Multimodal fusion using learned text concepts for image categorization