Andrew R. Haas and Nobuyuki Shimizu
Department of Computer Science, University at Albany
Albany, New York
Several authors have considered the possibility of a robot (real or simulated) that accepts directions in human language. Modern computational linguists test their programs using large collections of natural text or speech, but programs that follow directions have seldom been tested in this manner. We describe the collection of a corpus of 847 sets of directions for navigation through a virtual office building, written by 89 different individuals. After collecting 650 examples, we built a program that can follow these directions. We collected 197 more sets of directions and used them to test the program; it was able to reach the desired destination 79% of the time. This program is simple; we describe data from our corpus that suggests possible extensions.
Andrew R. Haas and Nobuyuki Shimizu
Department of Computer Science, University at Albany
Albany, New York