next up previous contents
Next: Evaluation Up: NLG Methodologies Previous: Whole system methods   Contents

Process methodologies

Some researchers argue that it is precisely the investigation of control regimes for generation that is central to the uniquely computational investigation of texts and their creation that NLG offers. The style of NLG pursued by Meteer and McDonald (e.g., [Meteer: 1992,McDonald, Vaughan and Pustejovsky: 1987]) is the primary example of this. Meteer draws on the apparent `efficiency' of natural language production by humans--i.e., fast, fluent, without ``talking oneself into corners''--to motivate the particular architectural decisions made in her account. Others focus similarly on process issues simply because of their background in computer science; however, NLG systems that are developed only by computer scientists often suffer from a lack of linguistic motivation. Issues are then not `theorized' linguistically, making their correlation with other phenomena and mechanisms more difficult.

Particular kinds of approaches focusing on process issues impose various additional constraints on the process of generation. Perhaps the most widespread of these is that of incremental generation (e.g., [Kempen: 1977,de Smedt and Kempen: 1987,de Smedt: 1990,Reithinger: 1991,Abb, Günther, Herweg, Lebeth, Maienborn and Schopp: 1996]). Here the task for the NLG system is to use its resources in order to build up texts progressively in time as uttered, making use of information from other levels of abstraction as soon as it becomes available. This is usually motivated on grounds of psycholinguistic plausibility. A somewhat related approach is that of generation using parallel processing methods (cf. [Kitano: 1990,Ward: 1992]).

A different style of approach focuses on reversible accounts where the linguistic resources defined must function both for generation and analysis (cf. [Neumann: 1991,Strzalkowski: 1994]). Although this appears clearly desirable on grounds of economy, true bidirectionality has proved difficult to achieve in practise. A variant of this is to allow self-monitoring (e.g. [Neumann and van Noord: 1994]) where the resources available are used to construct both sentences and the possible analyses of those sentences simultaneously--thereby avoiding the introduction of unwanted syntactic ambiguities. Finally, there are approaches which accept that there are pervasive interactions between information of various kinds and which therefore seek to ensure that the processes of generation can support such interactions (cf. [Emele, Heid, Momma and Zajac: 1992,Elhadad, McKeown and Robin: 1996]). In all these different styles of approach, the additional tasks set may have implications for the resources used: that is, resources developed independently of the extra constraints imposed may not support the required task.


next up previous contents
Next: Evaluation Up: NLG Methodologies Previous: Whole system methods   Contents
bateman 2002-09-21