Planning Text for Advisory Dialogues

Johanna D. Moore, Cécile ParisACL 1989 · DOI

Abstract

This paper introduces a text planning architecture for interactive advisory dialogues that maintains a rich representation of its own communicative intentions. By integrating Intentional, Attentional, and Rhetorical structures into a hierarchical plan, the system can reason about its previous utterances to handle follow-up questions and clarify misunderstandings. The approach is demonstrated through the Program Enhancement Advisor (PEA), showing how plan-based generation enables more flexible and robust human-computer interaction than traditional schema-based methods.

Introduction

In the late 1980s, Natural Language Generation (NLG) systems typically produced static, "one-shot" text. If a user encountered a confusing explanation, these systems lacked the internal "self-awareness" to explain why they chose a particular phrasing or to offer an alternative strategy. Moore and Paris identified that a successful advisor must be able to reason about its own utterances and handle dialogue failures gracefully. They proposed that a system needs to understand which parts of its explanation were intended to persuade, which were background, and how to pivot if the user remains confused.

Methodology: Hierarchical Expansion Planning

The core of the system is a top-down hierarchical expansion planner inspired by Sacerdoti’s work. It constructs a tree that simultaneously tracks three distinct layers of discourse:

  1. Intentional Structure: The high-level goals of the speaker (e.g., "persuade the user to replace a function").
  2. Attentional Structure: The objects and concepts currently in focus.
  3. Rhetorical Structure: Based on Rhetorical Structure Theory (RST), this defines the relationships between text segments, such as "Motivation," "Contrast," or "Elaboration."

The planner uses "Plan Operators" which consist of effects (goals achieved), constraints (user model conditions), a nucleus (core information), and satellites (supporting information).

A hierarchical plan tree showing the integration of intentional goals and rhetorical relations.

Case Study: The Program Enhancement Advisor (PEA)

The system was implemented within the Program Enhancement Advisor (PEA), a tool that recommends Lisp code improvements. For instance, it might suggest replacing setq with setf to enhance maintainability.

A sample dialogue between the system and a user demonstrating follow-up questions and clarification.

As shown in the sample dialogue, when a user asks "What is a generalized variable?", the system doesn't just repeat itself. It looks back at its plan tree, identifies that the previous "Class-Ascription" strategy might have failed, and can then select a different operator, such as providing examples (e.g., car and cdr of a cons).

Key Insights

  • Plan Repair: By recording the rationale behind every utterance, the system can perform "plan repair" when a user is confused. If a specific "Motivation" satellite fails, the system can try a different rhetorical tactic.
  • User Modeling with Assumptions: The planner makes assumptions about user knowledge. If a follow-up question indicates a lack of understanding, the system updates its user model and avoids repeating the same mistake in the future.
  • Flexibility: This approach is significantly more flexible than template-based "schema" systems, allowing for dynamic, multi-sentential explanations tailored to the ongoing conversation.

Conclusion

While the planning mechanism is general, the plan operators require significant domain-specific engineering. Additionally, the system relies on external components like the Penman system for linguistic realization. Despite these limitations, the work laid a foundational framework for modern conversational agents by emphasizing that text generation is a goal-directed, interactive process rather than a static output task.

Summary updated

BibTeX
@inproceedings{moore1989planning,
  author = {Johanna D. Moore and Cécile Paris},
  title = {{Planning Text for Advisory Dialogues}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1989},
  doi = {10.3115/981623.981648},
}
Figure