Generating Formatted Texts: the role of document structure
To date, most researchers in NLP have assumed a more or less one to one mapping between the presentational structure (i.e. layout)of a document and its discourse structure:<P>
To generate a text they typically build a text plan that is a rhetorical structure and then render it as text by faithfully following the order in which the propositions occur in the plan, making each one a sentence -- with maybe a bit of aggregation here or there. The result is<P>
- a very limited number of possible texts (often only one);<P>
- texts that are fairly stilted;<P>
- the use of little, if any, meaningful formatting.
To analyse a piece of text, they break it down into clauses and try to build a hierarchical structure with rhetorical relations as the glue. The result is that it is rarely possible to build from this a rhetorical structure that conforms to current discourse theories (e.g., RST), or one that would, if it were input to a generator, reproduce the same text. The usual reaction to this is to lay the blame on the particular theory being used.
I will argue that the key to resolving these issues is to recognise the role of document structure in the generation and comprehension of texts, and the distinction between document structure and rhetorical structure. What NLG systems are actually generating, or what the discourse researcher is actually analysing, is not a rhetorical structure, but a document structure -- that is, a structure composed of textual elements like chapters, sections, paragraphs, orthographic sentences and clauses, indented lists and the like. This structure may, of course, be isomorphic with the rhetorical structure, but it is very often not. Moreover, cases where it is not are not necessarily evidence of inferior texts.
At the ITRI, we have developed a theory of document structure (Scott et al, forthcoming, Computational Linguistics) based on the notion of a text grammar, as formulated by Nunberg (1999). This structure is a realisation of the underlying rhetorical representation; it is here where the decisions on how to divide the message into textual units, how to order them, and how to lay them out, are represented.
In this talk, I will present the theory of document structure and some exemplars of how it has been used in an NLG system. I will then show how and why disentangling discourse structure from document structure will help avoid the problems outlined above.