TeamV at LT-EDI 2026: Multilingual Hate Speech Span Detection and Counter-Narrative Generation via Few-Shot In-Context Learning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

TeamV developed a system for the LT-EDI 2026 Shared Task on Counter-Narrative Generation, specifically targeting homophobic and transphobic comments. Their approach addresses two subtasks: Hate Speech Span Detection in English, Tamil, and Hindi, and Counter-Narrative Generation in English and Tamil. The system utilizes the Qwen3-Max large proprietary language model, enhanced by few-shot in-context learning (ICL) and robust post-processing. This methodology achieved significant results on the CodaBench leaderboard. For Task 1, TeamV secured 1st place across all three languages, with macro F1 scores of 0.5338 in English, 0.5272 in Tamil, and 0.5478 in Hindi. In Task 2, their generated counter-narratives ranked 1st globally in English with an 87.47% average score and 5th in Tamil. The paper details their prompting methodology and span-matching pipeline.

Key takeaway

For NLP Engineers developing multilingual content moderation systems, this work highlights the effectiveness of few-shot in-context learning with large language models. You should consider Qwen3-Max or similar proprietary LLMs, coupled with robust post-processing, to achieve high accuracy in hate speech span detection and generate effective counter-narratives across languages like English, Tamil, and Hindi. Implement a detailed prompting methodology to optimize performance.

Key insights

Few-shot in-context learning with large LMs excels at multilingual hate speech detection and counter-narrative generation.

Principles

• Large language models offer strong multilingual reasoning.
• Few-shot ICL enhances proprietary LLM performance.
• Robust post-processing is crucial for task accuracy.

Method

TeamV's system uses Qwen3-Max with few-shot in-context learning and robust post-processing for hate speech span detection and counter-narrative generation across multiple languages.

In practice

• Apply few-shot ICL for multilingual NLP tasks.
• Integrate post-processing for LLM output refinement.
• Evaluate LLMs on specific hate speech detection metrics.

Topics

• Hate Speech Detection
• Counter-Narrative Generation
• Few-Shot Learning
• In-Context Learning
• Qwen3-Max
• Multilingual NLP

Best for: Research Scientist, AI Scientist, NLP Engineer, Prompt Engineer

Related on AIssential

• Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification — Prompt sensitivity and demonstration quality critically impact LLM performance in metalinguistic classification tasks.
• Comparison of Modern Multilingual Text Embedding Techniques for Hate Speech Detection Task — Supervised models with multilingual embeddings and gradient boosted trees excel at hate speech detection.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.
• Detecting HIV-Related Stigma in Clinical Narratives Using Large Language Models — LLMs can effectively detect HIV-related stigma in clinical notes, with encoder models outperforming generative models in zero-shot settings.
• A Multitask Transformer for Offensive Language Detection and Target Identification in HateBR — Multitask learning improves offensive language detection consistency but may reduce fine-grained target attribution.
• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.