Generator tasks can be organized in terms of the kinds of information with which they are concerned--or, adopting a `decision' metaphor (cf. [McDonald: 1980]), with the kinds of decisions they have to make. In this section we have further differentiated the tasks of NLG along two dimensions of linguistic specification: stratification and metafunction--two dimensions drawn from the functional semiotic approach of Systemic Functional Linguistics (cf. [Halliday: 1978]). Information can be more or less abstract (strata: register, semantics, lexicogrammar, graphology/phonology), and it can be concerned with different kinds of meanings (metafunction: ideational, interpersonal, or textual).
Accounts naturally differ in their segmentation of the overall space of NLG tasks; although often there is more difference in terminology than difference in content. Regardless of any particular segmentation the general selections and possibilities--and hence the abstract tasks--that are entailed by the linguistic system remain. A reader/hearer draws the same conclusions from the appearance or non-appearance of a particular kind of variation regardless of whether this is `intentional' on the part of the system that produced the text or not: a lack of motivation cannot alter the fact that other selections are in general possible in the language used. For human interpreters, generated texts such as our examples above consist of sentences that are read as instances of natural language and, as such, readers will interpret them as possessing this semantics regardless of whether the generation system itself used that semantics in their construction. This can mislead a user if the consequences of the semantics actually implemented differ too widely, or are too restricted, compared to those expected. It is thus advisable to be aware of the semantic commitments that individual phrasings of clauses and other linguistic units enter into even when a generation system itself has not used those commitments.
The `input' adopted to any system must also be seen against this background: to the extent that an input diverges from the semantic specifications that are linguistically motivated, the less likely it is that a generation system will be able to control the corresponding areas of linguistic variation appropriately. Those aspects of the background that an input makes explicit determine those areas that may be flexibly reflected in the texts generated; those areas that do not appear remain outside the control of the generator and the resulting texts will be fixed, or predetermined, in just those areas. Any restriction in the linguistic variation that can be controlled by an NLG system also limits both that system's flexibility and its possibilities for application. Knowing how an input and a linguistic semantics differ from each other provides a firm basis for evaluating the functionality that a generation system can reasonably aim for in a particular context of use; the costs and benefits of application can then be assessed more accurately.