Full Paper Back to Papers