Consistent dendrification: Trees from categories

1. lntroduction I shall start by taking a fairly simple Combinatory Categorial Grammar (CCG) of the type developed by Steedrnan over the past decade or so (e.g. Steedman 1996) including rules of functional application, and functional composition. I shall have nothing to say about functional substitution in this paper, and shall assume that there are type-raised categories in the lexicon (e.g. S/(S\NP)). I shal! also assurne, following Steedman, that syntactic symbols such as S, NP, S\NP are in fact abbreviations for feature bundles. From a Phrase Structure Grammar (PSG) perspective, a CCG derivation that uses functional composition, if interpreted as building a structural level cf representation, can give rise to some very strange looking trees containing some very unusual node labels. Whereas certain labels correspond to PSG ones (e.g. VP = S\NP), others do not (e.g. S/NP), Furthermore, because certain analyses require a rule cf composition, such trees and labels will be required. 1f there is anything at all "real" about traditional PSG categories for languages such as English, then on the faceof it, CCG fails to capture them. There is a related point. Ifthese strange categories such as S/NP need to be assembled, then one would expect tl1at some lexical items would require such a category either as an argument or as the result. But, tltere seem to be curiously few such words and possibly no verbs. What we shall do in this paper is examine how CCG categories can correspond to trees (cf. Joshi & Kulick 1996 and Henderson 1992 for otlter approaches). We shall see that interpreting a lexical CCG category as a partial description of a tree using a number of very simple principles will allow a number of "natural" distinctions to fall out without being slipulated. In particular, subjects but not objects will be immediately dominated by the S, different types of "empty" categories will be predicted; and structural differences between raising and control verbs will be observed. lf the lexicon is constrained so that the categories can be interpretedas trees in tlte mannerwe shall describe, and ifduring the course ofa successful derivation such trees can tllen be combined with other trees, then we shall say that the lexicon is constrained by a principle cf "Consistent Dendrification". 176 2. Hypothesising Trees As a start towards interpreting a lexical CCG category (e.g. X\Y) as a partial description of a tree, we shall assume that a category …