next up previous contents
Next: Interactions between tasks/modules Up: From application to NLG Previous: Content interpretation   Contents

Summary

Once some application information or input specification is selected for expression, it must be recoded appropriately for generating surface sentences or utterances that contribute successfully to an unfolding text or dialogue. This can be seen as a process of progressive refinement of the information initially selected for presentation. All stages of refinement may have consequences for the global tasks of content selection and organization. The component subtasks contributing to such refinement can be outlined by separating out the kinds of information and organizations involved according to the levels of abstraction required by distinct linguistic strata. Note that no NLG system can avoid taking some position with respect to these kinds of organization even if this is not done explicitly. If particular organizations are overlooked, ruled out by inadequate theory, or simply omitted, then those selections that are present in the NLG system are hardwired.

The only general statement that can be made about the `input' to generation systems is in terms of linguistic stratification. The input can be drawn from different linguistic strata, accordingly varying in degree of `abstraction'. Certain systems adopt very low levels of abstraction--barely different from a sparse syntactic structure or sentence template [Lavoie and Rambow: 1997]; others adopt very high levels of abstraction including communicative goals such as `be persuasive' [Hovy: 1988a]. Which strata are adopted for providing input depends on the purpose of the component: there is no a priori correct stratum which is the input. Any notion of the `input' to the generation process being identified with some `content' to express is therefore also problematic; the question of selecting text content is a difficult one that probably does not have a single `answer'. Some core content may be selected on the basis of a recognized information need on the part of the intended audience, but this may well need further enrichment or reduction depending on what the user will understand, what the system wishes to communicate, and the informational requirements of the various kinds of organization that contribute to any complete text. Some `inputs' also mix levels of abstraction in their specifications--this can either be deliberate, generally in the service of usability, or accidental. Accidental mixing most often occurs with inputs defined extrinsically to an NLG component since modelling principles for application knowledge are still sufficiently unrefined as to confuse linguistic and non-linguistic information of various kinds within single representations (for some of the problems with this, see the useful discussion in: [Lang: 1991]).


next up previous contents
Next: Interactions between tasks/modules Up: From application to NLG Previous: Content interpretation   Contents
bateman 2002-09-21