Modelling the Persuadee in Asymmetric Argumentation Dialogues for Persuasion

Anthony HunterIJCAI 2015

In the realm of artificial intelligence, persuasion is often modeled as a symmetric debate between two experts. However, real-world applications—such as health coaching apps or educational tools—frequently involve an asymmetric relationship where a sophisticated system interacts with a human user who may only provide limited feedback. This paper, presented at IJCAI 2015 by Anthony Hunter, establishes a formal theoretical foundation for asymmetric persuasion dialogues. By combining abstract argumentation with probabilistic user modeling, the framework allows a system to strategically navigate a conversation to change a user's mind, even when the user's internal beliefs are not fully known.

The Framework of Asymmetric Dialogues

The core challenge in automated persuasion is the "Natural Language Understanding" bottleneck. It is difficult for current systems to parse complex, free-form counter-arguments from a human. Hunter addresses this by proposing a restricted dialogue format where the system performs the heavy lifting—presenting arguments and asking targeted questions—while the user responds with simple "Yes/No" feedback.

The underlying structure of these dialogues is based on Abstract Argumentation. In this model, arguments are treated as nodes in a directed graph, and the relationships between them are defined by "attacks" (edges). For example, an argument for eating a banana might be attacked by an argument regarding its sugar content.

Figure 2: An example of an abstract argumentation graph where multiple lines of reasoning (G, H) converge on intermediate arguments (E, F, D) to eventually influence the goal argument (A).

Probabilistic User Modeling

Because the system cannot see inside the user's mind, it maintains a probabilistic user model. Instead of assuming a binary state (the user either accepts or rejects an argument), the system assigns a probability (ranging from 0 to 1) to the user's belief in each node of the argument graph. As the dialogue progresses, the system uses the user's feedback to update this distribution.

The system has three primary moves at its disposal:

  1. Posit ($A!$): Presenting an argument to the user to influence their belief.
  2. Query ($A?$): Asking if the user believes a specific argument to refine the user model.
  3. Terminate ($\perp$): Ending the dialogue once the goal is reached or utility is exhausted.

Mass Redistribution and Update Strategies

The "brain" of the system lies in how it updates its model after a move. The paper defines four distinct strategies for "mass redistribution," which represent different psychological profiles of a persuadee:

  • Naive: The user is highly impressionable and accepts whatever they are told.
  • Trusting: The user accepts the posit and automatically discounts any arguments that conflict with it.
  • Strict: A more skeptical profile where the user only accepts a new argument if they already disbelieve its attackers.
  • Ambivalent: The user is cautious and only partially shifts their belief when presented with new information.

These strategies are evaluated against various logical properties, such as whether they are "rational" (following the logic of the graph) or "credulous."

Figure 6: A comparison of the four update methods (Naive, Trusting, Strict, Ambivalent) against different logical properties like Rationality and Restrictiveness.

Optimizing the Dialogue

Persuasion is treated as an optimization problem. The system calculates the Expected Utility of various dialogue paths, seeking the sequence of moves that maximizes the probability of the user believing the "goal argument" (e.g., "I should exercise today").

A key insight is the role of Queries. While queries do not directly persuade the user, they reduce the "informational distance" between the system's model and the user's actual beliefs. By narrowing down what the user thinks, the system can select more potent arguments later in the conversation. The paper also introduces "update trees," proving that if a structural path of "defenders" exists for a goal argument, persuasion can be mathematically guaranteed under certain update rules.

Key Results and Limitations

The research provides several formal proofs regarding dialogue selection. For instance, it demonstrates that under "Naive" or "Trusting" models, repeating an argument provides no additional utility, allowing the system to prune its search space significantly. Furthermore, the authors suggest a method for utility normalization to account for "user fatigue," penalizing excessively long dialogues to ensure the system remains concise.

Despite its theoretical strengths, the model faces practical hurdles. The computational complexity of maintaining a probability distribution over all possible subsets of arguments is $2^n$, which may limit scalability. Additionally, the model assumes a static argument graph; in reality, users may possess "hidden" arguments unknown to the system. Future work will involve empirical testing to determine which of the four update strategies most accurately reflects real-world human behavior.

Summary updated

BibTeX
@inproceedings{hunter2015modelling,
  author = {Anthony Hunter},
  title = {{Modelling the Persuadee in Asymmetric Argumentation Dialogues for Persuasion}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure