Using default inheritance to describe LTAG

We present the results of an investigation into how the set of elementary trees of a Lexicalized Tree Adjoining Grammar can be represented in the lexical knowledge representation language DATR (Evans & Gazdar 1989a,b). The LTAG under consideration is based on the one described in Abeille et al. (1990). Our approach is similar to that of Vijay-Shanker & Schabes (1992) in that we formulate an inheritance hierarchy that efficiently encodes the elementary trees. However, rather than creating a new representation formalism for this task, we employ techniques of established utility in other lexically-oriented frameworks. In particular, we show how DATR's default mechanism can be used to eliminate the need for a non-immediate dominance relation in the descriptions of the surface LTAG entries. This allows us to embed the tree structures in the feature theory in a manner reminiscent of HPSG subcategorisation frames, and hence express lexical rules as relations over feature structures. Vijay-Shanker & Schabes (1992) have drawn attention to the considerable redundancy inherent in LTAG lexicons that are expressed in a flat manner with no sharing of structure or properties across the elementary trees 1. In addition to theoretical considerations, one practical consequence of such redundancy is that maintenance of the lexicon becomes very difficult. One way to minimise redundancy is to adopt a hierarchical lexicon structure with inheritance and lexical rules. Vijay-Shanker & Schabes outline such a view of an LTAG lexicon which is loosely based on that of Flickinger (1987) but tailored for LTAG trees rather than HPSG subcategorization lists. We share their perception of the problem and agree that adopting a hierarchical approach provides the best available solution to it. However, we see no need for the creation of a hierarchical lexical formalism that is specific to the LTAG problem. The use of hierarchical lexicons to reduce or eliminate lexical redundancy is now a fairly well researched area of NLP (Daelemans & Gazdar 1992; Briscoe et al. 1993) and a variety of formal languages for defining such lexicons already exist. One of the more widely known and used of these languages is DATR (Evans & Gazdar 1989a,b); in this paper we will show how DATR can be used to formulate a compact, hierarchical encoding of an LTAG lexicon. There are three major advantages to using an " off the shelf " lexical knowledge representation language (LKRL) like DATR. The first is that it makes it easier …