Please note: This PhD defence will take place online.
Peng Shi, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Jimmy Lin
Information access, which enables people to identify, retrieve, and use information freely and effectively, has attracted interest from academia and industry. Systems for document retrieval and question answering have helped people access information in powerful and useful ways. Recently, natural language technologies based on neural network have been applied to various tasks for information access. Specifically, transformer-based pre-trained models have pushed tasks such as document and passage retrieval to new state-of-the-art effectiveness. (1) Most of the research has focused on helping people access passages and documents on the web. However, there is abundant information stored in other formats such as semi-structured tables and domain-specific relational databases in companies. Development of the models and frameworks that support access information from these data formats is also essential. (2) Moreover, most of the advances in information access research are based on English, leaving other languages less explored. It is insufficient and inequitable in our globalized and connected world to serve only speakers of English.
In this thesis, we explore and develop models and frameworks that could alleviate the aforementioned challenges. This dissertation consists of three parts. We begin with a discussion on developing models designed for accessing data in formats other than passages and documents. We mainly focus on two data formats, namely semi-structured tables and relational databases. In the second part, we discuss methods that can enhance the user experience for non-English speakers when using information access systems. Specifically, we first introduce model development for multilingual knowledge graph integration, which can benefit many information access applications such as cross-lingual question answering systems and other knowledge-driven cross-lingual NLP applications. We further focus on multilingual document dense retrieval and reranking that boost the effectiveness of search engines for non-English information access. Last but not least, we take a step further based on the aforementioned two parts by investigating models and frameworks that can facilitate non-English speakers to access structured data. In detail, we present cross-lingual Text-to-SQL semantic parsing systems that enable non-English speakers to query relational databases with queries in their languages.