Planning Coherent Multisentential Text

Eduard H. HovyACL 1988 · OpenReview

Abstract

Though most text generators are capable of simply stringing together more than one sentence, they cannot determine which order will ensure a coherent paragraph. A paragraph is coherent when the information in successive sentences follows some pattern of inference or of knowledge with which the hearer is familiar. To signal such inferences, speakers usually use relations that link successive sentences in fixed ways. A set of 20 relations that span most of what people usually say in English is proposed in the Rhetorical Structure Theory (RST) of Mann and Thompson. This paper describes the formalization of these relations and their use in a prototype text planner that structures input elements into coherent paragraphs.

Introduction: The Problem of Coherence

In natural language generation (NLG), producing individual, grammatically correct sentences is a well-understood task. However, organizing these sentences into a coherent paragraph presents a significant challenge. Without a structured plan, a system given multiple facts to communicate might output them in a disjointed or random order.

Eduard Hovy identifies this as the "Problem of Coherence." For a paragraph with just seven input facts, there are $7!$ (5,040) possible permutations, most of which are nonsensical. The challenge lies in defining the logical "glue"—such as "therefore," "in particular," or "finally"—that makes a sequence of sentences feel logical to a human reader.

Methods: Formalizing Rhetorical Structure Theory

The core innovation of this research is the transformation of Rhetorical Structure Theory (RST) from a descriptive linguistic tool into a generative planning mechanism.

1. Relations as Planning Operators

In RST, text is composed of a Nucleus (the primary information) and a Satellite (supporting information). Hovy formalizes these relations (e.g., Purpose, Elaboration, Sequence) as AI planning operators. Each plan contains:

  • Nucleus/Satellite Constraints: Requirements the data must meet before a relation can be applied.
  • Intended Effects: The communicative goal (e.g., ensuring the hearer understands the purpose of an action).

2. Top-Down Hierarchical Planning

The system employs a top-down planner to organize text. It begins with a high-level communicative goal and recursively searches for RST relations that can satisfy that goal, breaking it down into subgoals until it reaches the raw data elements.

The planning process architecture, showing the recursive loop of expanding buds and growing the tree based on RST relations.

Key Insights and Structural Representation

By treating communication as a goal-oriented task, the system ensures that every sentence serves a specific purpose in changing the hearer's knowledge state. The recursive nature of RST allows for complex, multi-layered arguments where a satellite of one relation can serve as the nucleus for another.

A hierarchical RST tree showing how input records (a-g) are organized into a coherent structure using nested relations like Sequence, Elaboration, and Purpose.

Using rhetorical constraints drastically reduces the search space. For a 7-element paragraph, the system can narrow down thousands of permutations to just a few viable candidates. This separation of concerns allows a "Text Structurer" to handle the logic and order, while a separate generator (like the Penman system) manages grammar and syntax.

Results and Limitations

The prototype system demonstrated a qualitative leap in text quality across different domains, such as the Programming Enhancement Advisor (PEA) and naval logistics reports. Instead of disjointed lists, the system produced structured paragraphs with appropriate connectives.

However, limitations remain. Only a portion of the proposed RST relations have been fully formalized. Additionally, the system currently lacks "editorial" capabilities; it attempts to include every piece of input data provided, even if some information is redundant. Future work involves improving "sentence scoping"—deciding when to combine facts into a single sentence—and better integrating aggregation techniques to avoid repetitive text structures.

Summary updated

Abstract

Though most text generators are capable of simply stringing together more than one sentence, they cannot determine which order will ensure a coherent paragraph. A paragraph is coherent when the information in successive sentences follows some pattern of inference or of knowledge with which the hearer is familiar. To signal such inferences, speakers usually use relations that link successive sentences in fixed ways. A set of 20 relations that span most of what people usually say in English is proposed in the Rhetorical Structure Theory of Mann and Thompson. This paper describes the formalization of these relations and their use in a prototype text planner that structures input elements into coherent paragraphs.
BibTeX
@inproceedings{hovy1988planning,
  author = {Eduard H. Hovy},
  title = {{Planning Coherent Multisentential Text}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1988},
  url = {https://openreview.net/forum?id=SJ-KvieuZB},
}
Figure