Natural language generation (NLG) is a specialised subfield of natural language processing (NLP) that focuses on the development of artificial intelligence systems capable of generating human-language texts that meet a specific objective, such as summarising, translating, or explaining input data (Dong and Gong, Reference Dong, Li and Gong2022).

In recent years, the field has experienced a massive rise in both popularity and performance, largely driven by the impressive capabilities of large language models (LLMs) like ChatGPT. These advancements have brought NLG into the mainstream, making it possible to generate highly fluent and readable text with unprecedented ease (Gao, Hu and Yin, Reference Gao, Hu and Yin2025). This represents a paradigm shift from the conception of NLG presented in the earlier book by Reiter and Dale (Reference Reiter and Dale2000). Nevertheless, NLG is a rapidly evolving domain where today’s “state-of-the-art” technologies can quickly become obsolete. Recognising this, author Ehud Reiter intentionally moves away from providing a guide focused solely on the latest, fast-changing tools. Instead, this book offers a broad perspective on high-level conceptual and methodological guidelines of the fundamental aspects of NLG, such as requirements acquisition, evaluation, and safety, that are designed to remain relevant and useful for researchers and developers well into the future.

The book provides a comprehensive overview of the NLG field, focusing on data-to-text approaches (i.e., generating text from non-textual information as input data), organised into seven main chapters.

The first chapter provides a condensed overview of the whole book, helping the reader to be situated in the NLG task. NLG is cleverly and very simply defined as “the science of AI systems that can write”. Associated with this definition, two important issues are remarked upon: i) that NLG was long before the emergence of generative AI and LLM (e.g., ChatGPT); and ii) that it is not a closed and isolated research area, only relying on AI and NLP techniques, but it highly benefits from other areas, such as human-computer interaction, psychology, and linguistics; software engineers; and domain experts. The distinction between data-to-text and text-to-text is explained through two specific use case examples focused on generating weather forecasts and summaries from doctor-patient consultations, respectively. Although the different technologies to address NLG, either rule-based or machine learning and neural models, are deeply developed in further chapters, chapter 1 offers a high-level and conceptual discussion of them, focusing on their scalability and controllability, which offers a valuable initial contact with the potentials and limitations of each of these strategies. Moreover, issues related to the usefulness and effectiveness of NLG systems, such as requirements, evaluation, safety, testing, and maintenance, are then outlined, with references to the specific book chapters that address them in greater detail, especially the evaluation of NLG output, which is a challenging aspect, partly due to the stochastic nature of NLG systems. Evaluation, together with the aforementioned aspects, is crucial for enabling the reader to understand what NLG systems and applications need to do and take into account in order to be useful to real users and for real-world success. Examples of successful NLG applications in specific domains: journalism, business intelligence, summarisation, and health are also thoroughly discussed in chapter 7 of the book, but they are introduced here to maintain the author’s introductory and forward-looking purpose.

Without going into depth on ethical issues but acknowledging the importance of ensuring ethical behaviour in NLG systems, the author highlights only a few key concerns, mainly related to accessibility across languages and communities (e.g., disparities between low-resource and under-resourced languages and, similarly, for under-represented cultures), biases, unethical use cases (e.g., generating fake news), theft of intellectual property from generated content, and job losses due to automated text production. This allows the reader to further reflect on these aspects and the grey areas that may exist (e.g., NLG systems that are economical with the truth when communicating certain information).

In chapter 2, the author focuses on rule-based NLG, where algorithms and rules are designed to decide what is being generated and how. Although neural NLG is the dominant technique in recent years, the author argues that rule-based approaches remain the most effective solution for certain applications and provides the reader with a conceptual understanding of what a NLG needs, regardless of the type of technique later employed for generating text; thus, the relevance for including them in an independent chapter is well justified. The chapter is specifically centred on the different stages that are involved in the modular and sequential NLG pipeline architecture, ranging from signal analysis and data interpretation to the document planning, microplanning, and surface realisation necessary for transforming input data into text. For each of the stages, the author follows the same structure, revising its main principles and techniques and illustrating through two concrete examples – the Driving Feedback system (which produces reports on unsafe driving from GPS data) and Babytalk (a family of systems which generate summaries and reports about babies in a neonatal intensive care unit).

The chapter ends with template NLG approaches, which are also based on rules, but it constitutes an alternative to modular architectures to generate text directly from the input data.

The third chapter examines the shift from rule-based systems to models that learn to generate text directly from data. It distinguishes between three primary training paradigms: models trained from scratch using large task-specific datasets, fine-tuned models that adapt pre-trained models to specific applications, and prompted models that perform tasks based on direct instructions. To illustrate these paradigms, the author uses examples such as a simple NLG classifier that determines whether the article ‘a’ or an’ should be used given the following word, weather dialogues for fine-tuned neural models, and the use of prompted models to generate weather forecasts. Furthermore, the chapter traces the rapid technological evolution of the field, noting the progression from simple classifiers and n-gram models to the currently dominant transformer architecture (Vaswani et al.,. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017) and foundation models. The author highlights alignment techniques, such as Instruction Tuning and Reinforcement Learning from Human Feedback, which help models meet user expectations. A critical concern is the importance of training data, emphasising that the success of modern neural systems often depends more on the quality and acquisition of datasets than on the specific algorithms used. Finally, the chapter addresses practical challenges in neural NLG, including the problem of domain shift, the risk of factual inaccuracies, and the emerging need for consistent legal and regulatory frameworks.

Chapter Four explains the requirements needed for NLG systems to be useful in real-world applications. It first describes the four main quality criteria that a text should meet: it should be easy to read, factually correct, contain the information users actually care about, and be helpful for the task. For each criterion, the author provides helpful examples to illustrate its importance. The chapter then moves to more global criteria that an NLG system should meet beyond the quality of its outputs. It discusses requirements such as speed, cost, and privacy, as well as the importance of output consistency depending on the area of use. In some fields, it may be more important to report the minimum score obtained rather than the average score. Next, the author reviews how NLG can be used in different workflows, comparing the utility and requirements of systems in fully automated workflows, in workflows where humans always check and edit the outputs, and in setups where the model only provides a draft. The discussion focuses on how accurate, trustworthy, and cost-effective a system can be in each case. Moreover, the chapter examines the utility of systems that provide only text outputs, only graphics, or both, giving examples of relevant experiments. Finally, the chapter offers a helpful guide on how to identify system requirements using a four-stage pipeline. The chapter concludes by stating that NLG systems should consider not only the requirements of the final users but also those of all relevant stakeholders, and it highlights the risks of ignoring these requirements.

Chapter Five highlights the central role of evaluation in both research and real-world applications of NLG. The author argues that evaluation is not only about measuring text quality but also about determining whether a system truly meets user and stakeholder requirements and has a meaningful impact. The chapter introduces fundamental principles such as defining clear research questions, adopting a stakeholder perspective, and using rigorous experimental design, illustrating how to evaluate the real impact on an NLG system aimed at smoking cessation. Evaluation is framed as hypothesis testing, often supported by statistical methods, with particular attention to replication, representative test data, and ecological validity to ensure that findings generalise beyond laboratory settings. A good point of this chapter is the author’s acknowledgement of reporting negative results when NLG does not perform as expected. It then distinguishes between human and automatic evaluation. Human evaluation includes ratings, preference judgements, and task-based studies and is essential for assessing usefulness and quality, although it can be costly and complex. Automatic metrics, which can be classified into reference-based and referenceless metrics, are faster and cheaper but may not correlate well with human judgements, so their validity must be carefully examined. The author stresses the importance of using high-quality test sets and comparing systems against strong baselines or state-of-the-art approaches. For both human and automated evaluation, the author proposes a six-step experimental design: defining research hypotheses, selecting the study type, choosing participants (for human evaluation), selecting test materials, defining the experimental procedure, and determining the analysis method. Finally, the chapter discusses the importance of evaluating the real-world impact of NLG systems and conducting commercial evaluations that consider costs, benefits, risks, and return on investment. It concludes with ten practical tips for conducting rigorous and meaningful evaluations.

Chapter Six examines the challenges of deploying NLG systems responsibly in real-world contexts. It defines safety in terms of risk management and worst-case performance, highlighting dangers such as toxic or misleading outputs, hallucinations, harmful advice, privacy leaks, and outdated information, while noting that mitigation strategies, such as using safer models, putting humans in the loop, automatically detecting safety issues, and conducting red-teaming attacks, can reduce but not eliminate these risks. The chapter clarifies the distinction between software testing and evaluation, presenting testing as quality assurance focused on detecting bugs, handling edge cases, and ensuring systems meet client requirements despite output variability. Finally, it highlights that NLG systems require continuous maintenance, as domains, regulations, and user needs evolve, making safety checks, retesting, and adaptation an ongoing responsibility.

The final chapter of this book describes NLG use cases and applications in four specific domains: journalism, business intelligence, summarisation, and medical applications. Before diving into these domains, the author introduces general guidelines for developing successful NLG applications, including volume and scalability, data availability, accuracy, maintainability and adaptability, acceptability and trust, and conforming to genre and specialised language. As remarked by the author, these guidelines have to be adapted for each scenario, especially for real-world and commercial applications. The author first discusses how NLG systems can be useful for producing diverse types of news (i.e., general, sports, and financial news). The business intelligence use case is motivated by the fact that for some types of insights, it is best to communicate them in words instead of data visualisations. Further, the author points out text summarisation as a long-standing and relevant NLG use case, presenting different data sources from which summaries are useful: meetings, emails, news, and reviews, pointing out that LLMs are nowadays used to produce them. Finally, in the last use case on how NLG can be applied to medical applications, the author reviews different subscenarios ranging from reporting or explaining medical information to patients to helping practitioners make decisions to the generation of messages encouraging behaviour change (e.g., smoking cessation).

In contrast to previous NLG books published over the last decades (Krahmer and Theune Reference Krahmer and Theune2010; Stend and Bangalore Reference Stent and Bangalore2014), which were written primarily for advanced researchers or technical specialists and often emphasise detailed algorithmic approaches, this book blends foundational concepts in NLG with practical illustrations of how these ideas have been applied in real-world applications. It introduces core NLG principles alongside selected use case examples, allowing readers with very different backgrounds – linguistics, computer science, and business – and stages – students, developers, and researchers – to grasp the reasoning behind computational approaches to language generation and to implement and put into practice similar ideas according to their needs. In addition, the book encourages readers to think analogically about how automated text generation – particularly in data-to-text settings – can support decision-making and analysis based on textual or documentary data. One possible limitation, also acknowledged by the author, is that the rapid advancement of the field –particularly with the emergence of LLMs and generative AI – has significantly reshaped the way NLG is approached, and readers may find that the book provides limited detail on these recent developments. Nevertheless, this is not the intention of the author or the book, as such developments are difficult to anticipate or fully address in a single volume, particularly in a field that continues to evolve rapidly. Overall, the book can be recommended to a broad audience interested in understanding the practical potential of NLG. It is worth mentioning that, in addition to the book, it is also valuable that the author maintains a regularly updated blog, where he discusses topics related to the book and responds to readers’ questions, such as in this post: https://ehudreiter.com/2026/03/03/questions-from-readers-of-my-book/.