Discourse Structures for Text Generation

William C. MannCOLING 1984

TL;DR

This paper introduces Rhetorical Structure Theory (RST), a novel framework for understanding and generating organized text. RST posits that text is composed of nuclear and satellite parts, connected by specific rhetorical relations, offering a more comprehensive and functional approach to text organization than previous methods. This theory aims to bridge the gap between descriptive analysis of existing texts and the constructive generation of new ones.

What Problem Does This Paper Address?

Generating coherent and effective text is a significant challenge in computational linguistics. While text generation systems exist, they often lack a robust theory of text organization, leading to texts that may be grammatically correct but poorly structured and unconvincing. Existing approaches often focus on sentence-level generation or simplistic overall structures, failing to capture the nuanced ways writers organize their ideas to achieve specific communicative goals. This paper addresses the need for a comprehensive, functional, and formalizable theory of text organization that can guide the creation of more sophisticated text generation systems.

Methods

The paper introduces Rhetorical Structure Theory (RST), a descriptive theory of text organization. RST breaks down text into hierarchical structures composed of "spans," which can be either text spans (actual words) or conceptual spans (propositions). Each span is analyzed in terms of a nucleus (the most essential part) and satellites (supporting parts). These parts are connected by rhetorical relations (e.g., Motivation, Evidence, Elaboration, Concession) that define how the satellite contributes to the nucleus.

Key aspects of RST include:

  • Schemas: Predefined patterns of nucleus-satellite relationships that represent common rhetorical structures (e.g., Request, Evidence, Thesis/Antithesis). The paper illustrates how these schemas are applied to analyze a natural text.
  • Recursiveness: Schemas can be applied at various levels of text structure, from the entire document down to individual clauses, allowing for analysis of text at multiple scales.
  • Functionality: RST aims to be functional, explaining how text achieves its effects for the writer by defining the purpose of each relation.
  • Descriptive to Constructive: The paper outlines a "descriptive" version of RST for analyzing existing texts and proposes that this can be extended into a "constructive" version for text generation.

Datasets

The paper does not introduce or use specific datasets for training or evaluation in the traditional machine learning sense. Instead, it uses a single natural text, a newsletter article from California Common Cause, as an illustrative example to demonstrate the application of Rhetorical Structure Theory. The authors mention that they have applied the theory to "approximately 100 short natural texts, including administrative memos, advertisements, personal letters, newspaper articles, and magazine articles" to identify patterns and exceptions, but these are not presented as formal datasets.

Key Insights and Ideas

  • Text as Organized Argument: The core idea is that well-written text isn't just a sequence of sentences but a deliberate organization of parts to achieve specific communicative goals. RST provides a framework to analyze and replicate this organization.
  • Nucleus-Satellite Structure: The concept of a nuclear element and its supporting satellites is fundamental. This highlights that not all parts of a text are equally important, and their roles are defined by their relationship to the nucleus.
  • Rhetorical Relations as Communicative Tools: The defined relations (Motivation, Evidence, etc.) are not just grammatical connectors but represent specific communicative functions that writers employ to persuade, inform, or convince their audience.
  • Functionality for Text Generation: By understanding why a writer uses a particular structure (i.e., its function), RST provides a basis for generating text that serves specific writerly purposes, moving beyond mere grammatical correctness.
  • Scalability through Recursion: The recursive nature of RST allows it to analyze and describe text structures at all levels, from the entire document down to individual clauses, making it a versatile tool.

Key Results

The paper's primary contribution is the introduction and demonstration of Rhetorical Structure Theory (RST) as a powerful framework for text organization. The "results" are not quantitative performance metrics but rather the successful application of RST to analyze a natural text, revealing its underlying rhetorical structure.

  • Demonstrated Applicability: The paper successfully applies RST to a real-world text, breaking it down into a hierarchical structure of nuclei, satellites, and rhetorical relations, illustrating the theory's descriptive power.
  • Comprehensive Analysis: RST is shown to be more comprehensive than previous text organization approaches by accounting for various levels of structure, the functional roles of text segments, and the relationships between them.
  • Foundation for Generation: The paper argues that the descriptive version of RST is directly extendable to a constructive theory for text generation, providing a theoretical basis for building more intelligent text generators.

Limitations and Future Work

  • Informal Nature: The authors acknowledge that descriptive RST is currently informal, requiring human judgment to apply relations and identify structures. Formalizing this for automated analysis or generation is a significant challenge.
  • Complexity of Formalization: Reconciling RST with existing formal theories of speech acts, semantics, grammar, and human inference is a complex undertaking that has not yet been achieved.
  • Constructive RST Development: The paper outlines the expectation that RST can be augmented for text generation but does not present a fully developed constructive theory. Future work will focus on mapping RST onto AI goal-pursuit methods to create a programmable framework for generation.
  • Implicit Communication: While RST accounts for implicit communication through relations, a deeper exploration of this aspect is noted as outside the scope of the current paper.

Summary updated

BibTeX
@inproceedings{mann1984discourse,
  author = {William C. Mann},
  title = {{Discourse Structures for Text Generation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1984},
}
Figure