A Framework for Customizable Generation of Hypertext Presentations

Benoit Lavoie, Owen RambowACL 1998

TL;DR

This paper introduces PRESENTOR, a framework for building customizable hypertext presentation generators. It addresses the challenge of rapidly developing and tailoring information presentations by providing declarative languages for defining content, structure, and linguistic realization, making it easier for non-programmers to create domain-specific generators.

What Problem Does This Paper Address?

Generating effective presentations of complex information is a significant challenge. As the volume of data grows, there's a pressing need for systems that can quickly create and adapt presentations to specific user needs or standards. Traditionally, building such systems requires extensive domain-specific knowledge and programming for each generation subtask (macro-planning, micro-planning, realization, formatting). This paper tackles the problem of simplifying and accelerating this process by offering a unified, declarative framework that separates domain knowledge from the generation engine, enabling easier customization and reuse.

Methods

PRESENTOR employs a pipeline architecture for presentation generation, comprising several modules: a Generator Interface, Macro-Planner, Micro-Planner, Realizer, and Formatter. The core innovation lies in its use of declarative knowledge bases that are interpreted at runtime, allowing customization without altering the core modules. These knowledge bases include:

  • Environment Variables: For system configuration.
  • Exemplars: Schema-like structures that specify presentation content and structure at various abstraction levels (rhetorical, conceptual, syntactic, surface form). These are designed for both deep and shallow generation.
  • Rhetorical Dictionary: Maps rhetorical relations to linguistic expressions.
  • Conceptual Dictionary: Translates language-independent conceptual structures into language-specific syntactic structures.
  • Linguistic Grammar: Transforms deep syntactic structures into surface word forms and punctuation.
  • Lexicon: Stores linguistic attributes of words.
  • Format Style: (Planned) For defining presentation formatting.

The framework emphasizes a common presentation structure that facilitates module integration and allows for the reuse of domain knowledge across different applications.

Datasets

Not discussed in the paper. The paper focuses on the framework's architecture and knowledge representation, not on specific datasets used for training or evaluation.

Key Insights and Ideas

  • Declarative Customization: The most significant insight is the power of declarative languages for defining presentation logic. This allows non-programmers to customize generators by modifying knowledge bases rather than code, drastically reducing development time and effort.
  • Unified Exemplar Structure: The concept of "exemplars" that integrate conceptual, syntactic, and surface form specifications is a key idea. This unified approach supports both deep and shallow generation and simplifies the representation of presentation content.
  • Pipeline Architecture for Practical Applications: While acknowledging debates about pipeline architectures in NLG, the authors argue for its practical effectiveness in real-world applications, highlighting its ability to minimize interdependencies between modules.
  • Separation of Concerns: The clear distinction between the generator engine and domain-specific knowledge bases is crucial for reusability and maintainability.

Key Results

The paper reports that PRESENTOR has been successfully used in several application domains, including weather forecasting, object model description, and system requirements summarization. The framework's portability (implemented in Java and C++) and its declarative nature are highlighted as key strengths contributing to its success. Specific quantitative results are not provided, but the qualitative success in diverse domains demonstrates the framework's efficacy.

Limitations and Future Work

The authors acknowledge a significant limitation: the configurability of PRESENTOR at the micro-planning level is restricted to lexicalization and rhetorical relation realization. Pronominalization rules are hard-coded heuristics, which can be problematic as they are often domain-specific and may require code changes when porting to new domains.

Future work includes adding declarative specifications for formatting styles. The authors also mention the development of a complementary framework, EXEMPLARS, which offers programmatic control, and express the intention to integrate features from both approaches.

Summary updated

BibTeX
@inproceedings{lavoie1998framework,
  author = {Benoit Lavoie and Owen Rambow},
  title = {{A Framework for Customizable Generation of Hypertext Presentations}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1998},
}
Figure