A Method for Development of Dialogue Managers for Natural Language Interfaces

Arne JönssonAAAI 1993

Introduction

In the early 1990s, the field of dialogue management was largely divided between two competing philosophies. The "strong AI" or plan-based approach attempted to model a user's deep psychological intentions and underlying goals to drive conversation. Conversely, linguistic-based models focused on the structural properties of discourse. In his 1993 AAAI paper, Arne Jönsson proposed a pragmatic alternative: a method for building efficient, task-oriented natural language interfaces (NLIs) using customized dialogue grammars derived from empirical data.

Jönsson’s work addresses the practical gap between theoretical general discourse models and the specific requirements of "simple service systems," such as database query tools. The core argument is that complex reasoning for phenomena like irony or deep intentionality is often unnecessary for task-specific applications. Instead, developers can achieve high "habitability"—the ease with which a user can interact with the system—by focusing on the sublanguage actually used in specific domains.

Methodology: The Sublanguage Approach

Jönsson’s methodology centers on customization through empirical observation, moving away from universal models toward domain-specific efficiency. The development process relies on several key pillars:

Wizard of Oz (WoZ) Simulations

Before a system is fully implemented, a human "wizard" simulates the interface's responses. This allows researchers to collect a corpus of authentic user interactions within the target domain. By observing how users naturally phrase requests and respond to system prompts, developers can identify the specific linguistic phenomena the dialogue manager must handle.

Hierarchical Dialogue Objects

The conversation is modeled as a hierarchy of objects:

  1. Dialogue: The entire interaction.
  2. Discourse Segments: Cohesive units of the conversation focused on a specific sub-task.
  3. Moves: Individual speech acts or utterances (e.g., a question or an update).

Dialogue Grammar and Focus Management

Rather than employing computationally expensive planning algorithms, the system utilizes a context-free grammar (CFG) to manage the interaction. The grammar identifies the "functional role" of an utterance—such as a Question (Q) or an Update (U)—and predicts the appropriate response, such as an Answer (A).

To handle linguistic nuances like ellipsis and pronouns (e.g., "How much is it?"), the system maintains a "dialogue tree" and a "scoreboard." Using simple heuristics, it tracks which objects and properties are currently in focus, allowing for rapid resolution of references without deep-history searches.

Empirical Evaluation: CARS and TRAVEL

The method was tested using a corpus of 30 dialogues involving 30 subjects (primarily computer novices). These interactions were spread across two distinct domains:

  • CARS: A database retrieval system for used cars interfacing with an INGRES database.
  • TRAVEL: A more complex system for retrieving information about charter trips to the Greek Archipelago, including scenarios for ordering trips.

Key Insights and Results

The "Overkill" Argument and Local Context

One of the paper’s most significant findings is that many complex human interaction phenomena accounted for in general discourse theories rarely occur in task-oriented NLIs. Jönsson argues that if a phenomenon does not appear in the WoZ corpus, the dialogue manager should not be burdened with the logic to handle it.

Furthermore, the study revealed that dialogue context is remarkably local. In the CARS study, 95% of user initiatives could be understood using only the information within the current discourse segment. Only 5% of cases required a search of the broader dialogue tree, suggesting that simple, local focus management is sufficient for most service-oriented tasks.

Efficiency and Habitability

The empirical results underscored the efficiency of the grammar-based approach:

  • Structural Simplicity: Between 60% (CARS) and 83% (TRAVEL) of interactions followed a straightforward "Initiative/Response" pattern.
  • Resolution: In the CARS application, 52% of user initiatives were fully specified and required no context, while 43% were resolved using local context.
  • User Experience: Despite the simplified underlying model, subjects reported feeling "very comfortable" with the system. This validates the idea that a dialogue manager does not need to be psycholinguistically plausible or mimic human behavior perfectly to be effective; it simply needs to be predictable and efficient.

Limitations and Conclusion

While highly effective for "simple service systems," Jönsson acknowledges potential limitations. As tasks become more procedural—such as the multi-step ordering phase in the TRAVEL system—the dialogue grammar must become more sophisticated. Additionally, the system’s performance can falter when user initiatives are overly vague, requiring more robust clarification sub-dialogues.

Ultimately, Jönsson’s work demonstrates that a robust, habitable NLI can be built from a relatively small number of dialogues. By "saturating" a context-free grammar with empirical data from WoZ simulations, developers can create dialogue managers that are both computationally lean and highly responsive to user needs.

Summary updated

BibTeX
@inproceedings{jonsson1993method,
  author = {Arne Jönsson},
  title = {{A Method for Development of Dialogue Managers for Natural Language Interfaces}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure