A further methodology is to work on complete generation systems that attempt to address all areas and aspects of the generation process. Complete generators are obliged to attempt a combination of horizontal (achieving breadth of coverage) and vertical (achieving control of that coverage) methods. There has also been a concentration on `complete' natural language processing systems, which include not only generation functionality, but also, at least, natural language understanding, knowledge representation, and inferential capabilities. This has, particularly in Germany, produced a number of substantial systems, including: WISBER [Horacek: 1990], HAM-ANS [Busemann: 1988], LILOG [Novak: 1991] and, most recently, VERBMOBIL [Kay, Gawron and Norvig: 1994].
It is sometimes maintained that this is a useful strategy in that it forces all aspects of the language problem to be considered, thereby reducing the danger of unfillable `black holes' (i.e., non-dischargeable homunculi). It can equally be maintained, however, that it forces compromises in approaches that are still too immature to be stable, resulting in components of limited functionality that are difficult to extend. Advances here do seem to have been limited to the development of a set of techniques and experiences useful in building such systems; deeper theoretical advances for NLG have been considerably fewer and, arguably, not necessarily won from the experience of integrating the generation component into the large-scale system itself. This has often been caused by an underestimation of the effort and impact of incorporating NLG functionality. Considerations of knowledge representation, rhetorical organization, mapping between levels, and between alternative logical forms are essential aspects of NLG.
Somewhat paradoxically, then, `complete' natural language systems have been based more on parsing technologies augmented with semantic inferential capabilities: a relatively narrow basis when compared with that necessary for generation. Integrating generation into systems of this kind can require costly reengineering of the host system or reduction in the scope and power of the generation component. A more successful strategy would certainly be to allow the requirements of NLG to influence the initial design more strongly. Relevant arguments here in the context of the LILOG project are offered by Novak [Novak: 1991].