A Rational Reconstruction of the PROTEUS Sentence Planner

Graeme RitchieACL 1984

TL;DR

This paper presents a "rational reconstruction" of the PROTEUS sentence planner, a system for generating English text describing games of "tic-tac-toe." The authors refine Davey's original program by structuring its components more clearly and formalizing its sentence planning process into three distinct phases: move annotation, sentence segmentation, and case-frame linking. This work aims to provide a cleaner, more general method for discourse generation, building upon existing ideas while offering a more robust framework for future research.

What Problem Does This Paper Address?

Generating natural-sounding, coherent text from structured data, such as game events, is a complex challenge in natural language generation (NLG). The original PROTEUS system, developed by Davey, demonstrated impressive fluency in describing "tic-tac-toe" games, but its internal workings, particularly the sentence planning component, were not fully detailed. This paper addresses the need for a clearer, more structured understanding of how PROTEUS achieved its discourse generation capabilities. The authors aim to provide a "rational reconstruction" – a cleaner, more general, and formally described version of the sentence planner, making its underlying principles more accessible and extensible for future NLG research.

Methods

The paper outlines a three-phase sentence planning process:

  1. Move Annotation: This phase involves analyzing the sequence of game moves and annotating each move with tactical information. Heuristic rules are used to determine which move the system would have made and which rule generated the actual move. This annotation includes information about generating heuristics, tactically equivalent alternatives, involved lines of play, and any "better" moves that could have been made.
  2. Sentence Segmentation: Here, the annotated moves are grouped into clusters, with each cluster forming a single sentence. This segmentation follows a set of ordered guidelines, prioritizing single moves, pairs of moves, and specific tactical situations (like mistakes, combined attacks/defenses, or thwarted attacks) to determine sentence boundaries. Each move is also tagged with its tactical relationship to preceding moves (e.g., "consequence," "thwart," "mistake").
  3. Case-Frame Linking: In this final phase, the annotated and segmented move descriptions are used to construct the internal sentence structure. Case-frame structures are computed and linked, representing different aspects of a move (e.g., "take" a square, "threaten" an opponent). These case-frames are then reformulated into "functional descriptions" suitable for functional unification grammar, incorporating links that determine syntactic structure, lexical choices (binders/connectives), and grammatical features like modality and aspect.

Datasets

The paper uses the game of "tic-tac-toe" as its domain. The "dataset" consists of sequences of game moves, represented by square numbers (1-9), along with the name of the player making the move. The paper does not introduce new datasets but rather uses the game's history as input for the described process.

Key Insights and Ideas

  • Decomposition of Sentence Planning: The paper's core contribution is the clear, three-phase decomposition of the sentence planning process (annotation, segmentation, linking). This structured approach makes the system's logic more understandable and potentially easier to adapt.
  • "Rational Reconstruction" as a Methodology: The authors propose "rational reconstruction" as a valuable method for understanding and refining complex AI systems. It involves creating a cleaner, more general, and formally described version of an existing system's core ideas, rather than a direct replication.
  • Tactical Information as the Basis for Discourse: The fluency of the generated text is shown to derive significantly from the detailed tactical analysis of game moves, rather than solely from syntactic processing.
  • Heuristics for Text Generation: The system employs heuristic rules not only for playing the game but also for guiding the text generation process, demonstrating a tight coupling between game strategy and discourse planning.
  • Flexibility in Textual Representation: The case-frame linking phase highlights the system's ability to generate multiple linguistic descriptions for a single move, reflecting the non-deterministic nature of natural language expression.

Key Results

The paper describes the process and structure of the rational reconstruction of the PROTEUS sentence planner. It does not present quantitative results in terms of accuracy or fluency metrics. The key outcome is the detailed explanation of the refined three-phase sentence planning mechanism and its underlying principles, which are presented as a more general and robust framework for discourse generation. The paper implies that this reconstruction maintains the fluency of the original Proteus system while offering a cleaner, more formal architecture.

Limitations and Future Work

  • Potential Loss of Fluency: The authors acknowledge that the process of cleaning up and generalizing Proteus's ad hoc measures might lead to minor losses in fluency, where specific effects were gained through less systematic means in the original program.
  • Domain Specificity: The "tic-tac-toe" domain has a convenient canonical form for representing events, which simplifies the initial move representation. The authors suggest that for more general domains, the planning process might need to start at a later stage, assuming moves already have some form of "descriptions."
  • Future Work: The paper explicitly states that the aim of the current project is to use this sentence planner as a "back-end" for experimenting with functional unification grammar. This suggests future work will involve integrating the generated functional descriptions with a more sophisticated syntactic generation component. The authors also hint at the need for more abstract models for domains where event representations are not as straightforward as in "tic-tac-toe."

Summary updated

BibTeX
@inproceedings{ritchie1984rational,
  author = {Graeme Ritchie},
  title = {{A Rational Reconstruction of the PROTEUS Sentence Planner}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}

Loading related content...

Figure