This 1993 paper from the AAAI proceedings remains a foundational look at how AI systems can move beyond simple "data dumps" to provide sophisticated, human-like justifications for their actions. By formalizing argumentation as a set of plan-based "communicative acts," the work demonstrates how AI can dynamically generate persuasive natural language that accounts for a user's existing beliefs and goals.
The Problem: Expert Systems as "Black Boxes"
In the early 1990s, knowledge-based systems (expert systems) were becoming increasingly capable but remained largely opaque to their users. While these systems could reach complex conclusions, they struggled to explain why those conclusions were valid or why a user should follow a specific recommendation.
The author identifies a critical gap: existing systems were proficient at generating descriptions (explaining what something is), but they lacked a computational model for argumentation—the act of convincing a user to believe a proposition or perform an action. Most explanations at the time were either "canned" text templates or dry lists of inference rules that lacked rhetorical cohesion and failed to consider the hearer's mental state.
A Plan-Based Approach to Argumentation
The paper proposes treating communication as a sequence of purposeful acts within a hierarchical planning framework. Using a classical AI planning approach, an "argument" is treated as a high-level goal that can be decomposed into smaller, manageable sub-acts.
As shown in the hierarchy above, the system categorizes argumentation into three primary branches:
- Deduce: Using logical structures like categorical or hypothetical syllogisms.
- Induce: Providing evidence, illustrations, or analogies.
- Persuade: Motivating action by indicating desirable consequences or plans.
Plan Operators and Preference Metrics
Each communicative act is defined by a formal operator consisting of four main components:
- Constraints: Hard requirements for the act to be applicable (e.g., the subject must be a valid proposition).
- Preconditions: The mental state of the user (e.g., does the user already believe this? Do they have the necessary knowledge to act?).
- Effects: The intended change in the user's mental state (e.g., the user now believes $X$).
- Decomposition: The specific steps to take, such as claiming a point and then providing supporting evidence.
Since multiple strategies can often argue the same point, the system employs "preference metrics" to choose the optimal plan. It prioritizes plans that are "cognitively economical" (requiring fewer subgoals) or those that mirror effective human discourse patterns, such as preferring deductive reasoning over induction when possible.
Application and Results
The approach was demonstrated within the Knowledge-based Replanning System (KRS), a resource allocation system used for Air Force mission planning. The system generates justifications for why certain mission parameters, such as aircraft types or targets, are in conflict.
Before this research, the KRS system produced a numbered list of facts (as seen in the figure above). While accurate, this "baseline" required the user to manually connect the dots to understand the underlying logic. In contrast, the new plan-based system generates cohesive paragraphs:
"You should replace F-111e aircraft with F-4g aircraft... A conflict between the aircraft and the target... motivates replacing [them]."
By integrating the text planner with the KRS system, the author proved that high-level communicative goals could be mapped directly to low-level semantic relations in a complex, real-world domain.
Conclusion and Future Directions
The core insight of this work is treating an argument not as a static piece of text, but as a strategic action intended to change a hearer's state. By distinguishing between convincing (changing beliefs) and persuading (changing goals), the system can tailor its strategy to the specific needs of the user.
While foundational, the author noted several areas for future work, including better representation of user intentions, accounting for how context changes the "force" of an argument, and modeling multi-functional communication where a single act might describe, compare, and persuade simultaneously. Additionally, the paper hints at the need for multimodality—coordinating linguistic acts with visual cues like map highlights to create a truly unified argument.
Summary updated