Researchers have developed a novel system for classifying Nepali memes, achieving second place in the CHiPSAL 2026 shared task for hate speech detection and fourth place for sentiment analysis. Their approach utilizes the Qwen3-VL-8B-Instruct vision-language model, which natively supports Devanagari script, eliminating the need for separate OCR and translation steps. The system employs a two-stage training process involving LoRA fine-tuning and contrastive backbone fine-tuning, with strategies like oversampling and focal loss to manage class imbalance. This method demonstrates effective adaptation of large vision-language models for low-resource languages. AI
IMPACT Demonstrates effective adaptation of large vision-language models for low-resource languages, potentially enabling broader applications in multilingual content analysis.
RANK_REASON The item is a research paper detailing a novel system for meme classification using a vision-language model. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
