Inferring Conceptual Structures from Pictorial Input Data

Abstract

This thesis investigates the mechanisms a program may use to learn conceptual structures that represent natural language meaning. A computer program named Moran is described that infers conceptual structures from simulated pictorial input data. Moran is presented "snapshots" of an environment and an English sentence describing the action that takes place between the snapshots. The learning task is to associate each root verb with a conceptual structure that represents the types of objects that participate in the action and the changes the objects undergo during the action. Four learning mechanisms are shown to be adequate to accomplish this learning task. The learning mechanisms are described along with the conditions under which each is invoked and the effect each has on existing memory structures. The conceptual structures that Moran infers for seventeen senses of four root verbs are shown.