Look, Read and Enrich Learning from Scientific Figures and their Captions Jose Manuel Gomez-Perez, Raul Ortega Expert System Cogito Labs { jmgomez,rortega}@expertsystem.com Abstract Compared to natural images, understanding scientific figures is particularly hard for machines. However, there is a valuable source of information in scientific literature that until now has remained untapped: the correspondence between a figure and its caption. In this paper we investigate what can be learnt by looking at a large number of figures and reading their captions, and introduce a figure-caption correspondence learning task that makes use of our observations. Training visual and language networks without supervision other than pairs of unconstrained figures and captions is shown to successfully solve this task. We also show that transferring lexical and semantic knowledge from a knowledge graph significantly enriches the resulting features. Finally, we demonstrate the positive impact of such features in other tasks involving scientific text and figures, like multi-modal classification and machine comprehension for question answering, outperforming supervised baselines and ad-hoc approaches. 1 Introduction Scientific knowledge is heterogeneous and can present itself in many forms, including text, mathematical equations, figures and tables. Like many other manifestations of human thought, the scientific discourse usually adopts the form of a narrative, a scientific publication where related knowledge is presented in mutually supportive ways over different modalities. In the case of scientific figures, like charts, images and diagrams, these are usually accompanied by a text paragraph, a caption, that elaborates on the analysis otherwise visually represented. In this paper, we make use of this observation and tap on the potential of learning from the enormous source of free supervision available in the scientific literature, with millions of figures and their captions.