Learning Structured Neural Semantic Parsers

<p>Semantic parsing, the task of translating user-issued natural language (NL) utterances (<em>e</em>.<em>g</em>., F<em>lights from Pittsburgh to New York</em>) into formal meaning representations (MRs, <em>e</em>.<em>g</em>., an SQL database query or a Python program), has become an important direction in developing natural language interfaces to computational systems. Recent years have witnessed the burgeoning of applying neural network-based semantic parsers in various tasks and domains. However, meaning representations typically exhibit strong syntactic structure, and are defined following domain-specific structured knowledge schemas (<em>e</em>.<em>g</em>., a database schema or Python API specification), which is not easily captured by standard neural sequence transduction models. Neural semantic parsers are also data-hungry, requiring non-trivial manual annotation effort by domain experts. These issues limit the scope of applications supported by a neural semantic parser, impeding the progress of applying the system to broader scenarios, especially those with diverse and complex structure of meaning representations.</p> <p>In this thesis, we explore developing neural semantic parsing models that could better capture the <em>structure</em> in various types of logical formalisms and knowledge schemas, while providing approaches to mitigate the cost of labeled data acquisition. The dissertation consists of three parts. The first part introduces a general-purpose parsing model with built-in syntactic knowledge of the grammatical structure of meaning representations. Next, in the second part, we investigate approaches to encode structured information in domain knowledge schemas (<em>e</em>.<em>g</em>., database tables) useful to understand user-issued utterances. Specifically, we focus on grounding elements in the schema (<em>e</em>.<em>g</em>., columns like departure_city in database tables, or functions like GetFlight(from=GetCityByName(·)) in API specifications) to their corresponding NL constituents (<em>e</em>.<em>g</em>., <em>from Pittsburgh</em>) in utterances. Finally, in the third part, we aim to improve the data efficiency of semantic parsers via semisupervised learning, while developing machine-assisted approaches to accelerate training data acquisition. </p>