Deep Compositional Captioning: Describing Novel Object Categories without Paired Training Data | IEEE Conference Publication | IEEE Xplore