Kalos - A System for Natural Language Generation with Revision

Ben E. Cline, Jane Terry NutterAAAI 1994

Introduction

Traditional Natural Language Generation (NLG) systems typically follow a "one-and-done" approach, attempting to produce final text in a single pass. This often leads to fragile architectures and repetitive, unpolished output. Presented at AAAI-94, Kalos introduces a paradigm shift by adopting a human-like "draft and revise" workflow. By utilizing iterative cycles to polish technical documentation, Kalos simplifies the generation process while significantly improving the quality of the final prose.

The Challenge of Monolithic Generation

Generating high-quality, multi-paragraph text is inherently difficult due to two primary factors:

  1. Monolithic Complexity: Systems that attempt to handle content selection, logical ordering, and grammatical correctness simultaneously are often fragile and difficult to scale.
  2. Lack of Polishing: Linear generation lacks a "big picture" view. Once a sentence is generated, it is rarely revisited, resulting in awkward repetitions and redundant facts.

Kalos addresses these issues by separating the initial generation of a rough draft from subsequent revision stages, reducing system complexity and enhancing readability.

System Architecture

The Kalos architecture is centered around a Uniform Knowledge Base, implemented using SNePS-2.1 semantic networks. This shared knowledge allows the system to track the underlying intent and data behind every sentence throughout the pipeline.

The Kalos Architecture, showing the interaction between the Generator, Revisor, and Knowledge Base.

The pipeline consists of three primary stages:

  • Deep Generator: Selects and orders raw concepts from the knowledge base based on a specific goal, such as describing a microprocessor.
  • Surface Generator: Translates these concepts into a series of simple, individual sentences.
  • Revision Module: This module is divided into two specialized components:
  • Conceptual Revisor: Focuses on content. It identifies and removes taxonomic redundancies (e.g., deleting a sentence about bus width if the chip's bit-size is already established) and reorders facts for logical flow.
  • Stylistic Revisor: Focuses on form. It performs "text polishing" through sentence compounding, anaphora (pronoun usage), and ensuring thematic progression—maintaining a consistent subject to improve narrative flow.

Key Insights: Task Decomposition and Interaction

A core insight of Kalos is that task decomposition allows for simpler initial generators. Because the revisor is designed to catch redundancies and awkward phrasing, the generators do not need to be "perfect" on the first pass. Furthermore, the architecture allows for surface-to-deep interaction; stylistic preferences can trigger the system to backtrack and reorganize the underlying conceptual structure.

Evaluation and Results

The system was tested using a domain-specific knowledge base for the Motorola M68000 microprocessor. The goal was to generate introductory technical documentation.

The transformation of the text through the pipeline was significant:

  • Initial Draft: Produced 49 simple, repetitive sentences (e.g., "The M68000 is a microprocessor. The M68000 has a data bus.").
  • Conceptual Revision: Reduced the count to 29 sentences by stripping redundancies.
  • Final Output: Resulted in a polished 14-sentence paragraph using sophisticated linguistic structures.
Sample final output from Kalos after stylistic revision, demonstrating cohesive technical prose.

Limitations and Future Directions

While effective, the current implementation of Kalos has limitations. Revisions are currently cumulative and irreversible; allowing the system to "undo" a revision could potentially optimize output further. Additionally, the pipeline is fixed, with conceptual revision always preceding stylistic revision. Future work may explore interleaving these stages or adding an ambiguity-checking module to ensure the generated text is not only polished but also clear and unambiguous.

Summary updated

BibTeX
@inproceedings{cline1994kalos,
  author = {Ben E. Cline and Jane Terry Nutter},
  title = {{Kalos - A System for Natural Language Generation with Revision}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure