POPEL-HOW: A Distributed Parallel Model for Incremental Natural Language Production with Feedback

Wolfgang Finkler, Günter Neumann

Introduction

In traditional natural language generation (NLG), the process is typically sequential: a "what-to-say" component (the planner) completes its task before a "how-to-say" component (the realizer) begins. However, human speech is inherently incremental and bidirectional. We often begin a sentence before the full thought is finalized, and grammatical constraints encountered mid-sentence can force us to adjust our conceptual planning.

POPEL-HOW, developed in the late 1980s, was a pioneering effort to bridge this gap. It introduces a distributed, parallel architecture that allows for incremental generation with continuous feedback between the conceptual and linguistic levels.

System Architecture

The core of POPEL-HOW is a distributed parallel system where linguistic representations are treated as autonomous "active objects." The system processes information through four distinct levels:

  1. Conceptual Knowledge Base (CKB): Represents raw ideas or concepts.
  2. Functional Semantic Structure (FSS): Maps concepts to semantic roles (e.g., agent, object).
  3. Dependency-Based Structures (DBS): Establishes syntactic relationships.
  4. Inflected Linearized Structures (ILS): Handles word ordering and morphology using the MORPHIX package.
Figure 1: The overall architecture of the POPEL-HOW system, illustrating the flow of information and feedback loops between the four processing levels.

Unlike centralized models, each segment of a sentence—such as a noun or verb—is an autonomous object. These objects attempt to map themselves to the subsequent level as soon as possible. If a lower-level object (e.g., at the DBS level) identifies a missing piece of information required by grammatical rules, it sends a request back up to the conceptual level. This bidirectional feedback allows linguistic constraints to actively influence the ongoing planning process.

Figure 2: Notation for active objects and communication links between levels, showing the hierarchical "father-son" relationships.

Key Concepts and Insights

Linguistic Constraints as Planning Inputs

In POPEL-HOW, grammar is not merely a final filter but an active participant in planning. If a chosen syntactic structure requires a specific case or preposition not yet planned, the system compels the planner to provide that detail immediately.

Monotonic vs. Non-monotonic Incrementation

The authors distinguish between monotonic growth (adding words to a sentence) and non-monotonic changes. The latter occurs when new information necessitates repairing or replacing a partially built structure, a common occurrence in spontaneous human speech.

Object-Oriented Linguistics

By treating parts of speech as concurrent objects with independent life cycles, the model achieves high parallelism. Different parts of a single sentence can be processed at different levels simultaneously, facilitating "premature" verbalization—outputting the start of a sentence while the end is still being planned.

Evaluation and Results

The system was evaluated within the XTRA dialog system, a natural language interface for expert systems. While standardized benchmarks were not common at the time, the authors demonstrated the system's efficacy through complex interaction scenarios involving tax forms and travel planning.

Key results included:

  • Incremental Speedup: Successful demonstration of starting output before conceptual planning is complete.
  • Output Flexibility: The order of input concepts directly influences lexical choice. For instance, knowing the "instrument" of an action early might lead the system to choose a specific verb (e.g., "flies") over a generic one (e.g., "goes by plane").
  • Effective Feedback: The model proved that the "how-to-say" component could successfully guide the "what-to-say" component to ensure linguistic soundness.

Limitations and Future Work

Despite its innovations, POPEL-HOW faced challenges. Distributed word choice remains difficult, as selecting the optimal word often requires a global context that independent active objects lack. Additionally, managing final word order in a parallel model can be complex if the original conceptual sequence is lost. Finally, while the system handles repairs, the authors noted that more sophisticated self-correction mechanisms are needed to achieve truly human-like output beyond simple repetition of corrected constituents.

Summary updated

BibTeX
@inproceedings{finkler1989popel,
  author = {Wolfgang Finkler and Günter Neumann},
  title = {{POPEL-HOW: A Distributed Parallel Model for Incremental Natural Language Production with Feedback}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1989},
}
Figure