Recursion in TEXT and Its Use in Language Generation

Kathleen R. McKeownAAAI 1983 · OpenReview

Abstract: In this paper, I show how textual structure is recursive in nature; that is, the same rhetorical strategies that are available for constructing the text's macro-structure are available for constructing its sub-sequences as well, resulting in a hierarchically structured text. The recursive formalism presented can be used by a generation system to vary the amount of detail it presents for the same discourse goal in different situations.

Introduction

In the early development of Natural Language Generation (NLG), systems often suffered from a lack of flexibility. When tasked with defining a term or explaining a concept, these systems typically produced static, one-sentence responses that failed to account for the user's expertise or the complexity of the subject matter. Kathleen McKeown’s 1983 research addressed this gap by proposing that human text is inherently recursive. By implementing this "textual recursion" in the TEXT generation system, McKeown demonstrated how AI could dynamically adjust the level of detail in its explanations based on specific discourse goals.

The Formalism: Predicates and Schemata

The TEXT system relies on a formalism built from two primary components: Rhetorical Predicates and Schemata.

  1. Rhetorical Predicates: These are the functional building blocks of a text. They represent specific communicative acts, such as identification (defining an object), analogy (comparing an object to something familiar), or particular-illustration (providing a specific example).
  2. Schemata: These serve as structural blueprints, defining how rhetorical predicates are typically sequenced in natural language. For instance, an "Identification Schema" might dictate a sequence starting with an identification predicate, followed by attribute descriptions, and concluding with an illustration.

The Innovation of Textual Recursion

The core contribution of this work is the introduction of recursion into the schema application process. In previous NLG systems, a predicate usually mapped directly to a single sentence. McKeown’s approach allows a predicate to "expand" into an entire schema of its own.

This recursive expansion is triggered by the system's assessment of the user’s knowledge. If the system is defining a complex object (e.g., a "Hobie Cat") and determines that the user is unfamiliar with a component term (e.g., "catamaran"), it can recursively call the Identification Schema to define the component within the larger explanation. This allows the system to generate a nested, hierarchical structure where a single point in a high-level explanation becomes the subject of its own sub-explanation.

Implementation and Results

The TEXT system was implemented using a database from the Office of Naval Research (ONR), containing structured data on military vehicles and weaponry, such as destroyers and projectiles. The system's ability to generate coherent, multi-paragraph text was demonstrated through several key examples:

  • The Hobie Cat Example: This demonstrated the system’s ability to scale detail. It could move from a concise three-sentence definition to a complex version where the initial sentence expanded into a four-sentence sub-sequence explaining the nature of a catamaran.
  • The Destroyer vs. Bomb Example: When asked to compare two vastly different objects, the system recognized the lack of shared context. It recursively defined the generic classes of each (surface ships vs. projectiles) before attempting a comparison, ensuring the final output was meaningful to the user.

Insights and Constraints

McKeown’s work suggests that text structure mirrors thought and task structures; just as a complex task has sub-tasks, a complex explanation requires sub-explanations. By using a finite set of rules (the paper utilizes four main schemata and ten predicates), the system can generate an infinite variety of texts tailored to the situation.

However, the paper also notes significant constraints. While a computer can handle deep recursion, human readers are limited by memory. If a sub-explanation becomes too lengthy or deep, the reader may lose track of the primary discourse goal. Furthermore, accurately modeling a user's goals and knowledge remains a significant challenge for automated systems.

Conclusion

By treating text as a recursive, hierarchical structure rather than a linear sequence of sentences, the TEXT system provided a framework for more natural and flexible AI communication. This approach allows for the generation of "well-planned writing" that can adapt its depth and complexity to meet the needs of its audience, a foundational concept in the evolution of computational linguistics.

Summary updated

Abstract

In this paper, I show how textual structure is recursive in nature; that is, the same rhetorical strategies that are available for constructing the text's macro-structure are available for constructing its sub-sequences as well, resulting in a hierarchically structured text. The recursive formalism presented can be used by a generation system to vary the amount of detail it presents for the same discourse goal in different situations.
BibTeX
@inproceedings{mckeown1983recursion,
  author = {Kathleen R. McKeown},
  title = {{Recursion in TEXT and Its Use in Language Generation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
  url = {https://openreview.net/forum?id=SyZFGg-_ZH},
}
Figure