Two Types of Planning in Language Generation

Eduard H. HovyACL 1988 · OpenReview

Abstract

As our understanding of natural language generation has increased, a number of tasks have been separated from realization and put together under the heading "text planning". So far, however, no-one has enumerated the kinds of tasks a text planner should be able to do. This paper describes the principal lesson learned in combining a number of planning tasks in a planner-realizer: planning and realization should be interleaved, in a limited-commitment planning paradigm, to perform two types of planning: prescriptive and restrictive. Limited-commitment planning consists of both prescriptive (hierarchical expansion) planning and of restrictive planning (selecting from options with reference to the status of active goals). At present, existing text planners use prescriptive plans exclusively. However, a large class of planner tasks, especially those concerned with the pragmatic (non-literal) content of text such as style and slant, is most easily performed under restrictive planning. The kinds of tasks suited to each planning style are listed, and a program that uses both styles is described.

Introduction

In the late 1980s, Natural Language Generation (NLG) reached a critical architectural juncture. Most existing systems relied on "top-down" planning, where the system decided exactly what to say and in what order before selecting any specific words. While effective for simple, factual reports, this rigid approach struggled with pragmatics—the nuanced adjustment of style, tone, and "slant" (bias) to suit a specific context.

Eduard Hovy argues that traditional planning is too narrow to handle the interleaved nature of human speech. Pragmatic goals like sounding "formal" or "sympathetic" cannot be treated as a single checklist item; they must influence every decision throughout the generation process. To address this, Hovy introduced PAULINE (Planning And Uttering Language In Natural Environments), a system that breaks away from rigid hierarchies by interleaving high-level planning with in-line realization.

Architecture and Methodology: Limited-Commitment Planning

The core innovation of PAULINE is a "limited-commitment" planning paradigm. Rather than committing to a full structure upfront, the system divides labor between two distinct planning styles that operate in tandem.

The architecture of PAULINE, showing the interaction between rhetorical goals and the stages of text generation.
  1. Prescriptive Planning (Top-Down): This handles the "big picture" content selection. Using schemas such as CONVINCE, RELATE, or DESCRIBE, the system identifies relevant information and establishes a general organizational flow.
  2. Restrictive Planning (Bottom-Up): This occurs "in-line" during the realization of sentences. When the system encounters multiple ways to express an idea (e.g., choosing between "a few students" and "a large number of students"), the restrictive planner evaluates these options against the current state of pragmatic goals.

To manage these processes, PAULINE utilizes a Monitoring component. It tracks the "satisfaction status" of various goals, such as brevity, formality, or sympathy toward a specific group. When goals conflict, the system employs a "least-satisfied" strategy, prioritizing the goal that has been most neglected in the text generated thus far.

Key Insights: Interleaving and Persistent Goals

Hovy’s work highlights several critical insights for advanced NLG:

  • Interleaving is Essential: Planning and realization should not be separate, sequential stages. Interleaving allows the system to capitalize on "fortuitous syntactic opportunities" that emerge during the word-selection process.
  • Persistent Pragmatic Goals: Unlike a goal to "mention a date," which is satisfied once, goals like "be arrogant" must remain active throughout the entire generation process to maintain stylistic consistency.
  • In-Line Decision Making: Style is often the cumulative result of many small, local decisions—such as adjective choice, conjunctions, and sentence length—rather than a single high-level structural choice.

Results and Evaluation

PAULINE was tested using a representation network of approximately 120 elements based on Conceptual Dependency. The primary case study involved generating accounts of a 1986 protest at Yale University regarding apartheid.

By toggling pragmatic parameters, the system could produce over 100 variations of the same underlying event. Examples included:

  • Neutral: A factual, detached account.
  • Formal/Pro-Protester: Utilizing complex syntax ("As a reminder to Yale..."), formal vocabulary ("erected"), and sympathetic phrasing ("concerned students").
  • Informal/Pro-University: Using colloquialisms ("It pisses me off"), shorter sentences, and biased labels ("shiftless students").

Limitations and Future Directions

While groundbreaking, Hovy acknowledges the "ad hoc" nature of PAULINE. To cover a wide range of pragmatic effects, many internal specialists relied on simplified "rule-of-thumb" methods rather than deep theoretical models. Future work suggests the need for even more sophisticated architectures to ensure that various generation tasks work harmoniously, moving toward a more integrated "what" (planner) and "how" (realizer) framework for human-like communication.

Summary updated

Abstract

As our understanding of natural language generation has increased, a number of tasks have been separated from realization and put together under the heading "text planning". So far, however, no-one has enumerated the kinds of tasks a text planner should be able to do. This paper describes the principal lesson learned in combining a number of planning tasks in a planner-realizer: planning and realization should be interleaved, in a limited-commitment planning paradigm, to perform two types of planning: prescriptive and restrictive. Limited-commitment planning consists of both prescriptive (hierarchical expansion) planning and of restrictive planning (selecting from options with reference to the status of active goals). At present, existing text planners use prescriptive plans exclusively. However, a large class of planner tasks, especially those concerned with the pragmatic (non-literal) content of text such as style and slant, is most easily performed under restrictive planning. The kinds of tasks suited to each planning style are listed, and a program that uses both styles is described.
BibTeX
@inproceedings{hovy1988two,
  author = {Eduard H. Hovy},
  title = {{Two Types of Planning in Language Generation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1988},
  url = {https://openreview.net/forum?id=ryZdLieuWr},
}
Figure