Inheritance Lexicon Theory

A number of approaches to lexical theory are emerging in computational linguistics which address these problems and attempt to integrate descriptions in the known lexical problem areas. A central role is played by the inheritance lexicon paradigm, initiated by Flickinger [Flickinger 1987]. Inheritance Lexicon Theory (ILT) has been developed in three main directions, each with slightly different linguistic assumptions and conventions for lexical representation.

HPSG:

In the HPSG lexicon, lexical signs are represented by AVM representations and classified into types, with more specific types inheriting generalisable properties from more general types. Lexical rules project a base lexicon on to a much larger (perhaps infinite) lexicon; the rules cover lexicon extension in morphology (inflections, derivations, compounds), syntax (complex subcategories such as passive), semantics (selectional conditions for disambiguating polysemy).

OOL:

In the Object-Oriented Lexicon (OOL) lexical items are represented as objects (classes and instances) in an object hierarchy, in which objects communicate by message-passing, and methods for handling the messages are defined for each object. More specific objects inherit general methods from more general objects, and therefore methods do not necessarily have to be fully specified for any given object. Object-oriented representations originated as a means of representing one type of semantic network in Artificial Intelligence, generally implemented as functions in LISP, but have resulted in well-known class-oriented programming languages such as SmallTalk, C++ and Java. The OOL concept was introduced by Daelemans [Daelemans 1987].

DATR:

DATR is a lexical knowledge representation language developed by Evans and Gazdar (summarised in [Evans & Gazdar 1996]). In DATR, the basic unit is the node (roughly comparable with the type in the HPSG approach and the object in the OOL approach) organised into a default inheritance hierarchy. Each node in the hierarchy is characterised by a set of attribute-value equations (more precisely, equations pairing attribute paths and values), in which any path may only occur once, and each value evaluates to a sequence of atomic constituent values (possibly null). The constituent values directly specified for particular nodes may be atomic, or inherited from more general nodes. Since a node may therefore inherit values from several other, more general nodes, but only if constrained by a unique attribute, DATR is said to have orthogonal multiple inheritance. Default inheritance means that a value of a given attribute may be specified more than once in an inheritance path, in which case the values at lower (more specific) nodes in the hierarchy override values.