Date of Award

Spring 2021

Document Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Radev, Dragomir


Natural language is a fundamental form of information and communication and is becoming the next frontier in computer interfaces. As the amount of data available online has increased exponentially, so has the need for Natural Language Interfaces (NLIs, which is not used for natural language inference in this thesis) to connect the data and the user by easily using natural language, significantly promoting the possibility and efficiency of information access for many users besides data experts. All consumer-facing software will one day have a dialogue interface, and this is the next vital leap in the evolution of search engines. Such intelligent dialogue systems should understand the meaning of language grounded in various contexts and generate effective language responses in different forms for information requests and human-computer communication.Developing these intelligent systems is challenging due to (1) limited benchmarks to drive advancements, (2) alignment mismatches between natural language and formal programs, (3) lack of trustworthiness and interpretability, (4) context dependencies in both human conversational interactions and the target programs, and (5) joint language understanding between dialog questions and NLI environments (e.g. databases and knowledge graphs). This dissertation presents several datasets, neural algorithms, and language models to address these challenges for developing deep learning technologies for conversational natural language interfaces (more specifically, NLIs to Databases or NLIDB). First, to drive advancements towards neural-based conversational NLIs, we design and propose several complex and cross-domain NLI benchmarks, along with introducing several datasets. These datasets enable training large, deep learning models. The evaluation is done on unseen databases. (e.g., about course arrangement). Systems must generalize well to not only new SQL queries but also to unseen database schemas to perform well on these tasks. Furthermore, in real-world applications, users often access information in a multi-turn interaction with the system by asking a sequence of related questions. The users may explicitly refer to or omit previously mentioned entities and constraints and may introduce refinements, additions, or substitutions to what has already been said. Therefore, some of them require systems to model dialog dynamics and generate natural language explanations for user verification. The full dialogue interaction with the system’s responses is also important as this supports clarifying ambiguous questions, verifying returned results, and notifying users of unanswerable or unrelated questions. A robust dialogue-based NLI system that can engage with users by forming its responses has thus become an increasingly necessary component for the query process. Moreover, this thesis presents the development of scalable algorithms designed to parse complex and sequential questions to formal programs (e.g., mapping questions to SQL queries that can execute against databases). We propose a novel neural model that utilizes type information from knowledge graphs to better understand rare entities and numbers in natural language questions. We also introduce a neural model based on syntax tree neural networks, which was the first methodology proposed for generating complex programs from language. Finally, language modeling creates contextualized vector representations of words by training a model to predict the next word given context words, which are the basis of deep learning for NLP. Recently, pre-trained language models such as BERT and RoBERTa achieve tremendous success in many natural language processing tasks such as text understanding and reading comprehension. However, most language models are pre-trained only on free-text such as Wikipedia articles and Books. Given that language in semantic parsing is usually related to some formal representations such as logic forms and SQL queries and has to be grounded in structural environments (e.g., databases), we propose better language models for NLIs by enforcing such compositional interpolation in them. To show they could better jointly understand dialog questions and NLI environments (e.g. databases and knowledge graphs), we show that these language models achieve new state-of-the-art results for seven representative tasks on semantic parsing, dialogue state tracking, and question answering. Also, our proposed pre-training method is much more effective than other prior work.