In this talk, I will address the problem of natural language grounding. We assume access to natural language documents that specify the desired behaviour of a control application. Our goal is to generate a program that will perform the task based on this description. The programs involve everything from changing the privacy settings on your browser, playing computer games, performing complex text processing tasks, to even solving math problems. Learning to perform tasks like these is complicated because the space of possible programs is very large, and the connections between the natural language and the resulting programs can be complex and ambiguous. I will present methods that utilize semantics of the target domain to reduce natural language ambiguity. On the most basic level, executing the induced programs in the corresponding environment and observing their effects can be used to verify the validity of the mapping from language to programs. We leverage this validation process as the main source of supervision to guide learning in settings where standard supervised techniques are not applicable. Beyond validation feedback, we demonstrate that using semantic inference in the target domain (e.g., program equivalence) can further improve the accuracy of natural language understanding.
展开▼
