Eliminating deceptions and mistaken belief to infer conversational implicature

Mark Lee, Yorick Wilks

Abstract

Conversational implicatures are usually described as being licensed by the disobeying or flouting of some principle by the speaker in cooperative dialogue. However, such work has failed to distinguish cases of the speaker flouting such a principle from cases where the speaker is either deceptive or holds a mistaken belief. In this paper, we demonstrate how the three different cases can be distinguished in terms of the beliefs ascribed to the speaker of the utterance. We argue that in the act of distinguishing the speaker's intention and ascribing such beliefs, the intended inference can be made by the hearer. This theory is implemented in ViewGen, a pre-existing belief modelling system used in a medical counselling domain.

Introduction

In human conversation, participants frequently employ non-literal language—such as irony or sarcasm—to communicate deeper truths. In linguistic terms, this is known as "flouting" conversational maxims. For an AI dialogue system, the primary challenge lies in distinguishing this intentional flouting from two other common scenarios: deception, where the speaker lies to mislead, and mistaken belief, where the speaker is simply incorrect.

Traditional computational models often assume a baseline of cooperation and truthfulness, leaving them ill-equipped to handle the nuances of conflict or irony. This paper presents a formal framework using the ViewGen belief-modeling system to distinguish these states by systematically analyzing the beliefs ascribed to the speaker.

Modeling Belief Environments

The core of the proposed approach is the use of nested belief environments. ViewGen, implemented in Prolog, represents beliefs not just as flat facts, but as nested structures (e.g., "The System believes that Simon believes that David is happy"). This allows the system to maintain a distinction between its own knowledge and the knowledge it attributes to other agents.

The system employs two primary mechanisms for belief representation:

  1. Default and Stereotypical Ascription: The system assumes others share its beliefs unless evidence suggests otherwise, or it uses role-based "stereotypes" to fill in gaps.
  2. Plan Recognition: Using a partial order clausal link planner, the system attempts to reconstruct the speaker's goal. When a statement contradicts known facts, a diagnostic process is triggered to determine why the "infelicity" occurred.
Figure 1: Graphical distinction between deception and mistaken belief

As shown in Figure 1, the distinction between deception and mistaken belief hinges on the speaker's internal state. In a mistaken belief, the speaker genuinely believes $P$ while the system knows $\neg P$. In deception, the speaker believes $\neg P$ but intends for the system to believe they believe $P$.

The Logic of Elimination

The paper posits that conversational implicature (like irony) is detected through a process of elimination. For irony to be successful, the speaker must believe that the hearer knows the statement is false. If the speaker isn't certain the hearer will catch the error, they risk being perceived as a liar or as being mistaken.

To recognize an implicature, the system must prove that the speaker intended for the hearer to recognize the violation of standard "felicity conditions." If the system can rule out both deception and mistaken belief, the statement is accepted as a pragmatically communicative act.

Figure 2: A recognized plan explaining Simon's infelicity

Figure 2 illustrates this process in a "toy" example where a user, Simon, reacts to the system deleting a subdirectory. By "blocking" the standard interpretation of Simon's utterance (that he is happy about the deletion), the system ascribes a goal of irony, successfully inferring the intended meaning.

Key Results and Formalization

The research successfully formalizes the belief states required for these three distinct phenomena:

  • Mistaken Belief: $bel(System, bel(Speaker, P))$ while the system knows $not(P)$.
  • Deception: $bel(Speaker, bel(System, bel(Speaker, P)))$ where the speaker wants the system to believe they believe $P$, despite actually believing $not(P)$.
  • Implicature (Irony): The speaker intends for the system to recognize that the standard rules for a sincere statement have been violated, prompting a search for a secondary, plan-based meaning.

Discussion and Limitations

While the framework provides a robust logical foundation, several challenges remain. The current definition of deception is relatively strict and does not account for complex, multi-step "gaslighting" strategies. Furthermore, while the system can detect that an implicature has occurred, providing a universal method to determine the exact implied message across all speech types remains an area for future work. Finally, implementing these logical rules in real-time systems requires balancing computational tractability with the underspecified nature of principles like "politeness" or "relevance."

Summary updated

Abstract

Conversational implicatures are usually described as being licensed by the disobeying or flouting of some principle by the speaker in cooperative dialogue. However, such work has failed to distinguish cases of the speaker flouting such a principle from cases where the speaker is either deceptive or holds a mistaken belief. In this paper, we demonstrate how the three different cases can be distinguished in terms of the beliefs ascribed to the speaker of the utterance. We argue that in the act of distinguishing the speaker's intention and ascribing such beliefs, the intended inference can be made by the hearer. This theory is implemented in ViewGen, a pre-existing belief modelling system used in a medical counselling domain.
BibTeX
@misc{lee1998eliminating,
  author = {Mark Lee and Yorick Wilks},
  title = {{Eliminating deceptions and mistaken belief to infer conversational implicature}},
  year = {1998},
  eprint = {cmp-lg/9806005},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/cmp-lg/9806005},
}
Figure