As mentioned above for text organization, any kind of structure may be `frozen' and so the simplest method for constructing sentences again involves parameterized templates. For our example biography text above, for example, these might include patterns of the form ``?X was born in ?Y on ?Z.'', where the symbols ?X, ?Y and ?Z are filled in (somehow) to match the present user query. This method essentially precompiles the possible texts: usually the templates are written by hand. Problematic with the approach is the fact that the system knows very little about the units it is manipulating. Even the simplest grammatical variation (for example, if ?X received for some reason a plural filler, such as ``The Smith twins'') requires additional techniques for altering the template. It is also difficult to ensure that the text created reads fluently; whenever the templates are flexible enough to allow iteration (for example, re-use over all the works of art created, or over all the places studied), then repetitive and stylistically poor (and hence communicatively less effective) texts result. The differences between, and relative benefits of, `full' natural language generation and template-based surface generation have been discussed in some detail by Reiter [Reiter: 1995]. Since the method is so simple, however, it has been pursued in some practically-oriented systems. The more sophisticated of these begin to mix fully generated text fragments with pre-stored templates and it is likely that roles will be found for both approaches within single systems and applications (cf. [Reiter, Mellish and Levine: 1995]); accordingly there have recently been proposals for rather sophisticated template-based generation [Busemann: 1996]. The drawbacks of templates for surface generation have to be carefully weighed however: they are intrinsically monolingual, have poor re-use properties, can be stylistically stilted when appearing in texts, and become difficult to maintain and manage as systems grow.
Busemann now draws a distinction between `shallow' and `template' generation [Busemann: 2000]. The former relies on partially frozen representations at all levels of linguistic abstraction: this permits an easier carry-over from template to full generation as required by the particulars of the generation application--the advantages of simple generation can be combined with the advantages of deep generation on demand. The notion of partially fixed representations has been argued to be similar to the treatment of linguistic phenomena in TAGs [van Deemter and Odijk: 1997] and so is not intrinsically less well-founded than other linguistically motivated treatments. This is also carried out in the proposed RAGs architecture, where it is suggested that there is no real difference between prespecified frozen information and partial results from earlier computations that are being made available as packages rather than requiring re-computation, apart from the inability to take the package apart again in the case of templates proper. Most large-scale generators are now also including provision for mixing fully generated and user-specified portions.