Repairing Miscommunication: Relaxation in Reference

Bradley A. GoodmanAAAI 1983

Introduction

In the early development of natural language processing (NLP) systems, a significant bottleneck was the rigid expectation of perfect input. Most systems operated under the assumption that users would provide precise, accurate descriptions of objects. However, human communication is inherently imprecise. Bradley Goodman’s 1983 research addresses this fundamental flaw by introducing a "relaxation" mechanism. This approach allows AI to repair miscommunications by strategically ignoring or modifying parts of an imperfect description to identify the intended referent in the physical world.

The Problem of Imperfect Description

Early natural language interfaces were notoriously fragile. If a user requested a "blue tube" when the actual object was violet, the system would typically fail, reporting that no such object existed. This "literal mismatch" problem ignored the reality that speakers and listeners often have differing beliefs, perceptual biases, and contextual understandings.

Goodman argues that reference failures—where a description does not perfectly match any available object—are a standard feature of human dialogue, not an error to be discarded. The challenge is to build a "tolerant" system capable of resolving these failures by understanding how and why a description might be slightly inaccurate, rather than simply failing on a literal mismatch.

Methodology: Reference Repair via Relaxation

The proposed framework centers on Reference Repair via Relaxation. When a literal search fails to find a match in the knowledge base, the system begins to "relax" (ignore or modify) specific components of the user's description until a likely candidate is identified.

The system utilizes a KL-One classifier, a prominent knowledge representation system of the era, to navigate a taxonomy of objects. This allows the system to understand hierarchical relationships (e.g., a "nozzle" is a type of "part").

Hierarchical Relaxation Rules

Relaxation is not a random process. The system employs two primary knowledge sources to prioritize which features to modify:

  1. Linguistic Knowledge: The system recognizes that certain linguistic structures are more "relaxable" than others. For instance, information contained in a relative clause (e.g., "the tube that has two outlets") is often treated as confirmatory or secondary to the head noun.
  2. Perceptual Knowledge: The system distinguishes between "active" features (those requiring high attention, like size or position) and "less active" features (like color). It might relax a color value from "red" to "pink" or "violet" before questioning the fundamental shape or category of the object.

Relaxation Strategies

The system can employ three main strategies:

  • Feature Dropping: Completely ignoring a descriptor (e.g., ignoring "blue" in "blue tube").
  • Value Weakening: Broadening a specific value (e.g., changing "large" to "medium-to-large").
  • Substitution: Replacing a value with a related or perceptually similar one.

The Water Pump Domain

The research was grounded in actual human protocols—transcripts of an expert instructing an apprentice to assemble a toy water pump. These dialogues captured natural human errors, such as mislabeling colors or confusing similar-looking parts.

Components of the toy water pump used in the study's protocols.

As shown in Figure 1, the water pump consists of several distinct parts like the "Main Tube," "Nozzle," "Spout," and "Air Chamber." In the protocols, subjects frequently made mistakes, such as calling the violet stand "blue" or confusing the "nozzle" with the "spout." These real-world errors provided the "ground truth" for the types of miscommunications the system needed to handle.

Key Insights: Miscommunication as a Feature

A central insight of Goodman’s work is that miscommunication is a feature of robust interaction. By shifting the focus from "perfect parsing" to "robust repair," the paper acknowledges that imperfect descriptions are standard.

The system also utilizes Contextual Focus and part-whole reasoning. If a speaker describes a "green plunger" but the plunger actually consists of a red handle and a blue cap, the system can infer that the speaker might be attributing a feature of a sub-part to the whole, or vice versa, to resolve the reference.

Results and Discussion

While the paper is primarily an architectural contribution, it successfully demonstrated a computational scheme capable of:

  1. Detecting reference failure.
  2. Proposing "near-match" candidates.
  3. Using linguistic and perceptual heuristics to select the most likely intended referent.

For example, the system could successfully navigate a request for a "long blue tube" when the object was actually a violet stand by recognizing that structural features (like the "two outlets" on the tube) are more reliable indicators than color.

Limitations and Future Work

Despite its success, the system faced several limitations. It was highly domain-specific, relying on a manually constructed KL-One knowledge base for the water pump task. Furthermore, the decision of when to relax versus when to ask for clarification involves complex pragmatic rules that were not fully explored. Goodman suggested that future work should move toward a more general theory of "extensional descriptions"—how people successfully pick out objects despite linguistic imperfections.

Summary updated

BibTeX
@inproceedings{goodman1983repairing,
  author = {Bradley A. Goodman},
  title = {{Repairing Miscommunication: Relaxation in Reference}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}

Loading related content...

Figure