Carl Pollard (pollard@ling.ohio-state.edu)
Wed, 29 Jan 97 20:14:53 EST
Except for PHONOLOGY, CONTENT, and CONTEXT, a case can be made for every feature being a junk slot, so you can always move stuff out of lexical entries (or phrasal types) into the relation definitions without changing the set of maximally specific sound-meaning correspondences. However, one must not lose sight of the fact that there are generalizations to capture having to do with notions of lexical identity, so that in some sense (or perhaps several senses), we need mechanisms for saying that two word instances are instances of the same word. Whatever mechanism is used to do this, it CANNOT be simply that both instances satisfy the same description out of some large disjunction of descriptions called the lexicon, unless those disjuncts themselves were crafted to reflect some linguistic notion of word identity. Here is an example. If we accept Bratt/Abeille/Godard/Miller/Sag's (did I leave anyone out) analyses of French tense auxiliaries and causatives, then there are flat (argument-attraction/ argument-composition) analyses for sentences like 1. J'ai mange'. 2. J'ai lave' le chien. 3. Nous ferons manger Jean. 4. Nous ferons laver le chien a` Marie. Now consider: 5. J'ai mange' et lave' le chien. 6. *Nous ferons manger Jean et laver le chien a` Marie. and suppose we wish to provide an account along the lines that the AVOIR that takes an intransitive V is "the same AVOIR" as the one that takes a transitive V, but the FAIRE that takes an intransitive V is "a different FAIRE" than the one that takes a transitive V. (I am not necessarily advocating such an account, I am just using it to make a metatheoretical point.) Now by way of providing such an account it will NOT suffice to set up the lexicon (conceived as a big disjunction of descriptions) in such a way that the instances of AI in (2) and (3) both satisfy the same disjunct but the instances of FERONS in (3) and (4) satisfy distinct disjuncts. This is because "satisfying such-and-such a disjunct" is not a linguistic property; there is no way to write a constraint that says: "in order for two word instances to cooccur in construction X, they must satisfy the same lexical description" (short of coding up the integers as feature structures and then referring to the Goedel numbers of the lexical descriptons in the constraint!). What it comes down to is that in order to provide a linguistic notion of word identity in an HPSG, word objects must contain at least one path whose value uniquely identifies the word of which that word instance is an instance. How the lexicon is encoded as a disjunction is irrelevant (indeed, with enough stuff pulled into the relation definitions, there won't even BE any disjuncts). Carl
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:36:03 PST