Vertical methods take one particular area defined with respect to some linguistic resource stratum and attempt to fill in sufficient details from that and related areas in distinct strata in order to demonstrate control of the original area selected. Examples include the lexicogrammatically defined areas of referring expressions (cf. [Appelt: 1985a,Dale and Haddock: 1991,Reiter: 1990,Horacek: 1995,Henschel, Cheng and Poesio: 2000]), tense (e.g., [Gagnon and Lapalme: 1996]), and conjunctions (e.g., [McKeown and Elhadad: 1991]). All of these grammatical phenomenon can be usefully linked to notions of discourse structure and to discourse development. In order to select an anaphoric `she', for example, the generator also needs to know that the `discourse referent' has been introduced sufficiently recently as to be recoverable by the hearer. `Recency' appears to be best defined with respect to discourse structure rather than purely temporal recency, although here it is clear that issues of `referential adequacy'--i.e., does the referring expression give enough information to allow identification of the referent--are also important. Issues of tense selection also show clear links with indications of text structure. In contrast to horizontal methods, vertical methods therefore require particular attention on the relations between levels of representation. Work on `lexical choice' is also often best interpreted as a vertical development methodology, as are many of the functional modules identified in the RAGS project in that they too relate to several distinct levels of linguistic abstraction.
In theory, areas can be selected for attention from any linguistic stratum. For example, there has also been focus on particular text types that have been, and continue to be, formative for NLG research. One such is explanation: here the NLG system assists in presenting explanations of some behavior to the user. When the explanation is of some internal computational processes, such as, for example, the recommendations from some expert system, then the generator can be in a good position to provide such information--given that the expert system has represented its decisions and their motivations sufficiently explicitly (cf. [McKeown and Swartout: 1987,Moore and Paris: 1993]). Explanations also arise in educational contexts and in tutoring systems (cf. [Woolf and McDonald: 1983,Cawsey: 1990,Suthers: 1991]). More recently, the text type of instructional texts has received considerable attention (e.g., [Not and Stock: 1994,Paris, Linden, Fischer, Hartley, Pemberton, Power and Scott: 1995,Kosseim, Tutin, Kittredge and Lapalme: 1996]); this includes user manuals and instruction leaflets for consumer goods. Our example generated text above is an example drawn from work on a further text type, that of biographies--particularly those found in encyclopedias [Bateman and Teich: 1995]. One of the most extensive attempts to date to study a set of phenomena explicitly within the vertical methodology has been the European Union research project DANDELION (e.g., [Bateman: 1994]), where connections have been drawn between a variety of lexicogrammatical phenomena and text types via semantic configurations.