A Sound and Complete Dialogue System for Handling Misunderstandings

Andreas Xydis, Christopher Hampson, Sanjay Modgil, Elizabeth BlackSAFA@COMMA 2022 · OpenReview

Abstract

In formal argumentation, agents often exchange incomplete arguments, known as enthymemes, to improve communication efficiency. However, this pragmatic shorthand frequently leads to misunderstandings when interlocutors fill in missing premises differently. This paper presents a formal dialogue system based on the ASPIC+ framework that explicitly models and resolves such misunderstandings. By introducing specific locutions for clarification and correction, the system ensures that dialogue outcomes remain logically consistent with the underlying complete arguments. We demonstrate the system's soundness and completeness, proving that the use of enthymemes does not compromise the logical rigor of the argumentation process.

Introduction

In formal logic and artificial intelligence, arguments are typically treated as mathematical proofs where every premise is explicitly stated. However, human communication relies heavily on enthymemes—incomplete arguments where "obvious" information is omitted for brevity. While efficient, this practice introduces the risk of misunderstanding: what is obvious to the speaker may not be obvious to the listener.

Presented at SAFA'22, this research introduces a formal dialogue system designed to identify and resolve these logical gaps. The system bridges the divide between the pragmatic shorthand used in real-world dialogue and the rigorous requirements of structured argumentation theory.

Modeling Enthymemes and Misunderstandings

The authors build their system upon the ASPIC+ framework, but they represent arguments as labeled, downward-directed trees rather than simple sets of premises. This structural approach allows the system to formalize misunderstanding not as a simple error, but as a mismatch between the intended argument of the sender and the assumed argument of the receiver.

The relationship between an assumed argument (a), an intended argument (b), and the resulting enthymeme (c).

As illustrated in the figure above, an agent may intend a complex logical structure (b), but only communicate a subset of it as an enthymeme (c). The receiver, lacking the full context, may reconstruct an assumed argument (a) that differs from the original intention. Misunderstandings occur when these hidden logical connections do not align.

The Dialogue System: Locutions and Framework

To resolve these mismatches, the paper introduces a set of specific dialogue moves (locutions) that allow agents to probe the reasoning of their peers:

  • assert: Proposing an enthymeme.
  • w.d.y.i. (What do you intend?): Requesting the full argument behind a shorthand assertion.
  • w.d.y.a.b. (What did you assume by...?): Questioning the receiver's interpretation of a move.
  • meant / agree: Correcting or confirming a specific interpretation.

These moves are tracked within a Dialogue Framework (DF), a graph structure that monitors the status of assertions—determining which are "in" (supported) or "out" (defeated) as the conversation progresses. This process "reifies" hidden relations, forcing agents to make their underlying premises explicit when challenged.

Key Results: Soundness and Completeness

The primary contribution of this work is the formal proof of the system's Soundness and Completeness. These properties ensure that there is no logical penalty for using shorthand in a dialogue:

  1. Soundness: If a claim is deemed acceptable at the end of a dialogue, its underlying intended argument is logically valid within the formal Argumentation Framework (AF).
  2. Completeness: If an argument is logically valid in the AF, there exists a sequence of dialogue moves that can establish its acceptance within the dialogue system.

Crucially, these results hold under the "Honest and Exhaustive" condition, where agents do not lie about their intentions and continue to make moves until no further relevant information can be shared.

Limitations and Future Work

While the system provides a robust foundation for resolving misunderstandings, it currently assumes that agents share the same underlying logic, even if their knowledge bases differ. Future research will explore scenarios where agents use different logical rules or act strategically (dishonestly) to gain an advantage. Additionally, the computational complexity of searching for these dialogue paths remains an open area for optimization.

Summary updated

BibTeX
@inproceedings{xydis2022sound,
  author = {Andreas Xydis and Christopher Hampson and Sanjay Modgil and Elizabeth Black},
  title = {{A Sound and Complete Dialogue System for Handling Misunderstandings}},
  booktitle = {SAFA@COMMA},
  year = {2022},
  url = {https://openreview.net/forum?id=oFOkbXj1hh},
}
Figure