Prototypes and idealizations in natural language shape descriptions

Natural language shape descriptions are used to describe the appearance of physical objects. Understanding of these mechanisms will aid in tasks such as information retrieval. A system was implemented which can read these descriptions and present three dimensional models of the system's understanding. These graphics may serve as user feedback or to generate none verbal metric index terms. In the approach taken here, language understanding is a process of building mental models based on shape prototypes. Specific linguistic constructs are used to evoke the models and modify them. Object descriptions are governed by six principles: 1) complex objects are described by identifying their parts and relations between the parts, 2) all relevant parts and relations are assigned names (nouns), 3) these named parts are further classified with a small number of prototype shape adjectives, 4) relations between parts are assigned names and act as prototypes, 5) once the prototype is established, deviations from the prototype are expressed, and 6) spatial language may be used to modify the mental models of shape. The computer system built around these principles 'reads' text descriptions of natural objects and produces graphic images of the objects being described. Subjects read the same text description and were able to recognize the image generated by the computer from among a series of other computer generated images. Performance on the task was as good as for a control group performing the same task using artist generated images.