Signalling the Interpretation of Indicrect Speech Acts

Philip R. Cohen

Introduction

In everyday communication, the literal meaning of an utterance often diverges significantly from its intended purpose. A question such as "Is the salt near you?" is rarely a request for information; it is an indirect request for action. These are known as indirect speech acts. While computational linguistics has long theorized that listeners use "plan recognition" to bridge the gap between literal form and intended goal, empirical evidence for the specific inference paths taken by listeners has been difficult to isolate. This paper argues that the use of pronouns to refer to actions—a phenomenon known as anaphora—serves as a critical signal for how these indirect acts are interpreted within a shared plan.

A Plan-Based Theory of Language

Cohen utilizes a plan-based framework where speaking and physical actions are treated uniformly. In this model, communicative acts are "operators" that change the mental states of participants (their beliefs and goals) just as physical acts change the state of the world. A key distinction is made between the surface act (the literal form of the utterance, such as a declarative statement) and the illocutionary act (the intended force, such as a request).

By treating language as a component of a larger planning system, the paper suggests that reasoning systems—whether human or artificial—do not require a specialized logic for language. Instead, communicative acts are simply steps taken to achieve a goal within a logical framework of mental and physical states.

Inference Paths: Helpful vs. Intended

When a speaker makes a statement that leads to a listener performing an action, how does the listener arrive at that decision? The paper compares two potential inference paths:

  1. The "Helpful" Path: The listener hears a fact, updates their internal model of the world, and decides to perform an action of their own volition to be helpful.
  2. The "Intended" Path: The listener recognizes that the speaker intended for them to perform a specific action as part of a mutual, shared plan.

Distinguishing between these paths is vital for understanding conversational coherence and building AI that can truly cooperate with human users.

Anaphora as a Diagnostic Tool

The central insight of the paper is that reference—specifically the use of the pronoun "it"—acts as a diagnostic for interpretation. Cohen analyzes a dialogue fragment from a task-oriented experiment where an expert instructs a novice to assemble a toy water pump.

In the example "OK, it's done," the pronoun "it" refers to the action of putting a tube on the pump. For this reference to be successful, the action must be in the "focus" of both participants. Cohen argues that this mutual focus is only established if the novice interpreted the expert's previous statement as an indirect request (the "Intended Path"). If the novice were merely being "helpful" without recognizing the speaker's specific intent, the action would lack the shared salience required for a simple pronoun like "it" to be understood. Thus, the successful use of anaphora provides empirical support for the plan-based model of speech acts.

Conclusion and Future Directions

The paper demonstrates that speaker-reference provides a "smoking gun" for the plan-based interpretation of language. It validates the idea that listeners are not just reacting to information but are actively reconstructing the speaker's plan and adopting it as a mutual goal.

While this framework provides a robust theoretical outline, the author acknowledges that further research is needed to address the computational complexity of these inferences. Future work should also explore "mutual signaling," where the expert confirms they have accepted the novice's interpretation, creating a continuous loop of intent signaling that maintains conversational coherence.

Summary updated

BibTeX
@inproceedings{cohen1980signalling,
  author = {Philip R. Cohen},
  title = {{Signalling the Interpretation of Indicrect Speech Acts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1980},
}
Figure