A Scene Corpus for Training and Testing Spatial Communication Systems

It is argued that a ‘scene corpus’ would be a useful tool for the training and testing of systems for grounded spatial communication in the same way that text corpora have been used for training and assessing other language processing systems. Such a scene corpus would need to allow a full range of spatial relationships to be expressed over a range of scale spaces. The scenes should be sufficiently complex to allow sequential spatial descriptions to be constructed. The integration of listener models and reference frame variation will be required. The interface design will need to allow for different implementations and corpus extensions. Initial steps in the design of a scene corpus are described.