MemeScouts@LT-EDI 2026: Asking the Right Questions - Prompted Weak Supervision for Meme Hate Speech Detection

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

Summary

The MemeScouts team introduced a Prompted Weak Supervision (PWS) approach for detecting hate speech in memes, addressing challenges posed by multimodality, subtle cultural cues, and multilingual contexts. Traditional end-to-end prompting of vision-language models (VLMs) often proves brittle for this task. Their PWS method decomposes meme understanding into targeted, question-based labeling functions with constrained answer options, specifically for homophobia and transphobia detection. Utilizing a quantized Qwen3-VLM to extract features, this technique significantly outperformed direct VLM classification. The approach achieved substantial gains for Chinese and Hindi, securing 1st place in English, 2nd in Chinese, and 3rd in Hindi in the LT-EDI 2026 shared task. Iterative refinement through error-driven labeling function expansion and feature pruning further enhanced generalization and reduced redundancy.

Key takeaway

For Machine Learning Engineers developing robust hate speech detection systems for multimodal content, especially in multilingual contexts, direct end-to-end VLM prompting often proves insufficient. You should consider adopting a Prompted Weak Supervision (PWS) approach, which decomposes complex tasks into targeted, question-based labeling functions. This method, demonstrated to outperform direct VLM classification, offers a more structured and effective pathway to improve accuracy and generalization across diverse languages and subtle cultural cues.

Key insights

Prompted Weak Supervision (PWS) with question-based labeling functions enhances VLM performance for multilingual multimodal hate speech detection.

Principles

• Decompose complex multimodal tasks into targeted questions.
• Constrain VLM answer options for specific sub-tasks.
• Iterative error-driven refinement enhances generalization.

Method

The Prompted Weak Supervision (PWS) method decomposes meme understanding into targeted, question-based labeling functions with constrained answers. It uses a quantized Qwen3-VLM for feature extraction, refined by error-driven LF expansion and feature pruning.

In practice

• Apply PWS to complex multimodal content analysis.
• Utilize quantized VLMs for efficient feature extraction.
• Implement error-driven refinement for labeling functions.

Topics

• Prompted Weak Supervision
• Hate Speech Detection
• Vision-Language Models
• Multimodal AI
• Multilingual NLP
• Qwen3-VLM

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Adapting Reinforcement Learning with Chain-of-Thought Supervision for Explainable Detection of Hateful and Propagandistic Memes — Reinforcement learning with CoT supervision significantly improves MLLM performance and explainability for hateful meme detection.
• 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.
• MAF: Multimodal Adaptive Few-shot Prompting for Sentiment Analysis with MLLMs — MAF dynamically adapts few-shot prompts for MLLMs in sentiment analysis by integrating multimodal cues and adaptive demonstration retrieval.
• Can AI Save India’s Dying Languages? | Building a Tulu LLM Without Training Data Ft. Prathamesh — Structured prompting and custom romanization enable LLMs to generate low-resource languages without extensive training data.
• Semi-Supervised Speech Confidence Detection using Pseudo-Labelling and Whisper Embeddings — A novel framework detects speaker confidence by fusing human-engineered features and Whisper embeddings, enhanced by pseudo-labelling for data scarcity.
• From Codebooks to VLMs: Evaluating Automated Visual Discourse Analysis for Climate Change on Social Media — VLMs can effectively analyze social media images for climate discourse, even with moderate per-image accuracy.

Open in AIssential →

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