Towards an AI Model of Argumentation

Lawrence Birnbaum, Margot Flowers, Rod McGuireAAAI 1980

Abstract

In this seminal 1980 paper from the Yale AI lab, researchers propose a computational model for human argumentation. By implementing the ABDUL/ILANA system, the authors demonstrate how an AI can adopt a specific point of view in a political debate, utilizing structured tactics and goal-directed memory search to defend its position and challenge opponents.

Introduction

Early Artificial Intelligence research often struggled with the transition from formal logic to the messy reality of human discourse. This paper addresses that gap by deconstructing argumentation into a strategic, computational process. Rather than exhaustive inference, the authors argue that human debate is driven by specific goals: winning a point or defending a belief.

The Problem: Beyond Blind Inference

By 1980, AI systems were proficient at "blind inference"—deriving all possible conclusions from a set of facts. However, human arguers are far more selective. We coordinate reasoning, memory, and conversational rules to achieve persuasive ends. The authors sought to model the sub-tasks of argumentation, the "geometry" of how points interact, and how a "point of view" (POV) influences the interpretation of historical data.

The ABDUL/ILANA System

The researchers developed ABDUL/ILANA, a program capable of natural language dialogue focused on the 1967 Arab-Israeli War. The system's architecture relies on several key components:

  • Argument Tactics: Heuristic rules for navigating debates. For instance, to challenge a claim, the system might attack the claim directly, attack its supporting evidence, or attack the logical link between the two.
  • Causal-Temporal (CT) Search: A memory retrieval method that searches for event chains to verify claims or find rebuttals.
  • Point of View (POV) Slanting: "Gut reaction" rules (e.g., "blame the opponent for negative outcomes") provide an initial stance, which the system then justifies through inferential memory.

The Geometry of Argument

Central to the paper is the concept of an Argument Graph. In this network, propositions serve as nodes connected by two primary relations: Support and Attack.

Figure 1: An argument graph illustrating support and attack relations between conflicting claims.

As shown in the figure above, the system represents conflicting perspectives (e.g., "The Arabs started the war" vs. "The Israelis started the war") and the evidence cited for each. By treating the argument as a graph, the AI can use search rules to identify weak points in an opponent's logic or strengths in its own.

Key Insights

  • Reasoning is Goal-Directed: Inference is performed only when it serves to understand an attack or formulate a counter-argument.
  • Rebuttals as Side-Effects: A powerful insight of the model is that rebuttals are often discovered during the process of trying to understand an opponent's claim. Searching memory to verify a statement may "accidentally" reveal contradictory facts.
  • Interpretation Over Facts: The authors demonstrate that arguments are rarely about raw data but about the interpretation of those facts. Both sides of the AI use the same database but apply different "sympathies" to reach divergent conclusions.

Results and Limitations

The ABDUL/ILANA system successfully demonstrated the ability to identify unstated claims, select appropriate tactics based on the "geometry" of the graph, and engage in natural language interaction. However, the system remained limited by its hand-coded knowledge base and the modular separation between its linguistic and argumentative components.

Conclusion

This work represents a foundational step in modeling the strategic nature of human communication. By moving from logic to tactics, the authors provided a blueprint for AI that does not just "think," but argues with purpose.

Reference: Birnbaum, L., Flowers, M., & McGuire, R. (1980). Towards an AI Model of Argumentation. Proceedings of the First National Conference on Artificial Intelligence (AAAI-80).

Summary updated

BibTeX
@inproceedings{birnbaum1980towards,
  author = {Lawrence Birnbaum and Margot Flowers and Rod McGuire},
  title = {{Towards an AI Model of Argumentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure