Dual-branch Prompting for Multimodal Machine Translation
Researchers have developed D2P-MMT, a novel diffusion-based framework designed to improve multimodal machine translation (MMT) by making it more robust to irrelevant visual information. This approach utilizes a dual-branch prompting strategy during training, incorporating both authentic and reconstructed images to foster cross-modal interactions. A key innovation is a distributional alignment loss that ensures consistency between the two branches, bridging the gap between training and inference. Experiments on the Multi30K dataset show D2P-MMT outperforms existing state-of-the-art methods. AI
IMPACT This research could lead to more reliable machine translation systems that better leverage visual context, improving accuracy in real-world applications.