UNISON: Unpaired Cross-Lingual Image Captioning

Image captioning has emerged as an interesting research field in recent years due to its broad application scenarios. The traditional paradigm of image relies on paired image-caption datasets train the model a supervised manner. However, creating such for every target language is prohibitively expensive, which hinders extensibility technology and deprives large part world population benefit. In this work, we present novel unpaired cross-lingual method generate captions without relying any caption corpus source or language. Specifically, our consists two phases: (1) auto-encoding process, utilizing sentence parallel (bitext) learn mapping from scene graph encoding space decode sentences language, (2) cross-modal unsupervised feature mapping, seeks map encoded features modality modality. We verify effectiveness proposed Chinese generation task. comparisons against several existing methods demonstrate approach.