• Keine Ergebnisse gefunden

To conclude this chapter, we summarise our preferred choices in the five areas discussed above where linguistic deduction algorithms can be varied.

Direction of processing: A bottom-up algorithm is preferred because of its po-tential for data-driven and robust linguistic deduction. Wherever possible, a directed algorithm that combines the benefits of bottom-up and top-down processing should be chosen.

Selection function: The choice of a selection function is not really the focus of this thesis. For simplicity, we assume that the goals of a clause have already been ordered, and always use the first unblocked goal.

Memoing: We choose memoing in the form of Earley deduction, which has all the advantages that chart parsing has for NL parsing, and is especially well suited for search strategies such as best-first processing.

Constraints: We use the constraint language defined in section 1.1.1 and check all constraints immediately.

Search Strategy: We choose to employ a best-first search strategy, based on a probabilistic preference model elaborated in section 4.1.

Before presenting a processing model that instantiates these choices in chapter 3 (bottom-up Earley deduction), we will detour and show in chapter 2 how principle-based grammars such as GB andhpsgcan be brought into a rule-based form that can be processed efficiently by the proposed deduction system.

From Principle-Based

Grammars to Rule-Based Grammars

2.1 Principle-Based Versus Rule-Based Gram-mars

This chapter is concerned with the claim that there are two different kinds of grammatical theories: rule-based grammars in which rules describe how grammat-ical constituents areconstructed and principle-based grammars in which universal principlesconstrain the possible constituents.

First, we define more precisely what is meant by principle-based and rule-based grammar.

Principle-based grammar A set of (universal or language-specific) principles must hold of all constituents described by a grammar. A few very general rules may exist for forming the objects to which the principles are applied.

The prime example for this kind of approach is Government-Binding Theory (gb) [Chomsky, 1981; Chomsky, 1986; Chomsky, 1993].

Rule-based grammar Every constituent must instantiate one rule of the gram-mar. Examples of rule based grammatical theories/formalisms arelfg,dcg, andpatr.

A reason for the move from rule-based to principle-based grammars is the desire to provide a theory of language that exhibits explanatory adequacy, not just observational adequacy (accounting for the observed data) or descriptive adequacy

43

(assignment of the right structural descriptions to sentences). Principle-based grammars achieve explanatory adequacy by showing that a large variety of phrase structures can be predicted from the interaction of a smaller number of general principles.

Government-Binding Theory is the best example of a principle-based syntactic theory, but the use of general principles also plays an increasing role in grammar theories arising from the tradition of phrase structure grammars. Generalised Phrase Structure Grammar (gpsg) [Gazdar et al., 1985] uses phrase structure rules which are augmented with principles such as theHead Feature Principle. In Head-Driven Phrase Structure Grammar(hpsg) [Pollard and Sag, 1987; Pollard and Sag, 1994] phrase structure rules are largely abandoned, and the grammar rests on the lexicon, and on principles which constrain the possible structures.

A big advantage of the principle-based approach is the compactness of the representation of linguistic knowledge. To obtain the effect of k principles with nj degrees of freedom, which can be stated inO(n1+n2. . .+nk) lines of code, a rule-based approach has to useO(n1∗n2. . .∗nk) rules (cf. [Berwick, 1991]).

While it is possible to process principle-based grammars directly in a type inference system such as cuf or tfs, such direct processing suffers from serious efficiency problems, due to the fact that many inference steps must be performed at runtime. In order to overcome these efficiency problems, many implementations ofhpsggrammars often use rule-based formalisms such asaleoralep, and take one of the following approaches to the handling of principles:

the grammar rules are written in such a way that they respect the principles of the grammar, or

the principles are added to the grammar rules as procedural attachments.

The first approach has the disadvantage that the compactness of the gram-matical specification permitted by principle-based formalisms is lost. The second approach has the disadvantage that it increases the amount of processing needed at runtime because the principles often have disjunctive formulations for different types of structure.

We will present a formalism which permits a principle-based statement of the grammar, and uses program transformation techniques (partial deduction) in order to derive a rule-based grammar. This happens by specialising the principles of the grammar to particular phrase types and eliminating as much disjunction as possible at this stage. In the most extreme case, this step can transform the grammar into its disjunctive normal form (DNF), but in practice DNF is often either undesirable or impossible. It may be undesirable when it leads to a very large number of phrase structure rules, and impossible when it would lead to an infinite number of rules because of the use of recursively defined relations (such as append/3) in the principles of the grammar. Therefore, the outcome of the

program transformation is a grammar with a number of phrase structure rules to which some constraints and goals are attached which cannot be reduced at compile time1because they need to be instantiated during the processing to allow a terminating (and often deterministic) computation.

The converse operation, transformation of rule-based grammars to principle-based ones, may not be possible because there may be no commonalities between the rules, i.e. no generalisations about these rules that could be expressed by prin-ciples. Even if there are commonalities, their detection is beyond the capabilities of today’s grammar learning systems.

In the following sections, we will use examples fromgband hpsgto illustrate the transformation of a principle-based grammar into a rule-based grammar by means of partial deduction techniques.

Since pure principle-based and pure rule-based grammars are not strict oppo-sites, but rather the endpoints on a continuum, we will investigate what practical gain in terms of runtime efficiency can be achieved by transforming a grammar that is closer to the principle-based end of the scale into one that is near the rule-based end. As the starting point for this investigation, we use a grammar ofhpsg that was originally encoded in thealeformalism.