Modeling knowledge about possession transfer

Abstract

If we are to successfully create intelligent machines, it is essential to learn how to ground abstract notions, such as possession, in the physical world. In this work, I develop a model for the knowledge about possession transfer, which ties the abstract world to the physical world. The model grounds itself in spatial and time understanding, by making use of Borchardt's work on time space representations. The model identifies a list of 11 prominent possession transfer verbs and establishes a hierarchy to classify the other pertinent verbs. It also defines 6 dimensions for the possession space spanning physical possession, mental state, desire, IOU, money, and moving party. 19 TSR learning templates are developed as the representation for all the cases of all the prominent possession transfer verbs. The salient features of the verbs and their representations are identified. With these salient features, a decision-making tree is created. Near-miss learning is demonstrated to be a good learning technique for the system via 2 descriptive examples. I address the 10 questions and answers that the system can answer with my representation. In addition, 5 questions are addressed which cannot be answered. The correlation between the representation and visual events is discussed and explained with an example, proving how my representation can serve to aid a visual system in understanding the visual events it perceives in the environment.