Skip to main content
eScholarship
Open Access Publications from the University of California

Towards Democratizing Data Science with Natural Language Interfaces

  • Author(s): Su, Yu
  • Advisor(s): Yan, Xifeng
  • et al.
Abstract

Data science has the potential to reshape many sectors of the modern society. This potential can be realized to its maximum only when data science becomes democratized, instead of centralized in a small group of expert data scientists. However, with data becoming more massive and heterogeneous, standing in stark contrast to the spreading demand of data science is the growing gap between human users and data: Every type of data requires extensive specialized training, either to learn a specific query language or a data analytics software. Towards the democratization of data science, in this dissertation we systematically investigate a promising research direction, natural language interface, to bridge the gap between users and data, and make it easier for users who are less technically proficient to access the data analytics power needed for on-demand problem solving and decision making.

One of the largest obstacles for general users to access data is the proficiency requirement on formal languages (e.g., SQL) that machines use. Automatically parsing natural language commands from users into formal languages, natural language interfaces can thus play a critical role in democratizing data science. However, a pressing question that is largely left unanswered so far is, how to bootstrap a natural language interface for a new domain? The high cost of data collection and the data-hungry nature of the mainstream neural network models are significantly limiting the wide application of natural language interfaces.

The main technical contribution of this dissertation is a systematic framework for bootstrapping natural language interfaces for new domains. Specifically, the proposed framework consists of three complimentary methods: (1) Collecting data at a low cost via crowdsourcing, (2) leveraging existing NLI data from other domains via transfer learning, and (3) letting a bootstrapped model to interact with real users so that it can refine itself over time. Combining the three methods forms a closed data loop for bootstrapping and refining natural language interfaces for any domain.

The developed methodologies and frameworks in this dissertation hence pave the path for building data science platforms that everyone can use to process, query, and analyze their data without extensive specialized training. With such AI-powered platforms, users can stay focused on high-level thinking and decision making, instead of overwhelmed by low-level implementation and programming details --- ``\emph{Let machines understand human thinking. Don't let humans think like machines}.''

Main Content
Current View