Show, Tell and Discriminate: Image Captioning by Self-retrieval with Partially Labeled Data | Synapse