SAJI_English@LT-EDI 2026: Detection of Homophobia and Transphobia in Internet Memes Using Zero-Shot Learning

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

Summary

Jishnu Bandyopadhyay, Saloni Kushwaha, Deepawali Sharma, and Aakash Singh developed a zero-shot learning method to detect homophobic and transphobic content within internet memes. This approach addresses the growing challenge of identifying harmful messages in multimodal formats, which are prevalent among younger social media users. Their method, which secured rank 5 in the LT-EDI 2026 shared task, employs two large language models, Qwen2.5-VL-3B-Instruct and Meta-Llama-3-8B-Instruct. These LLMs are utilized to generate descriptions of memes and subsequently classify them based on the meme dataset provided for the LT-EDI 2026 challenge. The system achieved a macro F1-score of 0.55 for English language memes, demonstrating a practical step towards maintaining healthier online environments by flagging and removing abusive content.

Key takeaway

NLP Engineers developing robust hate speech detection systems for multimodal content should note that zero-shot learning with LLMs offers a competitive approach. You can integrate models like Qwen2.5-VL-3B-Instruct and Meta-Llama-3-8B-Instruct to generate descriptions and classify complex abusive content. This strategy enhances your system's adaptability to new online harms, improving moderation efficiency.

Key insights

Zero-shot learning with LLMs can effectively detect multimodal hate speech in internet memes.

Principles

• Multimodal content like memes complicates hate speech detection.
• Zero-shot learning offers a viable approach for new content types.

Method

Employ two LLMs (Qwen2.5-VL-3B-Instruct, Meta-Llama-3-8B-Instruct) to generate meme descriptions, then classify for homophobia/transphobia in a zero-shot manner.

In practice

• Apply zero-shot LLM techniques to new hate speech categories.
• Integrate multimodal hate speech detection into content moderation.

Topics

• Zero-shot Learning
• Large Language Models
• Hate Speech Detection
• Internet Memes
• Multimodal AI
• Online Content Moderation

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• I Know What You Meme, Even If it Emerged Today: Understanding Evolving Memes through Open-World Knowledge Acquisition — Meme understanding requires dynamic, open-world knowledge acquisition beyond static model parameters.
• Multi-Label Text Classification with Scikit-LLM — Scikit-LLM enables zero-shot multi-label text classification using LLMs with a scikit-learn-like workflow.
• Memes-as-Replies: Can Models Select Humorous Manga Panel Responses? — LLMs show promise in understanding social cues for humor, but struggle with visual information and subtle wit in meme reply selection.
• 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.
• 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.
• ArtNet: A JEPA-Like Articulatory Predictive Framework for Robust Zero-Shot Phoneme Recognition — ArtNet uses articulatory feature prediction and VIB to achieve robust zero-shot phoneme recognition across languages.

Open in AIssential →

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