Dialectal Toxicity Detection: Evaluating LLM-as-a-Judge Consistency Across Language Varieties

2024/11/17 by Fahim Faisal, Md. Mushfiqur Rahman, Faisal, Fahim +3 · 3 citations
Arts and Humanities · Computer Science · Health Professions · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Interpreting and Communication in Healthcare #Natural Language Processing Techniques #linguistics and terminology studies

paper · pdf · doi:10.48550/arxiv.2411.10954

openalex publication_date 2024/11/17 · openalex created_date 2024/11/21 · openalex updated_date 2026/07/28

Abstract

There has been little systematic study on how dialectal differences affect toxicity detection by modern LLMs. Furthermore, although using LLMs as evaluators ("LLM-as-a-judge") is a growing research area, their sensitivity to dialectal nuances is still underexplored and requires more focused attention. In this paper, we address these gaps through a comprehensive toxicity evaluation of LLMs across diverse dialects. We create a multi-dialect dataset through synthetic transformations and human-assisted translations, covering 10 language clusters and 60 varieties. We then evaluated three LLMs on their ability to assess toxicity across multilingual, dialectal, and LLM-human consistency. Our findings show that LLMs are sensitive in handling both multilingual and dialectal variations. However, if we have to rank the consistency, the weakest area is LLM-human agreement, followed by dialectal consistency. Code repository: \urlhttps://github.com/ffaisal93/dialect_toxicity_llm_judge

Cited by

• A Comparative Benchmark of a Moroccan Darija Toxicity Detection Model (Typica.ai) and Major LLM-Based Moderation APIs (OpenAI, Mistral, Anthropic)
• Measurement Without Validity: The Compounding Reliability Problem in Agentic AI Evaluation
• Not Like Us, Hunty: Measuring Perceptions and Behavioral Effects of Minoritized Anthropomorphic Cues in LLMs

Related

• Translate, then Detect: Leveraging Machine Translation for Cross-Lingual Toxicity Classification
• Toxicity Red-Teaming: Benchmarking LLM Safety in Singapore's Low-Resource Languages
• From One to Many: Expanding the Scope of Toxicity Mitigation in Language Models
• Realistic Evaluation of Toxicity in Large Language Models
• PolygloToxicityPrompts: Multilingual Evaluation of Neural Toxic Degeneration in Large Language Models