Adapting MUMBLE: Experience with Natural Language Generation

Robert RubinoffAAAI 1986

Introduction

In the mid-1980s, Natural Language Generation (NLG) systems faced a significant performance bottleneck. The TEXT system, designed to answer database queries about naval vessels, utilized a two-stage process: a strategic component to decide content and a tactical component to determine linguistic form. While the strategic component was effective, the original tactical component—relying on a functional grammar—was computationally prohibitive, often requiring five minutes to generate a single sentence.

This paper describes the successful replacement of that bottleneck with MUMBLE, a deterministic, interpreter-based generation framework. The transition resulted in a dramatic 60-fold speedup, enabling the production of 57 complex sentences in the same time it previously took to generate one.

The MUMBLE Framework Adaptation

Adapting MUMBLE for the TEXT system required "impedance matching" between TEXT’s idiosyncratic rhetorical propositions and MUMBLE’s rigid, top-down deterministic model. This involved modifications across three core pillars:

1. The Interpreter and Morphology

MUMBLE’s interpreter was updated to handle generic concepts. Unlike previous applications that dealt with specific objects (e.g., "this house"), the TEXT system generated generic statements (e.g., "Ships are..."). The interpreter had to be modified to determine plurality by inspecting the surface structure tree during realization, ensuring that generic concepts were correctly inflected based on their syntactic role.

Figure 1: A clause structure in MUMBLE showing the mapping of the 'SHIP' concept to a subject position and the 'WATER-VEHICLE' concept to an object position within a verb phrase.

2. Procedural Grammar

The grammar in MUMBLE functions as a decision engine rather than a static set of rules. It uses procedures to enforce local constraints and maintain "grammar variables" (stacks) to track discourse focus. For the TEXT system, new syntactic structures such as appositives and relative clauses were implemented to support complex descriptions.

3. The Dictionary and Realization Classes

The most labor-intensive task was building the dictionary. The author developed "realization classes" (r-classes) to map TEXT’s message formalism into English. This process revealed that linguistic generalizations are rarely portable; templates from previous MUMBLE projects were largely inapplicable to the naval domain of the TEXT system.

Figure 3: A noun phrase (np) structure illustrating how data attributes like (TRAVEL-MODE SURFACE) are realized as modifiers within the phrase 'a water-going vehicle'.

Key Insights

  • Determinism as a Constraint: MUMBLE’s commitment to no-backtracking forces a top-down organization. While global structures can influence local decisions, local constraints cannot easily alter the global structure once the interpreter has moved past a certain point.
  • The Myth of General Realization: The study found that message formalisms are highly idiosyncratic. Dictionaries must be custom-built for new domains, as "general-purpose" realization classes often fail to capture the specific nuances required by different strategic planners.

Results and Performance

The integration was highly successful. The MUMBLE-based tactical component reproduced the complex rhetorical structures of the original TEXT system—including comparisons and attribute-based identification—with significantly improved efficiency. The new system produced all 57 sentences from the test set in approximately 5 minutes, compared to the 5 minutes per sentence required by the previous functional grammar.

Conclusion

The adaptation of MUMBLE demonstrates that deterministic, incremental generation is not only viable but highly efficient for complex, rhetorically-driven text planning. While the lack of portable realization classes remains a challenge for scaling NLG systems, the performance gains achieved through a procedural, interpreter-based approach are substantial.

Summary updated

BibTeX
@inproceedings{rubinoff1986adapting,
  author = {Robert Rubinoff},
  title = {{Adapting MUMBLE: Experience with Natural Language Generation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure