Talk2Data: High-Level Question Decomposition for Data-Oriented Question and Answering

Through a data-oriented question and answering system, users can directly "ask" the system for the answers to their analytical questions about the input tabular data. This process greatly improves user experience and lowers the technical barriers of data analysis. Existing techniques focus on providing a concrete query for users or untangling the ambiguities in a specific question so that the system could better understand questions and provide more correct and precise answers. However, when users have little knowledge about the data, it is difficult for them to ask concrete questions. Instead, high-level questions are frequently asked, which cannot be easily solved with the existing techniques. To address the issue, in this paper, we introduce Talk2Data, a data-oriented online question and answering system that supports answering both low-level and high-level questions. It leverages a novel deep-learning model to resolve high-level questions into a series of low-level questions that can be answered by data facts. These low-level questions could be used to gradually elaborate the users' requirements. We design a set of annotated and captioned visualizations to represent the answers in a form that supports interpretation and narration. We evaluate the effectiveness of the Talk2Data system via a series of evaluations including case studies, performance validation, and a controlled user study. The results show the power of the system.