TL;DR
This paper introduces a computational model for understanding and generating argumentative dialogs focused on planning alternatives. It proposes a knowledge-based approach using general, common-sense justification rules to infer and construct belief justifications, enabling systems to handle novel disagreements.
What Problem Does This Paper Address?
Discourse processing, particularly in argumentative dialogs, faces two major challenges: understanding the connections between utterances and generating coherent responses. This paper focuses on a specific class of argumentative dialogs where each response presents a belief that either justifies or contradicts a previously stated or inferred belief. The core problem is how a system can effectively understand the underlying reasoning behind these justifications and then produce its own coherent justifications, especially when dealing with novel arguments or disagreements. Existing models often focus on plan recognition or template-based responses, but this paper aims to address the deeper inferential and generative aspects of belief justification in argumentative dialogs.
Methods
The paper's approach is built on the assumption that dialog participants justify beliefs using instantiations of general, common-sense justification rules. For plan-oriented beliefs, these rules are essentially planning heuristics based on structural features of plans.
The model tackles two main tasks:
- Recognizing Justifications: This involves a forward-chaining search to find a chain of instantiated justification rules that contains the user's stated belief and justifies an earlier dialog belief or its negation. The process uses heuristics to prune search paths and identify the most relevant justification.
- Forming Justifications: This is approached via backward chaining. The model searches for justification rules that support a given belief and can be instantiated with existing beliefs. It prioritizes using existing justifications before attempting to construct new ones, aiming to discover agreements or disagreements with the participant's reasoning.
A key aspect is the use of both factual (objective judgments about planning relationships) and evaluative (subjective judgments about whether a plan should be executed) beliefs, with a set of general justification rules for each.
Datasets
Not discussed in the paper. The authors mention using the model in short argumentative dialogs in two domains: "day-to-day planning in the AI lab" and "removing and recovering files in UNIX," but these are described as domains for participation and experimentation rather than formal datasets.
Key Insights and Ideas
- Generalization is Key: The core idea is that participants use general, common-sense justification rules, not just specific knowledge. This allows systems to understand and generate justifications for novel arguments.
- Belief Justification as a Core Component: The paper argues that understanding and generating belief justifications are central to argumentative dialog, going beyond simple plan recognition.
- Bridging Understanding and Generation: The model uses the same set of justification rules for both recognizing existing justifications and forming new ones, creating a unified approach.
- Handling Novelty: By relying on general rules, the system can potentially handle belief justifications it has never seen before, as long as they are formed from these underlying rules.
- Evaluative Beliefs Matter: The inclusion of evaluative beliefs (e.g., "should be executed") is crucial for argumentation, as they represent subjective judgments that are often the subject of disagreement.
Key Results
The paper describes a fully implemented model written in Quintus Prolog. It includes 30 justification rules and over 400 beliefs about various plans. The model has been used to participate in short argumentative dialogs. The authors do not present quantitative results in terms of accuracy or performance metrics, focusing instead on the conceptual framework and the implemented capabilities.
Limitations and Future Work
The authors acknowledge several limitations:
- Plan Representation: The model treats plans as atomic units and focuses on "all or nothing" beliefs, limiting its ability to handle step-by-step reasoning, conditional planning, or beliefs not directly related to plans.
- Response Directionality: The current model can only understand responses that justify an earlier belief. It cannot yet understand responses that contradict an inferred justification for an earlier belief.
- Reactivity: The model is reactive, producing responses only when it perceives a disagreement. It needs to be extended to understand why it's in an argument and to be aware of the underlying goals of other participants.
Future work suggested includes addressing these limitations, particularly by extending the model to recognize more complex relationships and to be more proactive in dialog.
Summary updated