One general requirement for NLG systems is to construct various kinds of structure corresponding to the distinct strata identified above. The techniques applied for this typically differ depending on both the nature of the structures involved and the intended application/use. Thus lexicogrammatical structures are usually treated quite differently from, for example, text structures. This is not a logical necessity: attempts to represent linguistic information in terms of feature structures that can be combined by unification, classification, or some other form of logical inferencing (e.g. [Elhadad and Robin: 1992,Grover, Brew, Manandhar and Moens: 1994]), or approaches that adopt a single linguistic account throughout (e.g., [Bateman and Teich: 1995]), have also been attempted because of their greater promise of homogeneity. There are also approaches that deny the utility of defining distinct modules at all (cf., e.g., [Appelt: 1985b,Danlos: 1987,Ward: 1992]). However, most approaches still exhibit dividing lines in techniques employed between lexicogrammar, semantic information and text organization: this can be manifested in differences in representation, processing, temporal sequencing, or any combination of these. One particularly deep-seated distinction is that between the selection of content (`what to say') and the expression of that content (`how to say it'); this was introduced by Thompson [Thompson: 1977] in terms of the `strategy' and `tactics' of language production respectively. This distinction is commonly echoed in NLG discussions in the two `modules' of deep generation and surface generation (lexicogrammar) respectively.
All such distinctions simplify locally the construction of modules but also create significant interfacing problems between those modules. For example, it has been argued that any specific ordering of decisions can be shown to be inappropriate in the general case and that more flexible control regimes are necessary (e.g., [Danlos: 1987,Emele, Heid, Momma and Zajac: 1992,Elhadad, McKeown and Robin: 1996]). It may be that some semantic configuration is normally expressed by a particular grammatical structure, but in some exceptional case the lack of a necessary lexical item in a language may require a different grammatical structure; this establishes a dependency between choice of mapping and available lexical material. Similarly, once a rhetorical organization has been planned, it may be the case that the grammatical constructions available offer alternative structurings or that the particular content to be expressed exhibits further regularities that could be presented by suitable aggregations. Such dependencies can be found between almost any posited modules of the generation process.
Two techniques for dealing with this are the use of a blackboard architecture (e.g., [Nirenburg, Lesser and Nyberg: 1989]) where processes can run independently, posting their results and requirements for other processes in a commonly accessible data structure called the `blackboard', and the use of unification where partial structures are combined (unified) opportunistically as they become increasingly specified according to given constraints. The latter, while theoretically and formally the most attractive, is computationally expensive and needs fine control in order to remain usable (cf. [Elhadad and Robin: 1992]). More rigid control strategies that nevertheless attempt to avoid the problems of strict temporal sequencing are models involving `feedback' or `revision' from one component, e.g., lexicogrammar, to another e.g., text planning (e.g., [Hovy: 1988a,Reithinger: 1991,Rubinoff: 1992,Inui, Tokunaga and Tanaka: 1992,Robin: 1996]).
Whereas the division and temporal ordering between components and their decisions is criticized theoretically, it remains the case that an imposition of ordering achieves a significant simplification in design. A system that strictly orders the strategic and tactical phases of generation has what is termed a `pipeline' architecture (cf. [Reddy: 1979,Reiter: 1994]) and, despite many weaknesses, the simplicity of this architecture has made it the architecture of choice in more practical generation systems. The likelihood of success of the pipeline strategy can also be significantly improved by enriching either the input representation, or the generator's requirements, or both, to provide a better match: the richer the input representation, the more chances are offered for providing sufficient information to the next component to do its job without requiring backtracking.