Learning Algorithms for Broad-Coverage Semantic Parsing

Broad-coverage semantics holds great potential for improving a variety of downstream applications, such as question-answering, relation extraction, and paraphrasing. However, parsers for broad-coverage semantics, in their current formulation, are heavily reliant on features from predicted syntax. Even with the advancement of deep neural approaches, semantic parsers still incorporate syntactic pipelining to achieve their best performance.In this thesis, we argue that while syntax is crucial for broad-coverage semantic parsing, a reliance on predicted syntax through a pipeline is not sustainable. Predicted syntax tends to be error-prone and not robust to changes in domain. Moreover, discrete syntactic features which used to be the cornerstone for broad-coverage semantic parsing, are not amenable for incorporation in a pipeline involving neural architectures.