Lifecycle of neural semantic parsing
Humans are born with the ability to learn to perceive, comprehend and communicate with language. Computing machines, on the other hand, only understand programming languages. To bridge the gap between humans and computers, deep semantic parsers convert natural language utterances into machine-understandable logical forms. The technique has a wide range of applications ranging from spoken dialogue systems and natural language interfaces. This thesis focuses on neural network-based semantic parsing. Traditional semantic parsers function with a domain-specific grammar that pairs utterances and logical forms, and parse with a CKY-like algorithm in polynomial time. Recent advances in neural semantic parsing reformulate the task as a sequence-to- sequence learning problem. Neural semantic parsers parse a sentence in linear time, and reduce the need for domain-specific assumptions, grammar learning, and extensive feature engineering. But this modeling flexibility comes at a cost since it is no longer possible to interpret how meaning composition is performed, given that logical forms are structured objects (trees or graphs). Such knowledge plays a critical role in understanding modeling limitations so as to build better semantic parsers. Moreover, the sequence-to-sequence learning problem is fairly unconstrained, both in terms of the possible derivations to consider and in terms of the target logical forms which can be ill-formed or unexecutable. The first contribution of this thesis is an improved neural semantic parser, which produces syntactically valid logical forms following a transition system and grammar constrains. The transition system integrates the generation of domain-general (i.e., valid tree-structures and language-specific predicates) and domain-specific aspects (i.e., domain-specific predicates and entities) in a unified way. The model employs various neural attention mechanisms to handle mismatches between natural language and formal language—a central challenge in semantic parsing. Training data to semantic parsers typically consists of utterances paired with logical forms. Another challenge of semantic parsing concerns the annotation of logical forms, which is labor-intensive. To write down the correct logical form of an utterance, one not only needs to have expertise in the semantic formalism, but also has to ensure the logical form matches the utterance semantics. We tackle this challenge in two ways. On the one hand, we extend the neural semantic parser to a weakly-supervised setting within a parser-ranker framework. The weakly-supervised setup uses training data of utterance-denotation (e.g., question-answer) pairs, which are much easier to obtain and therefore allow to scale semantic parsers to complex domains. Our framework combines the advantages of conventional weakly-supervised semantic parsers and neural semantic parsing. Candidate logical forms are generated by a neural decoder and subsequently scored by a ranking component. We present methods to efficiently search for candidate logical forms which involve spurious ambiguity—some logical forms do not match utterance semantics but coincidentally execute to the correct denotation. They should be excluded from training. On the other hand, we focus on how to quickly engineer a practical neural semantic parser for closed domains, by directly reducing the annotation difficulty of utterance-logical form pairs. We develop an interface for efficiently collecting compositional utterance-logical form pairs and then leverage the data collection method to train neural semantic parsers. Our method provides an end-to-end solution for closed-domain semantic parsing given only an ontology. We also extend the end-to-end solution to handle sequential utterances simulating a non-interactive user session. Specifically, the data collection interface is modified to collect utterance sequences which exhibit various co-reference patterns. Then the neural semantic parser is extended to parse context-dependent utterances. In summary, this thesis covers the lifecycle of designing a neural semantic parser: from model design (i.e., how to model a neural semantic parser with an appropriate inductive bias), training (i.e., how to perform fully supervised and weakly supervised training for a neural semantic parser) to engineering (i.e., how to build a neural semantic parser from a domain ontology).