Natural language processing is often viewed as a complex modular system consisting of interacting components whose mode of operation may be radically different. The different components can be collected into two subsystems according to their different tasks:
In the case of language understanding, the basic task of the linguistic system is to determine a semantic representation of a given utterance from which the conceptual system can draw general inferences, e.g., in order to resolve anaphora, and to determine the speakers intention behind the utterance. In order to perform these tasks the conceptual component acts primarily on the basis of world knowledge, discourse model and situation knowledge. The basic knowledge sources of the linguistic system are the grammar and the lexicon of a language which represents in a declarative way the relation between well-formed utterances and their associated semantic representations. The central process for analysing the grammatical structure of a given utterance is called parsing. The output the parser delivers is the set of possibly all semantic representations that the grammar associates to that utterance.
In most natural language understanding systems NLU, i.e. systems where only language understanding is considered, a clean separation between the two systems is assumed. In this view, the linguistic system is modular in the sense, that it is the only component that is concerned directly with the form and content of the grammar while the conceptual system is the only one that is responsible for general inference. In these kind of models the semantic representation specified in the grammar serves as an intermediate representation between the linguistic and conceptual system.
This division of labour is also the basis of current natural language generation systems NLG developed in the area of artificial intelligence and computational linguistics (see e.g.,[Dale et al. 1990, Paris et al. 1991] for a collection of state-of-the-art reports) as well as in cognitive science (see e.g., [Levelt1989]). It is an increasing consensus that the input to an NLG should be of the form of a COMMUNICATIVE INTENTION, i.e. some goal that the speaker wants to communicate by means of natural language. To be able to produce an utterance that adequately communicates the speaker's goal several subtasks have to be performed, e.g.,
Most of the work of determining the content of a discourse is done by
the conceptual component. In
order to perform its tasks it also takes into account
world knowledge, discourse and dialog
knowledge as well as knowledge about the interlocutors (however, not
necessarily represented in the same way or the same account of knowledge).
The linguistic
system is responsible for realizing the content of a discourse determined by
the conceptual component in a natural language. On the basis of a
grammar and a lexicon the grammatical structure of a given content has to
be produced in order to determine a well-formed utterance. The
process that is responsible for this task will be called
grammatical generation or short generation.
Although the modular design of an NLG into a conceptual and linguistic component has been proven fruitful for the investigation of natural language production it is a matter of active debate what the input for the linguistic component exactly should look like. For example, many researchers (e.g., [Danlos1987, Appelt1985, Hovy1987, Finkler and Neumann1989, Reithinger1991, Neumann1991a]) have argued that the conceptual and linguistic decisions are strongly dependent upon each other; e.g., in the case of lexical gaps, choice between near synonymous or paraphrases a communication between the two phases is required. On the other hand in many approaches it is assumed that the conceptual component has to provide all information needed by the linguistic component to make decisions about lexical and syntactic choices (e.g., [McDonald1983], [McKeown1985], [Levelt1989], [McKeown et al. 1990]). They assume that feedback from the linguistic to the conceptual component would be exception rather than rule. The different views have lead to different architectures where the line between the conceptual and linguistic component have been drawn in different ways, i.e. there is no such clean separation between the two components as it is the case for most NLU systems.
Figure 5.1: The Architecture of an NLS using a reversible grammar.