Tellex, S. "Grounding Language in Spatial Routines"

Abstract

This thesis describes a spatial language understanding system based on a lexicon of words defined in terms of spatial routines. A spatial routine is a script composed from a set of primitive operations on sensor data, analogous to Ullman's visual routines. By finding a set of primitives that underlie natural spatial language, the meaning of spatial terms can be succinctly expressed in a way that can be used to obey natural language commands. This hypothesis is tested by using spatial routines to build a natural language interface to a real time strategy game, in which a player controls an army of units in a battle. The system understands the meaning of context-dependent natural language commands such as "Run back!" and "Move the marines on top above the flamethrowers on the bottom." In evaluation, the system successfully interpreted a range of spatial commands not seen during implementation, and exceeded the performance of a baseline system. Beyond real-time strategy games, spatial routines may provide the basis for interpreting spatial language in a broad range of physically situated language understanding systems, such as mobile robots or other computer game genres.