MELLA: Bridging Linguistic Capability and Cultural Groundedness for Low-Resource Language MLLMs
Summary
MELLA, a novel multimodal multilingual dataset, addresses the significant performance gap of Multimodal Large Language Models (MLLMs) in low-resource languages. Existing MLLMs often provide "thin descriptions" due to insufficient cultural groundedness and linguistic capability. This study introduces a dual-source data strategy, collecting 6.8 million image-text pairs across eight low-resource languages: Arabic, Czech, Hungarian, Korean, Russian, Serbian, Thai, and Vietnamese. MELLA comprises D_know (native web alt-text for cultural knowledge) and D_ling (MLLM-generated, translated captions for linguistic fluency). Experiments demonstrate that fine-tuning MLLM backbones like InternVL2-8B and QwenVL2-7B on MELLA significantly improves both cultural understanding (keyword accuracy) and linguistic skills (BLEU, ROUGE-L, METEOR), enabling "thick descriptions."
Key takeaway
For AI Scientists and Machine Learning Engineers expanding MLLM capabilities to low-resource languages, you should prioritize both linguistic fluency and cultural groundedness. Relying solely on machine translation for data will yield superficial outputs. Instead, adopt a dual-source data strategy, integrating native cultural context from sources like web alt-text with linguistically rich translated captions to build truly effective and inclusive multimodal AI systems.
Key insights
Effective low-resource MLLMs require both linguistic fluency and deep cultural understanding, beyond mere translation.
Principles
- Image meaning includes literal denotation and symbolic connotation.
- Translation-only data produces "thin descriptions" lacking cultural depth.
- Authentic, in-culture data is vital for cultural groundedness.
Method
A dual-source strategy collects native web alt-text for cultural knowledge (D_know) and MLLM-generated, translated captions for linguistic capability (D_ling), then combines them for unified fine-tuning.
In practice
- Source native web alt-text for cultural knowledge.
- Generate and translate MLLM captions for linguistic fluency.
- Combine diverse datasets for balanced MLLM enhancement.
Topics
- Multimodal LLMs
- Low-Resource Languages
- Cultural Groundedness
- Dataset Curation
- Supervised Fine-tuning
- Image Captioning
Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer
Related on AIssential
See Counsel's argued verdicts on the open AI decisions leaders are weighing →
Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.