How to Ask What to Say?: Strategies for Evaluating Natural Language Interfaces for Data Visualization

摘要

In this article, we discuss challenges and strategies for evaluating natural language interfaces (NLIs) for data visualization. Through an examination of prior studies and reflecting on own experiences in evaluating visualization NLIs, we highlight benefits and considerations of three task framing strategies: Jeopardy-style facts, open-ended tasks, and target replication tasks. We hope the discussions in this article can guide future researchers working on visualization NLIs and help them avoid common challenges and pitfalls when evaluating these systems. Finally, to motivate future research, we highlight topics that call for further investigation including development of new evaluation metrics, and considering the type of natural language input (spoken versus typed), among others.