Inference in DATR

Inference in DATR Roger Evans & Gerald Gazdar School of Cognitive and Computing Sciences University of Sussex, BRIGHTON BN1 9QN Abstract DATR is a declarative language for representing a restricted class of inheritance networks, permit- ting both multiple and default inheritance. The principal intended area of application is the representation of lexical entries for natural language processing, and we use examples from this domain throughout. In this paper we present the syntax and inference mechanisms for the language. The goal of the DATR enterprise is the design of a simple language that (i) has the necessary expressive power to encode the lexical entries presupposed by contemporary work in the unification grammar tradition, (ii) can express all the evident generalizations about such entries, (iii) has an explicit theory of infer- ence, (iv) is computationally tractable, and (v) has an explicit declarative semantics. The present paper is primarily concerned with (iii), though the examples used may hint at our stra- tegy in respect of (i) and (ii). 1 Introduction Inheritance networks ("semantic nets") provide an intuitively appealing way of thinking about the representation of various kinds of knowledge. This fact has not gone unnoticed by a number of researchers working on lexical knowledge representation, e.g. de Smedt (1984), Flickinger et al. (1985), Calder & te Linden (1987), Daelemans (1987a,1987b), Gazdar (1987) and Calder (1989). However, many such networks have been realized in the context of programming systems or programming languages that leave their precise meaning unclear. In the light of Brachman (1985), Ether- ington (1988) and much other recent work, it has become apparent that the formal properties of notations intended to represent inheritance are highly problematic. Although not discussed here, DATR has a formal semantics (Evans & Gazdar 1989) for which some completeness and soundness results have been derived. These results, and others (on complexity, for example) will be provided in a subsequent paper. There ate several prototype computational implementa- tions of the language, and non-trivial lexicon fragments for English, German and Latin have been developed and tested. 2 Syntax The syntax of DATR, especially the use of value- terminated attribute trees to encode information, derives from OATR (Shieber 1986). The language consists of strings of symbols drawn from the set SYM - {:, ",., -, --, <, >, (,)} and the sets ATOM and NODE, all of which are disjoint. A string is in OATR, (with respect to given sets ATOM of [atom]s and NODE of [node]s) iff it is a [sentence] as defined by the following set of rules: [sentence] ::= Inode]:[path] == [lvalue]. { [nodel:[path] = [value]. [lvaluel ::= [latoml I ( [lseq] ) [gvalue] ::= [gatom] ]l ( [gseql ) [value] ::= [atoml ( [seql ) [latom] ::= [desc] I [gatom] [gatom] ::= "[desc]" I [atom] - 66 - [desc] ::= ~node] I [lpath] / [node]:[lpath] Bseq] ::= [gseq], I [lseq] [desc] tlseq] [gseq] ::= [seal, ] I [gseq] "[desc]" [gseq] [seq] ::= el [value] [seq] [lpath] ::-- < [laseq] > [path] ::= < [aseq] > [laseq] …

Content truncated to 500 words