Multimodal multilingual models: improving image-text embedding alignment across languages with limited data
Rapid advancements in Visual Language Models (VLMs) have transformed multimodal understanding but are often constrained by generating English responses regardless of the input language. This phenomenon has been termed as Image-induced Fidelity Loss (IFL) and stems from limited multimodal multilingua...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/446130 |
| Acceso en línea: | https://hdl.handle.net/2117/446130 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning (Machine learning) Natural language processing (Computer science) Computer vision Deep learning Visual language models Multimodality LLMs Aprenentatge profund (Aprenentatge automàtic) Tractament del llenguatge natural (Informàtica) Visió per ordinador Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic |
| Sumario: | Rapid advancements in Visual Language Models (VLMs) have transformed multimodal understanding but are often constrained by generating English responses regardless of the input language. This phenomenon has been termed as Image-induced Fidelity Loss (IFL) and stems from limited multimodal multilingual training data. To address this, we propose a continuous multilingual integration strategy that injects text-only multilingual data during visual instruction tuning, preserving the language model's original multilingual capabilities. Extensive evaluations demonstrate that our approach significantly improves linguistic fidelity across languages (from 2.7 % to 88.7 % in German and from 4.4 % to 92.9 % in Spanish) without degradation in visual performance. We also explore model merging, which improves language fidelity but comes at the cost of visual performance. In contrast, our core method achieves robust multilingual alignment without trade-offs, offering a scalable and effective path to mitigating IFL for global VLM adoption. |
|---|