This chapter describes how NLG systems can be built using algorithms and rules which explicitly extract insights from texts, structure information into narratives, and create good linguistic expressions of information; this is called rule-based NLG. Rule-based NLG is mainly used for data-to-text.

Building rule-based NLG systems can be expensive, but this approach gives developers complete control over functionality. Rule-based systems will not go ‘off the rails’ and do crazy or inappropriate things (unless they have software bugs, which can be fixed).

This chapter includes the following sections:

• NLG Pipeline
• Examples
• Signal analysis
• Data interpretation
• Document planning
• Microplanning
• Surface realisation
• Template NLG
• Further reading and resources

Resources: Selected blogs

• Challenges of Surface Realisation
• How Should Different NLG Components Add Value?
• NLG vs Templates: Levels of Sophistication in Generating Text
• We Need Robust Ways to Select Content of NLG Texts
• What are the Problems with Rule-Based NLG?

Below blogs are more personal, reflecting on things I have done

• An Architecture for Data-to-Text Systems
• The story of simplenlg

Resources: Software

• Jinja (templates in Python)
• simplenlg (surface realisation in various languages)

Resources: Surveys

• Gatt and Krahmer 2018 (especially Section 2)