Extraction of Common Physical Properties of Everyday Objects from Structured Sources

Abstract

Commonsense knowledge is essential for the reasoning of AI systems, particularly in the context of action planning for robots. The focus of this paper is on common-sense object properties, which are especially useful to restrict the search space of planning algorithms. Popular sources for such knowledge are commonsense knowledge bases that provide the information in a structured form. However, the utility of the provided object-property pairs is limited as they can be simply incorrect, subjective, unspecific, or relate only to a narrow context. In this paper, we suggest a methodology to create a highly accurate dataset of object properties that are related to common physical attributes. The approach is based on filtering non-physical properties within commonsense knowledge bases and improving the accuracy of the remaining object-property pairs based on supervised machine learning using annotated data. Thereby, we evaluate different types of features and models and significantly increase the correctness of object-property pairs compared to the original sources.