Uncertainty and Ambiguity in Semantic Structures

Language is a powerful tool for communication and coordination, allowing us to share thoughts, ideas, and instructions with others. Accordingly, enabling people to communicate linguistically with digital agents has been among the longest-standing goals in AI. However, unlike humans, machines do not naturally acquire the ability extract meaning from language. One natural solution to this problem is to represent meaning in a structured format (e.g. a semantic parse, code, logic, etc.) and to develop models for processing language into such structures. In this talk, I will discuss two core challenges in parsing language into structure: uncertainty and ambiguity. I will first present results benchmarking the ability of semantic parsing systems to quantify uncertainty. I will then move on to a discussion of ambiguity in parsing, introducing a novel dataset and data generation framework. Time permitting, I will discuss recent work on using uncertainty to improve safety with a human in the loop.