Design for Dialogue Comprehension

William C. MannACL 1979 · DOI

Introduction

In the late 1970s, the field of natural language processing (NLP) faced a fundamental challenge: ambiguity. As computational systems became more sophisticated, they inevitably generated multiple valid interpretations for a single human utterance. In his 1979 paper, Design for Dialogue Comprehension, William C. Mann introduced the Dialogue Comprehension System (DCS). This framework moved beyond simple syntactic parsing to treat conversation as a series of goal-oriented "Dialogue Games." Central to this system was Preeminence Scheduling, a novel search strategy designed to identify the most likely interpretation of an utterance with high computational efficiency.

The Problem: The Ambiguity Trap

Mann identified a critical bottleneck in contemporary NLP systems. When a machine processes human speech, it often encounters several "correct" interpretations. At the time, systems typically handled this in one of two ways:

  1. First-find: The system accepts the first interpretation that meets a basic logical threshold. While fast, this method frequently leads to errors because the first valid interpretation is not necessarily the intended one.
  2. Bounded search: The system generates a wide range of interpretations and assigns each a score, eventually picking the highest. This is computationally expensive and often forces the system to make arbitrary distinctions between two nearly identical high-scoring options.

Mann argued that true comprehension requires understanding the underlying goals of the speakers and a decision-making process that only commits to an interpretation when it is demonstrably superior to its rivals.

The DCS Framework

The Dialogue Comprehension System is built upon three core architectural pillars:

1. Dialogue Game Theory

Rather than analyzing sentences in isolation, DCS views dialogue as a structured sequence of "turns" where participants raise and resolve specific issues. These patterns are formalized as "Dialogue Games," such as Permission Seeking or Information Seeking. In this model, every utterance is viewed as a formal speech act—either a "bid" to start a game or a move to "terminate" one. By identifying the active game, the system gains a context-bound set of expectations that narrow the search space for meaning.

2. Preeminence Scheduling

This is the system's most significant algorithmic innovation. Instead of calculating absolute quality scores for every possible interpretation, the system seeks a "threshold of demonstrable preferability." To optimize computation, the system employs a counter-intuitive strategy: it focuses its energy on refuting the current top-ranked interpretation and supporting the lower-ranked ones. If the leading candidate remains significantly better than its competitors despite these efforts, it is declared "preeminent." This "efficiency of refutation" ensures the system does not waste resources perfecting the score of a clear winner.

3. Hearsay-3 Architecture

DCS utilizes an evolution of the Hearsay-II "blackboard" model. This architecture features a global data structure (the blackboard) where specialized "knowledge sources" (KS) monitor for information relevant to their expertise. DCS specifically introduces a Scheduling Blackboard (SBB) to manage the logic of Preeminence Scheduling, allowing the system to dynamically prioritize which interpretations to investigate or challenge.

Methodology and Data

Mann emphasized the necessity of using "actual human dialogue" to avoid the biases inherent in laboratory-constructed examples. The DCS was designed and tested against high-stakes, real-world communication, most notably:

  • Apollo Astronaut Transcripts: Communications between Apollo astronauts and ground control provided a rich dataset of interpersonal, goal-driven dialogue.
  • Naturalistic Human Dialogues: Various transcripts from non-laboratory settings were used to capture the nuances of ongoing, unconstrained communication.

Key Insights

  • Comprehension as Goal-Seeking: The paper posits that syntax is secondary to intent. If a system cannot recognize the "game" being played, it cannot truly understand the turn, regardless of how perfectly it parses the grammar.
  • Thresholds Over Rankings: Mann argued against the practice of choosing a single interpretation when two are nearly equal. DCS is designed to remain uncommitted until a clear threshold of difference is established, preventing the system from making brittle, arbitrary choices.
  • Hierarchical Processing: The DCS blackboard is organized into six distinct levels: Text, Word-senses, Syntax, Propositions, Speech-acts, and Goals. This hierarchy allows the system to build a coherent goal structure for every utterance, from the phonetic level up to the speaker's ultimate objective.

Conclusion and Limitations

At the time of its presentation, the Dialogue Comprehension System was an ongoing implementation project. While it successfully integrated Speech Act theory with a unified computational model, it lacked quantitative benchmarks regarding accuracy and processing speed. Furthermore, the system's success remained dependent on the quality of its "knowledge sources" to generate initial interpretations. Nevertheless, Mann’s work on Preeminence Scheduling and Dialogue Games remains a foundational contribution to the study of goal-oriented communication and efficient search strategies in artificial intelligence.

Summary updated

BibTeX
@inproceedings{mann1979design,
  author = {William C. Mann},
  title = {{Design for Dialogue Comprehension}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1979},
  doi = {10.3115/982163.982186},
}
Figure