We can build NLG systems by using machine learning (ML) techniques to create language models which can be used in NLG. Most machine-learning in NLG uses neural models, that is models that are loosely inspired by how neurons in the human brain work. This chapter looks at different ML and neural approaches which are used in NLG, and discusses data and other issues.

The chapter focuses on high-level conceptual material. It does not attempt to give a detailed description of the latest language model technology, since this technology is changing and evolving very quickly.

The chapter includes the following sections:

• Examples
• Machine learning models for NLG
• Training data
• Issues
• Further reading and resources

Resources: Selected blogs

• Amateurs focus on models; professionals focus on data
• Challenges are Same for Neural and Rule NLG
• LLMs and Data-to-text
• Problems in using LLMs in commercial products
• Real-World Neural NLG

Resources: Software

• Huggingface (huge repository of ML models, software, tutorials, etc)
• NLTK (good library for non-neural ML for NLP)

Resources: Web sites

• Google’s blogs and research (biased towards Google products, but a great resource for learning about the latest technologies)