Semantic-head-driven Generation

In the third chapter, I discuss the generation problem for $\cal {R}$($\cal {L}$) grammars, and propose a semantic-head-driven generation technique which I claim is superior to some other approaches to generation that have been proposed previously. In order for this technique to be useful, the semantic structures should be defined in a lexical and semantic-head-driven fashion. Typically, the semantics of a phrase is a further instantiation of the semantics of the semantic-head of the phrase. And furthermore, the semantics of the other parts of the phrase, is determined by the semantic-head of the phrase. Semantic-head-driven generation proceeds by predicting the lexical head of a given phrase. This lexical head is then connected upward to the goal by successive rule applications.

Clauses which have a daughter with identical semantics as the mother node, are called chain-rules, and are, for the purpose of the generation meta-interpreter, represented as:

$$\mbox{\it cr}(\mbox{\rm Head},\mbox{\rm Mother},\langle \mbox{\rm D}_{1} \dots \mbox{\rm D}_{n}\rangle) \mbox{\tt :-} \phi.$$
On the other hand, rules without a head (such as lexical entries) are called non-chain-rules and are represented as

$$\mbox{\it ncr}(\mbox{\rm Mother},\langle \mbox{\rm D}_{1}\dots\mbox{\rm D}_{n}\rangle)\mbox{\tt :-} \phi.$$

The following meta-interpreter in figure 5.8 defines a simple instantiation of the semantic-head-driven processing strategy. In chapter 3 several improvements and variations of this strategy are discussed.

Figure 5.8: A simple version of the semantic-head-driven meta-interpreter. In chapter 3 some improvements of this algorithm are discussed.

The semantic-head-driven strategy is compared with the top-down oriented approaches of [108] and [17], and the chart-based approach of [83]. Semantic-head-driven generation is more general than those top-down oriented approaches as it handles certain linguistically motivated left-recursive analyses of subcategorization. On the other hand semantic-head-driven generation must be favored over Shieber's chart-based generator because it allows certain non-compositional analyses, such as those proposed for idiomatic constructions.

An important problem for semantic-head-driven generation are linguistic analyses which are based on a threading implementation of head-movement. I provide an analysis of verb-second in Dutch along those lines, which is problematic for semantic-head-driven generation (because it violates the assumptions about the construction of the semantic structures mentioned above). A further problem for semantic-head-driven generation is exemplified by an analysis of English raising-to-object constructions.

Although I do provide some ad-hoc solutions to these problems, it can be argued that a more general solution is available only, if we allow grammars in which phrases can be built with other operations than concatenation. For example, raising-to-object constructions can be analyzed using Bach's `wrap' mechanism. The resulting analysis is not problematic for semantic-head-driven generation. On the other hand, such more powerful operations on strings lead to an increased burden on the parser.