Plans, Inference, and Indirect Speech Acts

James F. Allen, C. Raymond PerraultACL 1979 · DOI

In their seminal 1979 paper, James F. Allen and C. Raymond Perrault addressed a fundamental challenge in computational linguistics: the gap between the literal form of an utterance and its intended communicative function. This work moved beyond purely syntactic or semantic analysis, proposing instead that language understanding is a process of plan recognition. By treating conversation as a goal-oriented activity, the authors provided a formal framework for how a listener can infer a speaker's underlying intent, particularly in the case of indirect speech acts.

The Problem of Indirect Speech Acts

In human communication, there is rarely a one-to-one mapping between a sentence's structure and its purpose. This phenomenon is known as the Indirect Speech Act. For instance, when a speaker asks, "Can you pass the salt?", they are rarely inquiring about the listener's physical capabilities; they are making a request.

Traditional linguistic theories of the era struggled to account for these variations without resorting to exhaustive, ad-hoc rule sets. Allen and Perrault argued that the "meaning" of such utterances is not found solely in syntax or dictionary definitions, but in the speaker's intent and the listener's assumption of cooperation. The challenge for artificial intelligence, therefore, is to move beyond literal parsing toward "plan inference"—reconstructing the speaker's hidden goals based on the evidence provided by their words.

Speech Acts as Planning Operators

To model this process, the authors treated speech acts as operators within a planning system, similar to the STRIPS architecture used in early robotics. Their model relies on several key components:

  1. Two-Tiered Operators: The system distinguishes between surface-level acts (the literal form, such as a "Yes/No Question") and illocutionary-level acts (the actual intent, such as a "Request").
  2. Plan Recognition Rules: The listener uses a set of plausible inference rules to reconstruct the speaker's plan. For example:
  3. Action-Effect Inference: If a listener believes a speaker wants to perform action X, they assume the speaker wants to achieve the effects of X.
  4. Know-Positive Inference: If a listener believes a speaker wants to know if a proposition P is true, they assume the speaker may want to achieve P.
  5. Intentionality and Gricean Logic: Following Gricean maxims of communication, the model assumes that a speaker doesn't just want a goal achieved; they want the listener to recognize that they want that goal achieved. This mutual recognition is the basis of successful communication.

The Cooperative Agent and the Obstacle Heuristic

A central contribution of the paper is the idea of the "helpful agent." The model assumes that a listener identifies "obstacles" in the speaker's inferred plan—steps the speaker cannot complete alone—and provides a response specifically designed to overcome those obstacles.

This explains why humans often provide more information than was strictly requested. If a traveler asks a station clerk, "When does the Windsor train leave?", and the clerk responds, "3:15 at Gate 7," the clerk has identified that not knowing the gate is a future obstacle to the traveler's ultimate goal of boarding the train. By providing the gate number, the clerk proactively assists the traveler’s plan.

Implementation and Results

The authors validated their theoretical model through a computer program that simulated a clerk at a train station information booth. While the system operated within a constrained domain, it successfully handled a variety of complex dialogue behaviors, including:

  • Idiomatic Requests: Correcting interpreting "Can you pass the salt?"
  • Non-idiomatic Indirect Acts: Understanding nested intents, such as "John asked me to ask you to pass the salt."
  • Information-Seeking: Responding to "Do you know when the train leaves?" with the actual time rather than a literal "Yes."

The implementation demonstrated that a small number of independently motivated plan-inference rules could account for a vast range of human linguistic behavior, suggesting that the "intelligence" of a dialogue system should reside in its planning and inference layers rather than just its syntactic parser.

Limitations and Legacy

Despite its foundational status, the 1979 paper acknowledged several limitations. The authors noted the difficulty of identifying the "literal force" of a sentence from features like intonation or mood. Furthermore, the model focused on primary goals, whereas real-world sentences often perform multiple speech acts simultaneously. Finally, the plan-based approach requires the system to possess a deep, structured model of the world (e.g., how trains and stations function), which remains a significant scaling challenge for general-purpose AI.

Nevertheless, Allen and Perrault’s work remains a cornerstone of pragmatics in NLP, establishing the principle that to understand what someone says, a machine must first understand what they are trying to do.

Summary updated

BibTeX
@inproceedings{allen1979plans,
  author = {James F. Allen and C. Raymond Perrault},
  title = {{Plans, Inference, and Indirect Speech Acts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1979},
  doi = {10.3115/982163.982187},
}
Figure