Processing of 3-D Spatial Relations for Virtual Agents Acting on Natural Language Instructions

We are developing a system in which a user can control virtual agents through natural language dialogue. We particularly focus on the semantic analysis of spatial expressions in natural language instructions. Spatial expressions are relative to the speaker’s viewpoint and the indicated positions can change dynamically. In this paper we propose a semantic representation using lambda structures. Some features of lambda structure enable us to keep the speaker’s viewpoints in such semantic representations undetermined, to keep track of changing positions, and to encapsulate representations like English prepositions. With this representation, we can handle this kind of spatial expressions efficiently. We also state the method of translating natural language instructions to this representation.