ARG-ST on phrases (long)
Ivan A. Sag
sag at csli.stanford.edu
Sun Jan 21 01:18:17 UTC 2001
Hi Everybody,
It's really nice to see such a lively discussion on this list! Let
me add my 2 cents worth.
1. I think it's important to keep in mind that there is something empirical
to be accounted for here (or at least I've always thought there was). I believe
that there are no human languages where a verb V selects for the category of,
semantically selects for, assigns case to, assigns a semantic role to, or
agrees with an element A in a configuration like (1):
(1) S
/ \
VP
/ \
V S
/ \
NP VP
/ \
V' A
Maybe someone wants to argue with this on empirical grounds, for
example, about the accessibility of the subject NP within the
complement S in (1). I could imagine that there are controversial
cases. But I think there is a universal property of human languages
--- selectional locality --- that a linguist should want to present an
account of.
Of course there's tough adjectives that select for a gappy complement and
the gap may be an unbounded distance away, but the whole SLASH analysis
encodes this kind of information in a way that lets it be locally selected.
More on this below.
2. HPSG-2 made a move that tried to provide THE BEGINNINGS OF an account of
selectional locality. By treating SUBCAT values as lists of synsem objects
(not a list of signs, as it was in HPSG-1) and stating the Subcategorization
Principle as a universal principle requiring structure sharing of the
appropriate members of the head daughter's SUBCAT list with the SYNSEM values
of the selected complements, the specific version of UG embodied in HPSG-2
made much information about the complements inaccessible -- unselectable.
As a linguist, I found this very satisfying --- a step in the right direction;
a step toward accounting for (I won't say explaining) selectional locality
in terms of a proposed feature geometry and a particular constraint.
3. So somehow, about once or twice a year (every year), someone comes
up with some problem in some language where it looks like you could
solve things just fine if only SUBCAT (or ARG-ST) were a head feature,
or propagated up to the S-level by some other principled means. To me,
this looks like moving backwards, away from the goal of accounting for
selectional locality. If ARG-ST is a HEAD feature, then of course
essentially the whole (depending on what specific assumptions are made
about adjuncts) substructure of a complement becomes `visible' for
selection, e.g.:
(2) S
[AS <[4],[3]>]
/ \
/ \
[4] VP
[AS <[4],[3]>]
/ \
/ \
V [3]S
[AS <[4],[3]>] [AS <[1],[2]>]
/ \
[1] VP
[AS <[1],[2]>]
/ \
/ [2]
V'
[AS <[1],[2]>]
V can select for an S complement whose VP contains an ablative object:
(3) [AS <NP, S >]
[AS <NP,NP[abl]>]
Of course someone might try to tell another story about why there aren't
verbs like this, but I thought the Subcategorization Principle of HPSG-2
was at least a reasonable attempt...
4. I think it's worth trying to have an account of selectional
locality, and so I think it's worth looking for alternatives to the
SUBCAT/ARG-ST as HEAD feature analyses . Carl brings up the fact that
the HPSG-2 SUBCAT theory doesn't circumscribe the selectable CONTENT
information. I agree. But I wouldn't take that as a reason for
abandoning the HPSG-2 approach to selectional locality. Rather I would
seek an alternative account of CONTENT that allows us to circumscribe
the semantic information that is selectable. Moreover, it looks to me
that the prospects for finding such an alternative CONTENT theory are
not so bad. MRS, for example, might fill the bill. If the LISZT (RELS)
attribute is not part of CONTENT, but rather is pulled to the outer
level of the sign, then perhaps the remaining MRS features -- KEY and
TOP (or LTOP), for example --- provide the makings of a semantic
selectional locality hypothesis. This might not be the only option,
but I'd like to know if it holds up.
5. Carl also brings up Abeille-Godard's analysis of tense
auxiliaries. Here I think he's right again, but I would advocate a
different course of action. The problem here is actually a
consequence of the way A&G (2000 -- Syn & Sem vol. 32, ed. by Borsley) develop
their theory of lite elements. They bifurcate phrases into lite and
non-lite phrases and allow certain modified participles and
coordinated participles to be lite phrases. So in an example like (4),
(4) Paul a achete' et lu ton dernier livre
Paul has bought and read your last book
the coordinate participle [achete' et lu] is a lite phrase. But since
the theory of auxiliary selection articulated in Abeille, Godard and
Sag (1998 -- in Complex Preds..., ed. by Hinrichs et al.) says that
the tense auxiliaries select for the ARG-ST value of their complement
(rather than the COMPS value), therefore phrases, at least the lite
ones like [achete' et lu], must have ARG-ST values in order to be
selected by the auxiliary a(voir). Carl, tell me if I've
misconstructed your argument here.
But again, I would prefer to look for a revision in the A&G theory of lite
elements that would allow us to say that only words and lite things have
ARG-ST. The French evidence, as I understand it, shows only that lite
elements would require ARG-ST if auxiliaries select via ARG-ST, as AGS
argue (AGS could also be wrong about that, I hasten to add...).
6. So, I am assuming that, as linguists, we need an account of
selectional locality, though we might have to clarify what exactly the
domain of locality is. Many of the cases that get brought up, for
example, involve embedded subjects. Serbo-Croatian control, as well as
control in more than one Salish language, as I understand it (Tom
Hukari -- help me out here...), involves obligatory coindexing into a
SATURATED controlled compelement clause. That is, the subject of such
a clause has to be coindexed with the controller, but, given HPSG-2 or
HPSG-2+ assumptions, a control constraint, if stated as a constraint
on a verb class, couldn't have access to a verb's semantics and the
subject of its SATURATED complement. The notion of `Distinguished
Argument' that is used in Construction Grammar (see also Sag and
Pollard's (1991 -- in Language) use of EXT-ARG) might provide enough
extra information to deal with most of the cases like this that people
keep discovering. Moreover, if the value of DIST-ARG or EXT-ARG were
an index, then the extra information transmitted up to be locally
selectable would exclude the case of the embedded subject. Maybe
that's right; I don't know, but it would be interesting. If ARG-ST
is a HEAD feature, these interesting questions won't get asked anymore.
7. OK. So maybe I've been deluding myself all these years. I thought
the HFP was part of a hypothesis that certain information selected
at the phrasal level had consequences for the lexical head of the
phrase and that other information would be absent at the phrasal level
and hence unavailable for selection. Likewise, the SLASH-based
treatment of unbounded dependencies, I've always thought, made claims
about what information is transmitted up through such constructions:
`being gappy or not' one of the properties of a complement that is
made available for selection. Actually the prediction is that all the
local information of the gap is available for selection, e.g. case and
content (see the discussion of Hukari and Levine 91 - in
NLLT). And this hypothesis seems to me to have been confirmed in
spades by the discovery of all those languages (Irish, French,
Spanish, Yiddish, Moore, Icelandic, Kikuyu, Chamorro, Palauan, Thompson
River Salish, etc) where the global information about extraction is
locally selectable by an element within the dependency, just as the
SLASH-based analysis predicts.
8. So that's why I remain interested in the locality argument. We
have the beginnings of a theory of selectional locality going
here. Let's work together to refine it and clarify exactly what the
right notion of locality is...
Happy New Year,
Ivan
-------------------------------------------------------------------------
Ivan A. Sag
Professor of Linguistics
Director: Symbolic Systems Program (2000-2001)
Email: sag at csli.stanford.edu
WWW: http://www-csli.stanford.edu/~sag/sag.html
Dept. of Linguistics CSLI - Ventura Hall
Fax: 650-723-5666 Fax: 650-725-2166
Office: MJH 040B Office: Cordura 228
Phone: 650-725-2323 Phone: 650-723-2876
-------------------------------------------------------------------------
Stanford University - Stanford, CA 94305 USA
-------------------------------------------------------------------------
More information about the HPSG-L
mailing list