This chapter discusses what users and stakeholders want NLG systems to do. The academic community has not paid a lot of attention to requirements, but we cannot build useful NLG systems without knowing what is important to users. Meaningful evaluation also needs to be informed by what users care about.

The chapter starts by looking at quality criteria, such as accuracy and readability of generated texts; understanding which quality criteria are important to users is an essential part of requirements analysis. It then looks at workflows (is the NLG system on its own, or does it collaborate with a human author), and whether users want information presented in words, graphs, or both. It concludes by discussing some methodologies for understand user requirements.

This chapter includes the following sections:

• Quality Criteria: Texts
• Quality Criteria: System
• Workflow
• Text and graphics
• Requirements acquisition
• Further Reading

Resources: Selected blogs

• How do I Build an NLG System: Requirements and Corpora
• INLG: What real-world NLG users want
• We need to understand what users want!
• Why is adoption of AI in healthcare so slow?

Resources: Talks

• NLG Requirements (Ehud) (PDF)