The Structure and Generality of Spoken Route Instructions

A robust system that understands route instructions should be able to process instructions generated naturally by humans. Also desirable would be the ability to handle repairs and other modifications to existing instructions. To this end, we collected a corpus of spoken instructions (and modified instructions) produced by subjects provided with an origin and a destination. We found that instructions could be classified into four categories, depending on their intent such as imperative, feedback, or meta comment. We asked a different set of subjects to follow these instructions to determine the usefulness and comprehensibility of individual instructions. Finally, we constructed a semantic grammar and evaluated its coverage. To determine whether instructiongiving forms a predictable sub-language, we tested the grammar on three corpora collected by others and determined that this was largely the case. Our work suggests that predictable sub-languages may exist for well-defined tasks.