ABSTRACT
Symbols are the basis of many forms of human communication and are constantly being developed to communicate novel ideas in new contexts. The fact that certain symbols are more effective than others has been experimentally studied and documented, but the question of whether an algorithm could predict the effectiveness of a set of symbols or produce a new set of symbols has received little attention. The principal contributions of this work are 1) the first formal quantitative measurement of the effectiveness of a set of visual symbols and 2) the first algorithm and methodology for the generation of superior symbols to represent a set of discrete concepts. The foundation of the symbol generation algorithm is an information theoretic model for the evaluation of simple connected-line drawings. Concepts from Shannon and Al- gebraic Information Theory provide a foundation for the model and are combined with concepts from vision science, such as Structural Information Theory, to address the com- plexities of human perception. This theoretical model is used to establish a formal grammar for a class of visual symbols and perceptually motivated quantitative measures of symbol complexity and similarity. Given an input of discrete symbolic data and a prior probability distribution, the symbol generation algorithm minimizes these measures of complexity and similarity in a set of symbols and generates a mapping from the input data to these symbols. The performance of the symbol generation algorithm is evaluated through formal and empirical means. Formal evaluation is accomplished through analysis of the produced symbol sets using models from perceptual psychology and vision science. The algorithm is evaluated empirically by measuring the performance of human subjects in a visual task using the symbols. |