Catalogue of Artificial Intelligence Techniques
Rhetorical Structure Theory
Keywords: STRIPS, generation, natural language generation
Categories: Natural Language
Author(s): Alistair Knott
Rhetorical Structure Theory (RST) aims to provide an account of the phenomenon of text coherence. While many linguistic theories are concerned primarily with the structure of single sentences, RST explores the ways in which clauses and sentences can combine to form larger units of text, such as paragraphs, reports, or articles. The primary assumption underlying the theory is that a coherent text contains more information than is provided by its clauses and sentences individually. The additional information emerges from the way these units are related together. For instance, two clauses can be related as cause and effect, or as elements in a temporal sequence, or as premise and conclusion. RST proposes a set of some 23 rhetorical relations, such as volitional cause, contrast and justify, which it claims are sufficient to analyse `the vast majority' of coherent texts. These relations are defined functionally; that is to say, in terms of the effects the writer intends to achieve on the reader by the juxtaposition of the two related units. While relations can sometimes be signalled linguistically (for instance by sentence or clause connectives), they are defined without any reference to surface linguistic forms. The units linked by relations are known as text spans. Spans can be of any size, from single clauses upwards. Relations map onto texts in the first instance via structures called schemas, which group a small number of adjacent text spans in various different ways. The spans grouped by a schema are treated as a new composite span, which can then itself be related to other spans by other rhetorical relations. In this way a hierarchical pattern of schemas is built up, known as a rhetorical structure tree. A coherent text is characterised as one which can be described by such a tree. RST has recently found widespread application in the field of natural language generation. The hierarchical/compositional nature of rhetorical relations, as well as their definition in terms of the writer's goals, make them well-suited for use as planning operators, in the tradition of classical planning systems like STRIPS. Current debate about RST centres around the choice of the set of relations to be implemented and on how they should be defined, as well as on the question of what other mechanisms (if any) are necessary in addition to relations for a complete account of text coherence.
- Moore, J. and Paris, C., Planning Text for Advisory Dialogues: Capturing Intentional and Rhetorical Information
Computational Linguistics 19 (1993), 651--694.