Strategies for Adding Control Information to Declarative Grammars

STRATEGIES FOR ADDING CONTROL INFORMATION TO DECLARATIVE GRAMMARS Hans Uszkoreit University of Saarbrticken and German Research Center for Arlfficial Intelligence (DFKI) W-6600 Saarbriicken 11, FRG uszkoreit@coli.uni-sb.de Abstract Strategies are proposed for combining different kinds of constraints in declarative grammars with a detachable layer of control information. The added control information is the basis for parametrized dynamically controlled linguistic deduction, a form of linguistic processing that permits the implementation of plausible linguistic performance models without giving up the declarative formulation of linguistic competence. The information can be used by the linguistic processor for ordering the sequence in which conjuncts and disjuncts are processed, for mixing depth-first and breadth-first search, for cutting off undesired derivations, and for constraint-relaxation. 1 Introduction Feature term formalisms (FTF) have proven extremely useful for the declarative representation of linguistic knowledge. The family of grammar models that are based on such formalisms include Generalized Phrase Structure Grammar (GPSG) [Gazdar et al. 1985], Lexical Functional Grammar (LFG) [Bresnan 1982], Functional Unification Grammar (bUG) [Kay 1984], Head-Driven Phrase Structure Grammar (I-IPSG) [Pollard and Sag 1988], and Categorial Unification Grammar (CUG) [Karttunen 1986, Uszkoreit 1986, Zeevat et al. 1987]. Research for this paper was carried out in parts at DFKI in the project DIsco which is funded by the German Ministry for Research and Technology under Grant-No.: 1TW 9002. Partial funding was also provided by the German Research Association (DFG) through the Project BiLD in the SFB 314: Artificial Intelligence and Knowledge-Based Systems. For fruitful discussions we would like to thank our colleagues in the projects DISCO, BiLD and LIIX)G as well as members of audiences at Austin, Texas, and Kyoto, Japan, where preliminary versions were presented. Special thanks for valuable comment and suggestions go to Gregor Erbach, Stanley Peters, Jim Talley, and Gertjan van Noord. The expressive means of feature term formalisms have enabled linguists to design schemes for a very uniform encoding of universal and language-particular linguistic principles. The most radical approach of organizing linguistic knowledge in a uniform way that was inspired by proposals of Kay can be found in HPSG. Unification grammar formalisms, or constraint-based grammar formalisms as they are sometimes called currently constitute the preferred paradigm for grammatical processing in computational linguistics. One important reason for the success of unification grammars I in computational linguistics is their purely declarative nature. Since these grammars are not committed to any particular processing model, they can be used in combination with a number of processing strategies and algorithms. The modularity has a number of advantages: • freedom for experimentation with different processing schemes, • compatibility of the grammar with improved system versions, • use of the same grammar for analysis and generation, • reusability of a grammar in different systems. Unification grammars have been used by theoretical linguists for describing linguistic competence. There exist no processing models for unification grammars yet that incorporate at least a few of the most widely accepted observations about human linguistic performance. • Robustness: Human listeners can easily parse illformed input and adapt to …

Content truncated to 500 words