[Paper Review] Building Natural-Language Generation Systems

Ehud Reiter|ArXiv.org|May 2, 1996

Natural Language Processing Techniques8 references1,249 citations

TL;DR

This paper presents a structured framework for building natural-language generation (NLG) systems, outlining three core stages: content determination and text planning, sentence planning, and realization. It evaluates a spectrum of techniques—from hard-coded rules to AI-based planning—demonstrating that the optimal approach depends on application-specific needs such as fluency, syntactic variety, and resource constraints.

ABSTRACT

This is a very short paper that briefly discusses some of the tasks that NLG systems perform. It is of no research interest, but I have occasionally found it useful as a way of introducing NLG to potential project collaborators who know nothing about the field.

Motivation & Objective

• To provide a systematic overview of the three main stages in natural-language generation: content determination and text planning, sentence planning, and realization.
• To evaluate the trade-offs between different technical approaches—ranging from hard-coded solutions to AI-based planning—for implementing NLG systems.
• To guide practitioners in selecting appropriate techniques based on application requirements such as fluency, syntactic diversity, and available resources.
• To demonstrate that no single method is universally superior, but effectiveness depends on domain-specific constraints and goals.

Proposed method

• Content determination and text planning are performed simultaneously using techniques ranging from hard-coded programming logic to rule-based systems and planning algorithms.
• Sentence planning employs operations like conjunction, pronominalization, and discourse markers to enhance fluency without altering content.
• Realization is achieved through syntactic rules ensuring correct morphology, agreement, punctuation, and reflexive usage, with options ranging from full linguistic formalisms to simple template-based generation.
• Text-planning languages and macro packages are used to represent rhetorical structures and content flow, combining flexibility with programmatic power.
• The paper advocates for a pragmatic selection of techniques based on application needs, including the use of fill-in-the-blank templates for limited-domain systems.
• It emphasizes that effective NLG systems can be built using varying levels of linguistic sophistication, from minimal processing to complex formalisms like FUF and PENMAN.

Experimental results

Research questions

• RQ1 What are the key technical components required for building effective natural-language generation systems?
• RQ2 How do different approaches to content determination and text planning compare in terms of flexibility, robustness, and practicality?
• RQ3 To what extent can sentence planning improve text fluency without altering semantic content?
• RQ4 What are the trade-offs between using full linguistic formalisms and simpler template-based realization in NLG systems?
• RQ5 How should NLG system design be guided by application-specific requirements such as fluency, syntactic variety, and resource availability?

Key findings

• Content determination and text planning are typically performed simultaneously, with approaches ranging from hard-coded logic to advanced AI planning techniques.
• Sentence planning significantly enhances text fluency through operations like conjunction, pronominalization, and discourse markers, which preserve content while improving readability.
• Realization requires careful handling of morphology, agreement, punctuation, and reflexives, but can be effectively implemented using simple templates in low-variety domains.
• The choice of technique for each NLG stage depends critically on application-specific factors such as required fluency and syntactic diversity.
• No single method is universally superior; the most effective NLG systems result from selecting techniques that best match the domain’s constraints and goals.
• Practical NLG systems often combine multiple approaches—e.g., schema-based planning with template-based realization—to balance performance, cost, and maintainability.