How to solve some failures of LTAG.pdfVIP

How to solve some failures of LTAG Sylvain KAHANE LaTTiCe – TALaNa (Université Paris 7) sk@ccr.jussieu.fr Abstract. The paper presents a lexicalized dependency grammar which solves some failures of Lexicalized TAGs, such as the combinatorial explosion of the number of elementary trees and the non adequacy for the analysis of some constructions. Introduction Wide coverage grammars for natural languages have been developed in Lexicalized TAG (cf. Abeillé 1991, Candito 1999 for French and Paroubek et al. 1992, XTAG 1995 for English). These implementations have brought to the fore some failures of the formalism for natural language description which cannot be solved without adopting a descriptively more powerful formalism. These failures concern most of lexicalized grammars, including Categorial Grammars (CG). In this paper, we will present some of these failures and propose solutions in a lexicalized dependency grammar based on Nasr 1995, 1996. 1. Lexicalized grammars An LTAG is a particular case of lexicalized grammar (LG). A LG is a formal grammar that has the form of a lexicon: each lexical unit is associated to a set of elementary structures. The grammar has an operation of combination 1 and each sentence (= a string of word) can be associated to set of structures obtained by combinations of elementary structures associated to the words of the sentence. Formally, a LG is a 5-uple G = L, S, S F , ?, c where: - L is the lexicon; - S is the set of structures; it is an infinite set but it must be finitely defined; - S F is the subset of S of final structures; - ? is a many-to-many map from L to S; - c is the operation of combination of structures; it is a many-to-many map from S×S to S. 2 Below c(α,β) will be noted α.β. The operation c induces an operation c* from S* to S which associates to a sequence of structures of S all the structures of S obtained by combination of these structures. For instance, c*(α,β,γ) is all the structures obtained by the combinations (α