From RAGS to RICHES: Exploiting the Potential of a Flexible Generation Architecture

Lynne J. Cahill, John Carroll, Roger Evans, Daniel S. Paiva, Richard Power, Donia Scott, Kees van DeemterACL 2001 · DOI

Introduction

Traditional Natural Language Generation (NLG) systems have long relied on a rigid, three-stage pipeline: content determination, sentence planning, and realization. While this linear model provides a clear separation of concerns, it often fails to handle the complex, non-linear dependencies found in natural language. For instance, the choice of a referring expression (like a pronoun) may depend on the specific lexical choices made elsewhere or the overall rhetorical structure of a document. This paper introduces OASYS and RICHES, a framework and reference implementation designed to move beyond these architectural constraints using an event-driven "whiteboard" approach.

The OASYS Framework

The core contribution of this work is OASYS (Objects and Arrows SYStem), a software library that implements a whiteboard architecture. Unlike traditional pipelines where data is passed sequentially from one module to the next, OASYS utilizes a central, active database.

In OASYS, data is represented as a network of typed "Objects" connected by "Arrows." The system is fundamentally event-driven: when a module publishes or updates data on the whiteboard, it triggers events. Other modules can "listen" for these specific events and react accordingly. This allows for a highly flexible interaction model where modules can interleave their tasks, collaborate on shared representations, and handle mixed dependencies that would be impossible in a strict pipeline.

The RICHES Implementation

To demonstrate the feasibility of the OASYS framework, the authors developed RICHES, a proof-of-concept NLG system. RICHES is composed of several small, functionally motivated modules, including Media Selection, Document Planner, and Sentence Finaliser.

The flexibility of the architecture allows RICHES to support various control regimes:

  • Pipelined: Modules can still operate sequentially by waiting for "finalise" events.
  • Interleaved: Modules like Lexical Choice and Referring Expression generation can pass control back and forth as they incrementally build a sentence.
  • Concurrent: Multiple modules can monitor the whiteboard and begin processing as soon as their required input data becomes available.

Evaluation on PharmWeb

The RICHES system was evaluated using instructional texts from PharmWeb, a website providing patient advice for medications such as nose and eye drops. This domain provided a rich set of real-world examples to test the system's ability to generate complex, formatted output that includes both text and graphical elements. The goal was to determine if the flexible architecture could produce results matching the style and complexity of human-authored medical instructions.

Results and Discussion

The implementation of RICHES serves as a successful feasibility study for the OASYS framework. The system demonstrated the ability to map abstract rhetorical plans to concrete, formatted HTML documents.

Example output from the RICHES system showing generated instructions for nose drops, integrating text and diagrams.

Qualitatively, the results show that the whiteboard architecture can gracefully handle non-linear dependencies. For example, a realization module can effectively "ask" a planner for additional information by publishing a request to the whiteboard, triggering the planner to provide the necessary data. This transparency and functional definition of modules make them significantly easier to reuse or swap across different generation tasks.

Conclusion and Future Work

While the current implementation of RICHES is relatively straightforward, it proves that a flexible, event-driven architecture is a viable alternative to the traditional NLG pipeline. Future work will focus on "revision-based" architectures, where modules can refine previous outputs based on new information. The authors also plan to test the framework's limits with more linguistically diverse and complex generation challenges, further exploiting the potential of the OASYS whiteboard model.

Summary updated

BibTeX
@inproceedings{cahill2001from,
  author = {Lynne J. Cahill and John Carroll and Roger Evans and Daniel S. Paiva and Richard Power and Donia Scott and Kees van Deemter},
  title = {{From RAGS to RICHES: Exploiting the Potential of a Flexible Generation Architecture}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2001},
  doi = {10.3115/1073012.1073027},
}
Figure