Bayesian Reasoning in an Abductive Mechanism for Argument Generation and Analysis

AAAI 1998

TL;DR

This paper introduces NAG, a system that generates and analyzes arguments using Bayesian networks. NAG aims to create arguments that are both normatively sound and persuasive to a specific audience by modeling beliefs and reasoning in two separate Bayesian networks. This approach allows for a nuanced understanding of argument quality beyond simple logical validity.

What Problem Does This Paper Address?

The paper tackles the challenge of creating arguments that are not only logically correct but also convincing to a particular listener. Traditional argument generation often focuses on logical structure, neglecting the audience's existing beliefs and how an argument might influence them. Furthermore, analyzing arguments for their persuasive power and normative strength in a computationally tractable way is difficult, especially when dealing with complex, real-world knowledge. The authors aim to bridge this gap by developing a system that can generate arguments that are "nice" – meaning they are both normatively strong and persuasive.

Methods

NAG employs a novel approach that combines Bayesian reasoning with an abductive process and an attentional mechanism.

  • Bayesian Networks (BNs): NAG uses two primary Bayesian networks: one representing a normative model of correct reasoning and another representing a user model of the target audience's beliefs. These BNs are crucial for quantifying uncertainty and estimating the impact of arguments.
  • Abductive Process: Argument generation is framed as an abductive process. The system starts with a goal proposition and iteratively expands an "Argument Graph" (a subnetwork of the BN) by inferring plausible premises and links that support the goal. This is achieved through "Reasoning Agents" that consult various knowledge bases.
  • Attentional Mechanism: To manage the computational complexity of full BN propagation, NAG uses a focusing mechanism inspired by human attention. This mechanism identifies and activates relevant subnetworks within the larger BNs, effectively narrowing down the scope of reasoning to the most pertinent information for the current argument. This is implemented using semantic networks (SNs) that capture associative connections.
  • Generation-Analysis Cycle: The core of NAG's operation is an iterative cycle. The "Strategist" initiates argument generation, which involves the "Generator" extending the Argument Graph and the "Analyzer" evaluating its "niceness" (normative strength and persuasiveness). If the argument isn't satisfactory, the cycle repeats, refining the graph.

Datasets

Not discussed in the paper. The paper describes the system's architecture and methodology, using examples like the "asteroid impact" scenario for illustration, but does not detail specific datasets used for training or evaluation.

Key Insights and Ideas

  • Dual Models for Argument Quality: The most significant insight is the use of two distinct Bayesian networks – one normative and one for the user model. This allows NAG to simultaneously assess an argument's logical correctness and its potential to persuade a specific individual, acknowledging that these two aspects are not always aligned.
  • Attention as a Computational Strategy: Emulating human attention through spreading activation on semantic networks is a clever way to tackle the NP-hard problem of Bayesian inference in large networks. This significantly reduces computational load without drastically sacrificing argument quality.
  • Argument as Abduction: Framing argument generation as an abductive process, where the system seeks the best explanation (premises and inferences) for a goal proposition, provides a robust framework for building coherent and supportive arguments.
  • Iterative Refinement: The generation-analysis cycle allows for continuous improvement of arguments. If an argument is found to be weak, the system can automatically identify flaws and attempt to strengthen it, mimicking a human revision process.

Key Results

  • Reduced Generation Time: The evaluation showed that the attentional mechanism substantially reduces argument generation times. For instance, in the "asteroid" scenario, argument generation time with spreading activation was 12.5 CPU seconds, compared to 25 seconds without it. Similar speedups were observed across other test scenarios (Table 1).
  • Argument Quality Maintained: Crucially, this speedup did not come at a significant cost to argument quality. In most runs, the same arguments were generated with or without spreading activation.
  • Persuasiveness Demonstrated: A preliminary web-based evaluation with 32 respondents indicated that NAG's arguments were reasonably persuasive. Respondents showed a clear tendency to shift their beliefs towards target propositions after being presented with NAG's arguments, with significant shifts observed for key propositions.

Limitations and Future Work

  • Manual KB Construction: The initial semantic-Bayesian networks are currently built manually, which is a significant bottleneck for scalability. Future work could involve automating this process.
  • Limited Expressiveness of Argument Graphs: The current Argument Graph primarily incorporates causal or evidential links. The authors acknowledge that other types of relations might be important for a complete argument.
  • Simplistic User Model: While the user model is revisable, it's based on a single BN. More sophisticated models of human cognition and belief updating could be explored.
  • Natural Language Generation: The paper mentions that methods for rendering NAG's output in English are being considered, but the current output is primarily graphical. Developing robust NLG capabilities is a clear future direction.
  • Argument Presentation Optimization: The paper notes that NAG attempts to minimize the size of the Argument Graph during presentation, but further research into optimizing this process and generating more concise arguments is ongoing.

Summary updated

BibTeX
@inproceedings{1998bayesian,
  title = {{Bayesian Reasoning in an Abductive Mechanism for Argument Generation and Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1998},
}
Figure