For the case of cognitive generation the picture is unfortunately not that clear, because cognitive generation strategies have only recently been given significant consideration. However, from the work presented in [Kempen and Hoenkamp1987], [DeSmedt and Kempen1987], [Levelt1989], and [Pechmann1989], it seems to be evident that generation is also performed incrementally. Furthermore, all these approaches also assume that grammatical generation is competence-based, i.e., grammatical knowledge is used for the production of sentences. Moreover, there is psychological evidence that the same grammar is used for performing parsing and generation - as we have already outlined in the first chapter of this thesis.
In all of the work cited above it is assumed that the input is of semantic nature given as a functor/argument tree [Levelt1989]. However, it is assumed that the parts of the meaning of an utterance are passed in a piecemeal fashion. Since it is now possible that arguments can arrive before their functors or vice versa it is assumed that the grammatical structure is built up in a mixed bottom-up/top-down way (see [Kempen and Hoenkamp1987] and [DeSmedt and Kempen1987]).
Furthermore, for reasons of parsimony it is assumed that generation is also performed in a deterministic way [Kempen and Hoenkamp1987], i.e., not all grammatical possibilities for an utterance to be produced are explored simultaneously. However, because of the fact that humans are able to monitor what they have heard or said, it is assumed that in some situations some sort of re-synthesis (or revision) will take place [Levelt1989]. However, how these mechanisms are actually be performed has not been specified computationally. Thus, it seems to be an open question whether in the case of cognitive generation chart-based or backtracking strategies should be considered when performing the revision process.
We have shown in this thesis that methods for performing self-monitoring and revision during generation are suitably realized by means of a tight integration of parsing and generation. As we have made clear, such an approach can be made efficiently using the uniform tabular algorithm in combination with the item sharing approach. On the top of these methods we have developed a chart-based incremental monitoring strategy. The results of this thesis suggest that revision too benefits from the use of a tabular based strategy, since it can be performed purposively and efficiently.
In summary, there seem to be some evidence that our uniform computational model conveys some cognitive plausibility, and it would be of great interest whether there actually exists some empirical data that could support this assumption. In this thesis we have laid out important theoretical and practical foundations for achieving competence-based performance language models. We are strongly convinced that our work will help to realize ``natural'' natural language systems and we have outlined how our approach can valuable contribute to important new lines of future research directions.