Domain knowledge acquisition via language grounding

Abstract

This thesis addresses the language grounding problem at the level of word relation extraction. We propose methods to acquire knowledge represented in the form of relations and utilize them in two domain applications, high-level planning in a complex virtual world and input parser generation from input format specifications. In the first application, we propose a reinforcement learning framework to jointly learn to predict precondition relations from text and to perform high-level planning guided by those relations. When applied to a complex virtual world and text describing that world, our relation extraction technique performs on par with a supervised baseline, and we show that a high-level planner utilizing these extracted relations significantly outperforms a strong, text unaware baseline. In the second application, we use a sampling framework to predict relation trees and to generate input parser code from those trees. Our results show that our approach outperforms a state-of-the-art semantic parser on a dataset of input format specifications from the ACM International Collegiate Programming Contest, which were written in English for humans with no intention of providing support for automated processing.