Learning Physical Properties of Objects Using Gaussian Mixture Models

Abstract

Common-sense knowledge of physical properties of objects such as size and weight is required in a vast variety of AI applications. Yet, available common-sense knowledge-bases cannot answer simple questions regarding these properties such as “is a microwave oven bigger than a spoon?” or “is a feather heavier than a king size mattress?”. To bridge this gap, we harvest semi-structured data associated with physical properties of objects from the web. We then use an unsupervised taxonomy merging scheme to map a set of extracted objects to WordNet hierarchy. We also train a classifier to extend WordNet taxonomy to address both fine-grained and missing concepts. Finally, we use an ensemble of Gaussian mixture models to learn the distribution parameters of these properties. We also propose a Monte Carlo inference mechanism to answer comparative questions. Results suggest that the proposed approach can answer 94.6% of such questions, correctly.