Weakly Supervised Learning for Scalable Semantic Parsing
Abstract
Major Goals: One central challenge in natural language research is to do robust, wide coverage semantic analysis. Recently, significant progress has been made by developing algorithms for learning semantic parsers, which map sentences to rich, logical representations of their underlying meaning. State-of-the-art approaches can now learn highly accurate parsers for a number of benchmark problems. However, the general applicability of this work has been limited by the focus on somewhat idealized conditions, where the target domain is of limited size, each sentence is analyzed in isolation, and there is a primary focus on database query applications. We are developing a new framework for weakly supervised semantic parsing that solves these challenges by learning parsers from indirect, but easily gathered forms of supervision. For example, such an approach could learn to reason about a sentenceÕs possible meanings given the linguistic and situated context in which it was uttered. This type of reasoning is necessary for extending existing learning approaches to fundamentally new applications, such as understanding 3D spatial language. It is also, as we will see, crucial for building semantic parsers that scale to domains which are several orders of magnitude larger than previously considered. We will demonstrate these advantages within a number of applications, including question answering and grounded language understanding. The research is organized around common themes of (1) inducing linguistically rich probabilistic grammars (in the CCG formalism) that compactly model a wide range of complex linguistic phenomena and (2) introducing new techniques to learn from easily gathered training data (e.g., question-answer pairs). When combined, these advances will significantly improve the scale and general applicability of semantic parsers. Also, where necessary, we will investigate alternate methods of semantic supervision (including crowdsourcing of predicate argument structure in our recently introduced QA-SRL formalism) and deep learning methods that can complement (or even replace) the CCG grammars, while still being effectively learnable from the available supervision.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1610121
Entities
People
- Luke Zettlemoyer
Organizations
- Army Contracting Command
- United States Army
- University of Washington