The uniform tabular algorithm just developed together with a reversible grammar constitutes the grammatical competence base of a natural language system. The grammar declaratively describes the set of all possible grammatical well-formed structures of a language and the uniform algorithm is able to find all possible grammatical structures for a given input - at least potentially.
Grammatical knowledge and processing in this sense is necessary to render possible natural language processing. This is not only a principal emphasis in computational theories of human language ability but also in artificial intelligence [Wahlster1986]. However, if we accept the view that the primary motivation for developing natural language systems is to facilitate human-machine communication and/or to extend, clarify or verify theories that have been put forth in linguistics or cognitive psychology then it is necessary to investigate the relationship between linguistic competence and linguistic performance. This means that we have to consider the grammatical competence under the perspective of language use, for example, in order to explain how preferences, disambiguation, paraphrasing, incrementallity, monitoring and revision or active vs. passive language use can be explained on the basis of language competence.
The scientific goal followed in this chapter is to explain how such methods can be modelled by means of the uniform grammatical competence base. Of course, it would exceed the scope of this work to consider all of these aspects. Therefore, we concentrate ourself on the issue of monitoring, revision and paraphrasing during natural language generation. In the final chapter we outline how the uniform grammatical model can also be used to accommodate other aspects, like the use of preferences and fully incremental text processing.
It is a fact that speakers monitor what they are saying and how they are saying it [Levelt1989]. When they are making a mistake, or express something in a less felicitous way, they are able to make a repair. It is furthermore evident that speakers monitor almost any aspect of their speech, ranging from content to syntax to the choice of words to properties of phonological form and articulation (see [Levelt1989], page 497). In this thesis we are concerned with monitoring on the grammatical level and explain how revision can be performed in order to reduce the risk of generating ambiguous sentences that could be misunderstood by the listener. However, the methods we are going to develop can also be used to handle other monitoring and revision strategies. We demonstrate this by showing how the new methods are used during natural language understanding as a means of disambiguation. Furthermore, in the final chapter we outline how these methods in combination with preference strategies can be used to realize hearer adaptable strategies.
The underlying strategy followed in the methods to be presented is a tight interleaving of parsing and generation. It will turn out that the uniform tabular algorithm in combination with the item sharing approach leads to an elegant and effective specification of these performance-oriented methods. Furthermore, the methods can be straightforwardly be combined with the uniform algorithm in such a way that the original behaviour of the uniform algorithm is not affected. This means that it is possible to interrupt the monitoring and revision process in such a way that further processing proceeds in the normal un-monitored way. Thus, we achieve an elegant synergy of competence and performance-oriented processing.