Natural language generation often is characterized as a process in which the following two subprocesses can be identified. The conceptual part of this process decides what should be said in a given situation. The conceptual part thus constructs a `message', i.e. some sort of semantic representation. The other, grammatical, part of natural language generation then takes as its input such a semantic representation and decides how this meaning representation can be realized linguistically. For a discussion of this division of labor see for example [4], and also [57,91]; the distinction is also sometimes characterized in terms of a strategical and a tactical part.
The interest in this thesis will be in the second, i.e. grammatical (tactical), part of natural language generation. Thus, I will invariably assume that the conceptual part (or `planner') provides appropriate semantic structures -- our task will be to realize these semantic structures in natural language. A grammar defines the relation between semantic structures and phonological structures. Therefore, the task of generation is to compute this relation, given its definition in the form of a grammar.
Almost any modern linguistic theory assumes, that a natural language grammar not only describe the correct sentences of a language, but that such a grammar also describes the corresponding semantic structures of the grammatical sentences. Given that a grammar specifies the relation between phonology and semantics it seems obvious that the generator is supposed to use this specification. For example, for Generalized Phrase Structure Grammars (GPSG) [23][chapters 9 and 10] provide a detailed description of the semantic interpretation of the sentences licensed by the grammar. In my view, a generator based on GPSG should construct a sentence for a given semantic structure, according to the semantic interpretation rules of GPSG. However, [11] presents a generator, which is said to be based on GPSG, but which does not take as its input, as one would expect, a logical form, but rather some kind of control expression which merely instructs the grammatical component which rules of the grammar to apply. Similarly, in the conception of [22], the generator is provided with some kind of `deep structure' which can be interpreted as a control expression instructing the grammar which rules to apply. These approaches to the generation problem clearly `solve' some of the problems encountered in generation -- simply by pushing the problem away into the conceptual component. The generation problem I discuss in this chapter is somewhat more involved. I assume that the relation between semantic structures and phonological structures is defined by a declarative (constraint-based) grammar. The task of a generator is to compute, for a given semantic structure, the corresponding phonological structures, according to such a grammar.
As I will discuss in the next section, several approaches to the (grammatical) generation problem for constraint-based grammars are not completely satisfactory. As an alternative I propose a simple semantic-head-driven bottom-up generator (called BUG) which, as I will show, is superior in some respects to these others. 3.1 Furthermore I discuss some problems of BUG and some extensions to BUG. I also clarify the relationship of BUG to generators of the same family, such as the generators presented in [12], and [87].
I argue that the head-driven bottom-up approach should be favored because in this approach the order of processing is geared towards the input (head-driven) and the information available in lexical entries (bottom-up).