Parsing Parking Instructions for Self-driving Cars into Spatial Semantic Descriptions

摘要

This study is motivated by an attempt to develop a system that can control a self-driving car by words. To achieve this, the verbal instructions we provide to the car must be grounded onto the real-world phenomena. With this background, in this study, we extend the framework of semantic description clause (SDC) proposed by Kollar et al. (2010) by adding two new semantic categories, VIEW and STATE, so as to be able to ground more variety of the instructions for driving a car in real-world environment. We developed a pipeline of a CCG parser, a reranker of parse trees and a converter from parse trees into SDCs. The result of parsing the instructions with the extended CCG grammar shows 84.7% accuracy on 6,019 various Japanese driving instructions. We finally convert the parse trees into SDCs using conversion rules, resulting in 90.4% accuracy on 5,638 parse trees.