LowCLIP: Adapting the CLIP Model Architecture for Low-Resource Languages in Multimodal Image Retrieval Task
This addresses the problem of limited vision-language models for low-resource languages, offering an incremental improvement with specific gains in retrieval performance.
The research tackled multimodal image retrieval for low-resource languages like Azerbaijani by adapting CLIP with techniques such as synthetic data generation and augmentation, achieving state-of-the-art results like boosting ResNet50 MAP on MSCOCO from 0.70 to 0.80.
This research explores the development of multimodal vision-language models for image retrieval in low-resource languages, specifically Azerbaijani. Existing vision-language models primarily support high-resource languages, and fine-tuning them remains computationally demanding. To address challenges in vision-language retrieval for low-resource languages, we integrated the CLIP model architecture and employed several techniques to balance computational efficiency with performance. These techniques include synthetic data generation through machine translation, image augmentation, and further training the attention mechanisms of transformer-based models with domain-specific data. We integrated Multilingual BERT as a text encoder with image encoders like ResNet50, EfficientNet0, Vision Transformer (ViT), and Tiny Swin Transformer. Our study found that models like EfficientNet0 and Tiny Swin Transformer perform best on the datasets they were trained on, such as COCO, Flickr30k, and Flickr8k. Augmentation techniques boosted EfficientNet0 MAP on Flickr30k from 0.84 to 0.87 and ResNet50 MAP on MSCOCO from 0.70 to 0.80, contributing to a new state of the art in vision-language retrieval. We share our configurations and results to support further research. Code and pre-trained models are available at https://github.com/aliasgerovs/azclip.