Research map / Large language models / Language, safety and interpretability
A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TiALD, a human-annotated multi-task Tigrinya YouTube-comment benchmark (abusiveness, sentiment, topic) with iterative term-clustering sampling, and shows small fine-tuned encoders beat prompted frontier LLMs.
Problem
Abusive language detection resources and benchmarks are lacking for low-resource languages such as Tigrinya, which has mixed Ge'ez and Romanized social media text and a severe abusive-content imbalance.
Approach
Uses iterative word-embedding term expansion from seed words to build a diverse, balanced annotation pool, jointly annotates three tasks per comment, and includes both Ge'ez and Latin-script comments; a shared encoder with sigmoid multi-label heads is trained for all tasks.
Key results
- TiRoBERTa-base (multi-task) reaches 86.67% macro F1 on abusiveness, over 7 points above the best prompted LLM (Claude Sonnet 3.7 few-shot, 79.31%).
- Best multi-task TiRoBERTa-base achieves a TiALD Score of 64.81, versus 43.54 for the best few-shot LLM.
- Adding video title and generated description as context improves GPT-4o zero-shot abusiveness F1 from 71.05% to 74.70%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use joint multi-task training instead of single-task training (ablation) | TiRoBERTa-base · TiALD test, comment only | TiALD Score: 64.57 -> 64.81 | Helped |
| Use joint multi-task training instead of single-task training (ablation) | TiELECTRA-small · TiALD test, comment only | TiALD Score: 50.54 -> 52.30 | Helped |
| Use joint multi-task training instead of single-task training (ablation) | XLM-RoBERTa-base · TiALD test, comment only | TiALD Score: 51.74 -> 53.59 | Helped |
| Use joint multi-task training instead of single-task training (ablation) | TiRoBERTa-base · TiALD test sexist topic | F1: 31.78 -> 46.30 (+14.52 points) | Helped |
| Use joint multi-task training instead of single-task training (ablation) | TiRoBERTa-base · TiALD test neutral sentiment | F1: 40.0% -> 42.75% | Helped |
| Use joint multi-task training instead of single-task training (ablation) | TiELECTRA-small · TiALD test religious topic | F1: 0% -> 7.36% | Helped |
| Add video title to comment input (ablation) | TiRoBERTa-base single-task · TiALD test sentiment | Macro F1: 52.82 -> 54.97 (+1.5 points) | Helped |
| Add video title to comment prompt (ablation) | GPT-4o zero-shot · TiALD test abusiveness | Macro F1: 71.05 -> 75.59 | Helped |
| Add video title to comment prompt (ablation) | Claude Sonnet 3.7 zero-shot · TiALD test abusiveness | Macro F1: 59.20 -> 67.64 | Helped |
| Add video title and generated video description to comment prompt (ablation) | GPT-4o zero-shot · TiALD test abusiveness | Macro F1: 75.59 -> 74.70 | Hurt |
| Add video title and generated video description to comment prompt (ablation) | Claude Sonnet 3.7 zero-shot · TiALD test abusiveness | Macro F1: 67.64 -> 72.02 | Helped |
| Use few-shot examples instead of zero-shot prompting (ablation) | Claude Sonnet 3.7 · TiALD test abusiveness, comment only | Macro F1: 59.20% -> 79.31% | Helped |
Surprising findings
- Multi-task joint learning consistently improved over single-task training across encoder models.
- Smaller Gemma-3 4B outperformed frontier GPT-4o and Claude on sentiment and topic classification in zero-shot and few-shot settings.
Limitations
- Focuses on overt abusive language; implicit toxicity not covered.
- Single abusiveness label without granular subtypes.
- Sentiment and topic tasks show low macro F1 (about 53-55%), indicating difficulty; moderate inter-annotator agreement on topic (kappa 0.603).
Experimental setup
Datasets: TiALD (YouTube comments from 51 channels, 4.1M source corpus)
Benchmarks: TiALD
Metrics: macro F1 TiALD Score (average macro F1 across tasks) per-class F1 type-to-token ratio Cohen's kappa
Baselines: TiELECTRA-small TiRoBERTa-base AfriBERTa-base AfroXLMR-Large-76L XLM-RoBERTa-base GPT-4o Claude Sonnet 3.7 Gemma-3 4B LLaMA-3.2 3B
Builds on: word2vec Cohen's kappa AdamW
Closest papers by meaning
- Spanning the Spectrum of Hatred Detection: A Persian Multi-Label Hate Speech Dataset with Annotator Rationales (AAAI 2024 · similarity 0.87)
- Efficient Toxic Content Detection by Bootstrapping and Distilling Large Language Models (AAAI 2024 · similarity 0.86)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.85)
- TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection (AAAI 2026 · similarity 0.85)
- BingoGuard: LLM Content Moderation Tools with Risk Levels (ICLR 2025 · similarity 0.84)
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.