Automated Discourse Generation (also commonly known as `Text Generation' or `Natural Language Generation': NLG) is the branch of computational linguistics, or natural language processing, concerned with the question of how `texts' in natural human languages can be created automatically by machine. Both within and outside of natural language processing this is sometimes described as simply the `inverse' of the problem addressed by its elder siblings--natural language understanding and machine translation--but, in fact, NLG has developed out of a quite different set of concerns and has given rise to an almost disjoint subfield of computational linguistics with its own methods, goals and theoretical assumptions.
NLG is pursued for reasons both of theory and of application. On the theoretical side, it is pursued in order to reveal more about how language works and how it is structured. A computational account of the human linguistic system that is sufficiently well specified to allow the automatic generation of natural texts must combine a breadth and detail of linguistic description that is unusual in non-computational text linguistics, semantics and syntactic theory. This pushes the finer articulation of linguistic theory further along a broad front of issues. On the practical side, NLG attempts to provide usable solutions to problems of text production in practical contexts. NLG techniques are being investigated that enforce consistency during text production (e.g., consistent selection of terminology, of desired grammatical forms, `controlled languages', etc.), that maintain up-to-date texts in the face of changing situations and requirements, and that re-use information sources for the automatic creation of documents tailored to varieties of audience groups with differing informational and presentational needs. All of these areas represent major problems and opportunities in real-world document production.
The demand for natural language texts providing information of every conceivable kind is currently rocketting. Thus it is certain that NLG will form a key information technology in the future--there is scarcely an area of information presentation where beneficial applications of automatic text production cannot be imagined. As a consequence, there are a growing number of NLG systems now finding practical use, while the demands of real-world applications are also having an increasing impact on the approaches and questions raised in NLG work in general. Examples of established NLG applications include the generation of weather reports from meteorological data in multiple languages [Kittredge, Polguère and Goldberg: 1986,Coch: 1998], the generation of letters responding to customers' queries [Springer, Buta and Wolf: 1991,Coch, David and Magnoler: 1995], the generation of letters supporting tobacco smokers in attempts to escape their addiction [Lennox, Osman, Reiter, Robertson, Friend, McCann, Skatun and Donnan: 2001], and the generation of reports concerning, for example, project management [White and Caldwell: 1998] and the environment [Busemann and Horacek: 1997].
There is also a far wider range of applications already in use where textual presentations of information are created dynamically but where NLG techniques are not yet being applied. The adopted solutions in such cases range from the simplest so-called `mail merge' techniques, as offered by some word processors in order to repeatedly insert addresses or other textual information into a form letter, through to conditionalized paragraph inclusion and fill-the-slot text patterns. For the casual user, it is often not easy to tell the difference between texts prepared by hand, texts constructed using simple techniques, or full natural language generation using NLG technology. This is as it should be; a successful example of NLG should not draw attention to itself--it should simply be a perfectly natural piece of text production ideally meeting the needs and knowledge of the reader/hearer. This is, however, problematic for NLG research and development in that it makes the effort required for successful NLG harder to `see'. Since users are not generally aware of it unless something goes wrong, there is little appreciation both of the possibilities and the complexities of full natural language generation. This lack of appreciation spans both users and application developers who might see the utility of providing flexible, automatically produced natural language texts, but who are unaware of the kinds of complexities this can involve, the range of technological solutions that are available, and the level of investment required to `get it right'--that is, to find solutions that can scale-up as required rather than requiring re-engineering as desired language complexity grows.
NLG is therefore still very much at the beginning of its career; the full range of possible applications has barely been scratched. Current work is investigating the possible roles of NLG technology in an as diverse a range of areas as the automatic generation of user-appropriate technical documentation in multiple languages (cf. [Reiter, Mellish and Levine: 1995,Rösner and Stede: 1994,Peter and Rösner: 1994,Kruijff, Teich, Bateman, Kruijff-Korbayová, Skoumalová, Sharoff, Sokolova, Hartley, Staykova and Hana: 2000]), of instructional texts (cf. [Not and Stock: 1994,Paris, Linden, Fischer, Hartley, Pemberton, Power and Scott: 1995]), of patent claims (cf. [Sheremetyeva, Nirenburg and Nirenburg: 1996]), of information in computer-supported cooperative work (cf. [Levine and Mellish: 1994]), of patient health information and patient education documents for patients with differing levels of medical expertise (cf. [Cawsey, Binsted and Jones: 1995,DiMarco, Hirst, Wanner and Wilkinson: 1995]), of medical reports [Li, Evens and Hier: 1986,Cawsey, Webber and Jones: 1998,de Carolis, de Rosis, Andreoli, Cavallo and De Cicco: 1998], of virtual encyclopedia entries [Milosavljevic and Dale: 1996 a] and personalized, dynamic descriptions of museum exhibits [Oberlander, Mellish and O'Donnell: 1997,Androutsopoulos, Kokkinaki, Dimitromanolaki, Calder, Oberlander and Not: 2001], of dialogue contributions in appointment planning [Uszkoreit, Backofen, Busemann, Diagne, Hinkelman, Kasper, Kiefer, Krieger, Netter, Neumann, Oepen and Spackman: 1994,Busemann, Declerck, Diagne, Dini, Klein and Schmeier: 1997], and of natural language contribuions in interfaces to databases (cf. [Androutsopoulos, Ritchie and Thanisch: 1995]) or information systems (cf. [Bateman and Teich: 1995]), expert systems (cf. [Allgayer, Harbusch, Kobsa, Reddig, Reithinger and Schmauks: 1989]) and public assistance or information systems using spoken language (cf. [Davis and Schmandt: 1989,Kaspar, Fries, Schuhmacher and Wirth: 1995,Strik, Russel, van den Huevel, Cucchiarini and Boves: 1996]). NLG thus covers a very broad territory and its potential ramifications for how people and machines interact are immense.
Given both this potential and the range of interests involved, it should not be surprising that NLG has experienced very rapid growth over the past decade. This complicates the task of providing a thorough review of the field. Until the end of the 1980s it was still almost possible for a review to list the most significant NLG systems in existence. This would no longer be an effective strategy: the most extensive list of NLG systems is that of Bateman and Zock [Bateman and Zock: 2001], which gives details of over 300 systems and is regularly updated as new systems appear. The present review will therefore be considerably more focused. Our main concern will be set out the general problems and techniques that stand behind all attempts to develop and use NLG technology. Although particular problems can always be avoided or techniques simplified in the face of particular application demands, this is always accompanied by a corresponding loss of functionality. The real task in designing, or deciding upon, NLG technology or of particular theoretical positions is to understand how much functionality is being sacrificed by particular design decisions and at what cost. This is only possible against the background provided by a more general standpoint than that provided by particular system designs or application areas. We start, therefore, by introducing the basic problems addressed within NLG, and then move on to show the main techniques proposed for solving or avoiding those problems. We restrict attention further to problems inherent to the task of NLG as such rather than considering the problems of particular generation system implementations, although sufficient references will always be given to allow the interested reader to follow up on the details.