Ivan A. Sag (sag@csli.stanford.edu)
Mon, 27 Jan 97 12:01:04 PST
I think the issues are getting a little clearer now in the adverb extraction discussion. At least I think I see more clearly what is NOT at stake. Bob writes: >What we are in effect forced to take on with the particular LR >approach to adjunct at issue is the claim that the relation among the >tokens of _reads_ is the same in all three examples in (3): > >(3)a. Robin reads. > b. Robin reads slowly. > c. The book reads slowly. > >That is, there are three different, independent lexical entries. I agree with Bob that this consequence (requiring the examples in (3) to be treated in terms of three distinct lexical items) would require independent justification of a sort that has to this point not been provided. But I do not agree that the adverb-as-complement analysis has such lexical proliferation as a consequence. In particular, as Gertjan van Noord and Gosse Bouma showed several years ago in their paper `Adjuncts and the Processing of Lexical Rules', adverb addition lexical rules and the like need not be thought of as rules mapping lexical items into new lexical items. Rather, they can be formulated as recursive constraints that hold over a single lexical item. The basic idea is something like this (I'm taking a terminological or analytic liberty or two): B is a word if A is a lexeme and A and B stand in the add_adj relation: (1) word(B) :- lex(A), add_adj(A,B)... . The add_adj relation holds between two (verbal) signs just in case its 4-place counterpart relation holds between the ARG-ST lists of the two signs and the SEM value of the two signs: |sign | , |sign | (2) add_adj( |SEM X| |SEM Y| ) :- add_adj(L,M,X,Y). |A-S L| |A-S M| Further clauses of the definition go as follows (`<C|D>' designates a list whose first member is C, and whose remaining elements form the list D): (3) add_adj( < >, < >, A, A). (4) add_adj( <C|D>, <C|E>, A, B) :- add_adj(D,E,A,B). |adv | (5) add_adj(A, < |MOD |SEM X[ARG Y]|| | D > , Y, Z) :- add_adj(A,D,X,Z). So given the lexeme in (6a), we have the word in (6b) (more conditions could be added to (1) above, e.g. constraints determining the appropriate inflectional morphology.): |lexeme | |word | (6)a. |PHON ... | b. |PHON .. | |SEM S = love(1,2) | |SEM S' | |ARG-ST A = <NP-1,NP-2>| |ARG-ST A' | where add_adj(A,A',S,S') (6b) is a single lexical entry -- it is just a collection of constraints describing a certain set of feature structures of the type word. Now consider the definition of add_adj. If you take clauses (2)-(4) together, you get the effect of zero adverb additions. (7) is one of the ways of adding information consistent with (6b), just as combining with a 3rd sing feminine subject and 2nd person plur object would be a way of adding information consistent with (6b). |word | (7) |PHON .. | |SEM S = love(i,j) | |ARG-ST A = <NP-i,NP-j>| But there are other ways of satisfying (6b), namely by using clause (5). According to that clause, (8) is another way of satisfying (6b): |word | (8) |PHON .. | |SEM [1][ARG [love(i,j)]] | |ARG-ST <NP-i, NP-j, ADV >| | [SEM [1]] | In fact, there are infinitely many ways of satisfying add_adj, each of which involves adding an ADV to the ARG-ST. Note that (5) treats the recursive embedding of the adverb semantics just right. So, if we adopt a further constraint, call it Argument Conservation (I've rolled in the independent constraint that the SUBJ is a list containing just one NP).: (9) Argument Conservation: word: |SUBJ [1]<NP> | |COMPS [2] | |ARG-ST [1] (+) [2]| then any word described by (8) must also be described by (10): |word | (10) |PHON .. | |SEM [1][ARG [love(i,j)]] | |ARG-ST <NP-i, NP-j, ADV >| | [SEM [1]] | |SUBJ <NP-i> | |COMPS <NP-j, ADV > | | [SEM [1]] | Hence such a word will have to complement-select (an object and) an adverb that will provide the verb with an appropriate adverb-relation as its semantics. Note that all of this happens without multiplying words. If the verb in question combines with an adverb as complement, then the constraints will have to be satisfied in a way different from what would happen in the event that no adverb complement is present. So, as Gertjan and Gosse have shown, lexical rule analyses do not need to be cashed out by multiplying lexical entries. The adverb-as-complement analysis can treat (3a) and (3b) in terms of the same lexical entry. (3c) is perhaps a different lexical item, as Bob suggests; but perhaps not -- it may be possible to explain its further restrictions in terms of similar constraints -- I'm not sure. But I am sure that the adverb-as-complement analysis I have just sketched is exactly in accord with Bob's intuition. No gratuitous lexical items need be assumed. Exactly the same is true of the Complement Extraction Lexical Rule (CELR). Especially if, as Gosse has pointed out, one adopts the `lexical SLASH amalgamation' view I have been developing (`U' here is familiar set union; disjoint set union provides an empirically distinguishable hypothesis about SLASH amalgamation.): (11) LEXICAL SLASH AMALGAMATION: word: |ARG-ST <[SLASH [1]],...[SLASH [n]]> | |SLASH [1] U ... U [n] | Assuming (11), the effect of the CELR is achieved by the following revised formulation of Argument Conservation (`O' here is Reape's sequence union or `shuffle'; I am ignoring SLASH-binders here, e.g. tough.): (9)' Argument Conservation (revised): word: |SUBJ [1]<NP> | |COMPS [2] | |ARG-ST [1] (+) ([2] O <gap>)| (9)' simply allows one of the members of the argument structure list that would otherwise be realized as a complement to be absent from the COMPS list, as long as it is of type gap. But gaps are all specified as in (12) (12) gap: |LOCAL [1] | |SLASH {[1]}| Hence if an ARG-ST member is of type gap, it follows that the verb in question will have a SLASH element corresponding to the (LOCAL value of) the gap. That is (13) and (14) are both valid ways of satisfying a grammar that has both (6b), (11) and (9)': |word | (13) |SEM [love(i,j)] | |SUBJ <NP-i> | |COMPS <NP-j> | |ARG-ST < NP-i , NP-j, >| | [SLASH { }] [SLASH { }] | |SLASH { } | |word | (14) |SEM [love(i,j)] | |SUBJ <NP-i> | |COMPS < > | |ARG-ST < NP-i , NP-j, >| | [SLASH { }] |gap | | | |LOCAL [3] | | | |SLASH {[3]}| | |SLASH {[3]} | Further, note that SLASH Amalgamation, the revised Argument Conservation Principle in (9)' and the adverb analysis sketched above all interact to allow adverb extraction as a third way of satisfying the constraints of the grammar: |word | (15) |SEM [4][ARG [love(i,j)]] | |SUBJ <NP-i> | |COMPS <NP-j > | |ARG-ST < NP-i , NP-j, , ADV > | | [SLASH { }] [SLASH { } ] |gap | | | |LOCAL [3] | | | |SLASH {[3]}| | | |SEM [4] | | |SLASH {[3]} | Again, this is just another way of satisfying (6b), given the interacting constraints outlined above. Again, by interpreting the Adverb Addition Lexical Rule and the Complement Extraction Lexical Rule as constraints on words, rather than as procedures that multiply lexical entries, we are led to a unified, lexical analysis of complement and adverb extraction. Again, I believe that the analysis I have just sketched is exactly in accord with Bob's intuition. No gratuitous lexical items need be assumed. And finally, to address some of the issues Bob, Paul Kay and Dick Hudson bring up, there is no particular reason, as far as I can see, why adjunct selection must involve the same selection feature as complements. As suggested (but badly executed, I must say) in Pollard and Sag (1987), one could have a separate ADJUNCTS feature, giving us lexical entries that would obey (9)'': (9)'' Argument Conservation (revised, extended [:-)]): word: |SUBJ [1] | |COMPS [2] | |ADJUNCTS [3] | |ARG-ST [1] (+) [([2] (+) [3]) O <gap>] | This would enable one to distinguish ordinary adverbs (ADJUNCTS) from those that are semantic arguments (COMPS, e.g. word the letter BADLY; cf. *word the letter). Perhaps something more has to be said to deal with French pas correctly on this view of things.... Reactions? Ivan
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:36:03 PST