Learning Hybrid Models for Image Annotation with Partially Labeled Data