morphosyntactic feature geometries

[Susana Bejar]

Nice discussion. I thought I should pipe up, since my thesis is in fact an
exploration of some of these very topics. In particular, I explored
consequences for the theory of agreement if we assume a Harley+Ritter type
system for formal features in syntax. I tried to show that feature
hierarchies (which I see in terms of entailment relations) play a crucial
role in agreement systems and can explain some otherwise confounding
complex agreement patterns. I argued that conditions on agreement (i.e.
whether Match and Value succeed or fail)  are stated explicitly in terms
of whether or not an entailment relation exists between features of the
target and those of the controller.  A consequence of this view is that
the geometry turns out to be crucial to the determination of locality,
(anti)intervention, etc. in agreement systems. This turns out to be
extremely useful in accounting for languages where NPs exhibit
phi-dependent asymmetries with respect to whether or not they can control
agreement.

To continue with the going example, if an agr-head has a [group] probe
then the search for a goal can only be halted by an element in its domain
that is itself specified for some feature that entails [group]. In a
language with a plural/dual contrast, this means that both plural and dual
NPs in the domain of the probe will Match and halt it (see (1) and (2)).
However, a singular NP will not halt it (3), since no feature of a
singular NP entails [group]; let's assume it has only and [individuation]
feature.  (Note that I'm assuming a particular view of the relation
between [group] and [minimal], namely that [minimal] is a dependent of
[group], i.e.  [minimal] entails [group]. This is a controversial
assumption, but see the Cowper paper mentioned by Heidi for good arguments
in support of this).

(1)	----------------
	AGR		NP1		NP2
	[group]		[group]		...

(2)     ----------------
        AGR             NP1             NP2
        [group]         [minimal]        ...

(3)     ---------------------------------
        AGR             NP1             NP2
        [group]         [individ]         ...


As far as where the geometry lives, I haven't come across reasons not to
assume the simplest possible scenario: both syntactic and morphological
objects are built from the same inventory of features constructed by the
learner so terminal nodes and vocabulary items alike are built from
the same features, some of which stand in intrinsic
hierarchical/geometric/semantic relations to one another.

I think we have to be careful not to coflate separate issues when thinking
about feature systems. One question is privativity, another question is
whether something like a geometry is a valid way to represent intrinsic
relations between features. Likewise, with respect to underspecification,
we need to distinguish logical underspecification (which is what the H&R
system employs) vs. the kind of underspecification DM uses when assigning
features to vocabulary items. The latter is not logical
underspecification, but rather something extrinsically imposed in order to
derive the observed blocking patterns between competing vocabulary items.

In principle, I don't think having negative features precludes having
geometric relations between features. So if we did have negative values
in syntax, we could still have a (non-privative) feature geometry. The
negative values would simply be another way of node labelling:

(4)	+Individ
        /       \
    +group      (-group)
    /   \
 +min   (-min)


This does, however, preclude logical underspecification, so the question
is whether or not this is a desirable outcome. As far as the hypotheses
with respect to agreement that I outlined above, this would be a bad
outcome. Locality patterns in syntactic agreement systems suggest to me
that default values (or - values) are not specified in syntax. If they
were we wouldn't expect any phi-asymmetries in the kinds of NPs that can
control agreement. So we wouldn't expect the existence of languages where
plural and dual NPs can match a probe, but singular NPs cannot, since the
singular NPs would be specified as [-group] and therefore should
technicaly be able to match a probe for [group].

(5)	----------------
	AGR		NP1		NP2
	[group]		[-group]	...

(6)	----------------
	AGR		NP1		NP2
	[group]		[+group]	...

Under this view, if the default specification is necessary at all, it must
be supplied later.

In her last posting, Martha considers this question with respect to the
example of a conjoined subject like "Rolf and Martha". If I understand
correctly, I think Martha's suggestion is that If a verb in this context
needs to be marked as dual, not plural, then we need to assume negative
features in th syntax.

plural morphology <-> [+group, -minimal]
dual morphology <-> [+group, +minimal]

The implication, I think, is that without the possibility of a [-minimal]
specification, nothing will prevent the incorrect use of a plural marker
rather than a dual marker. I'm not sure I understand why this is a
necessary conclusion. Couldn't we just assume that vocabulary competition
ensures the correct outcome?

(5)

 Dual AGR node	  matching vocabulary items 	winning VI

[group], [min]	  plural <-> [group]
		  dual <-> [group, min]		dual <-> [group,min]


(6)

Plural AGR node    matching vocabulary items     winning VI

[group]		    plural <-> [group]		plural <-> [group]



Perhaps I'm missing the point! I'd welcome any clarification!

Best,

Susana Bejar