Formalizing Support

Creating an Ontology For Spatial Reasoning in Knowledge Bases

Team: Chongyang Gao, Chris Coleman, Evan Costa, Wangcheng Xu

In English, we can say that “the cup is on the table”, “the tattoo is on the arm”, and “the picture is on the wall.” While all three examples reflect physical support relations, the three examples are conceptually different instances of support. On one hand, the cup is on the table because it’s gravitationally supported - the table is solid, and unless the cup is glued down, we could feasibly move the cup such that it is no longer “on.” On the other hand, “the tattoo is on the arm” can be considered embedded support - the location of the tattoo is fixed with respect to the arm and we can’t suspend the support relation in the same way as moving a cup. Recent research grounded in cross-lingual and cognitive development studies defined five main subtypes of support relations: gravitational, embedded, adhesion, hanging, and point-attachment.

The objectives for this project with respect to representing support subtypes from a spatial/physical perspective are to (i) describe structural properties of support subtypes (ii) design reasonably-simple representations which describe whether a subtype holds given properties of the figure and ground, and (iii) formalize a process for handling any uncertainty we encounter. Then, given these representations, we can reason about figure/ground pairs in the context of specific support relation subtypes. Additionally, we could explore possible formalization to reinforce the support relation between the figure and the ground to meet the common expectations for their integrity and functionality to hold. This project aims at adding these support relations to a knowledge base called NextKB created by Northwestern University's qualitative reasoning group.

First We Had To Categorize Different Support Relations Using Knowledge in NextKB and Pre-Existing Research

Landau, B., Johannes, K., Skordos, D., & Papafragou, A. (2017). Containment and Support: Core and Complexity in Spatial Language Learning. Cognitive science, 41 Suppl 4, 748-779.

Second, we encoded these different instances of spacial relationships into NextKB in such a way that NextKB can reason to answer fundamental questions about the knowledge we have encoded into it. For example, what type of relative movements are possible with a cup on a table? Our code will allow the NextKB reasoner to reason that the cup can be moved upwards and sideways, but downwards since the table is directly below the cup, and the cup cannot move through the table.