Recent research on text generation has shown that there is a need for stronger linguistic theories that tell in detail how texts communicate. The prevailing theories are very difficult to compare, and it is also very difficult to see how they might be combined into stronger theories. To make comparison and combination a bit more approachable, we (Mann and Thompson) have created a book which is designed to encourage comparison.

A dozen different authors or teams, all experienced in discourse research, are given exactly the same text to analyze. The text is an appeal for money by a lobbying organization in Washington, D.C. It informs, stimulates and manipulates the reader in a fascinating way. The joint analysis is far more insightful than any one team's analysis alone. This paper is our contribution to the book.

Rhetorical Structure Theory (RST), the focus of this paper, is a way to account for the functional potential of text, its capacity to achieve the purposes of speakers and produce effects in hearers. It also shows a way to distinguish coherent texts from incoherent ones, and identifies consequences of text structure.

Rhetorical Structure Theory suggests that discourse has an ideational structure or hierarchy which depends upon the speaker's communicative intent (the speaker's discourse purpose). To capture this ideational structure discourse is marked-up with ideational realtions (rhetorical predicates). In Mann and Thompson's version of the theory such relations are asymmetrical forming nucleus-satellite constellations. This represents the notion that one constituent is more central (important or focused upon) than another. [Note: see Boonie Meyer's work for a discussion of symmetrical (equally weighted) verses asymmetrical rhetorical relations and see Hoey/Shiffrin/Pilkington for a discussion of how this may compare to the key discourse markers of co-ordination and subjunction in comparative and logically sequenced texts]. In this paper a monologue is marked-up to form a heirarchical network of relations between increasingly large text spans which serve to represent the overall intention of the speaker. A small set of frequently used relations (rhetorical predicates) is identified (however the overall set is open). The important feature of RST is that it allocates a function (relation) to the span based on the function it performs in the text (the effect intended by the author). This describes its utility for the generation of text by Intelligent Systems since it describes a method by which such systems may access knowledge-bases to retrieve and re-structure propositional statements contained therin to serve a particular discourse purpose. Such discourse purposes may be dependant on many things, such as the current teaching style, the student model (in terms of language likely to be understood, relevant analogies etc). A relation is defined by two fields; constraints and effect. Constraints can apply to the nucleus, the satellite and/or their combination. The effect field contains a statement as to the plausible effect intended by the author and the locus of effect (nucleusor nucleus-satellite). Examples of relations include: evidence, circumstance, motivation, justify and restatement. As with most (all?) linguistic theories, RST can produce multiple analyses of the same text. Hence we may encounter difficultieswith interpretation between those involved in the dialogue.