Content uploaded by John Frampton
Author content
All content in this area was uploaded by John Frampton on Feb 04, 2015
Content may be subject to copyright.
Remarks on \Derivation by Phase":
Feature Valuation, Agreement, and Intervention
by John Frampton, Sam Gutmann, Julie Legate, and Charles Yang
1
October 2000
In \Derivation by Phase" (2000, henceforth DBP), Chomsky revises the feature
system and the mechanisms of agreement and deletion developed in Chomsky (2000,
\Minimalist Inquiries", henceforth MI). Sp ecically, uninterpretable features on a
head nowenter the derivation \unvalued". A head which assigns structural case, for
example, will contain unvalued
-features and nominals will contain an unvalued
Case-feature. The Agree operation incorporates mechanisms for providing values
to unvalued features under appropriate conditions. Later, at the end of a phase,
the newly valued features will be recognized as such and deleted from the syntactic
representation that is the current state of the continuing derivation. These newly-
valued features do persist long enough to be part of the material handed over to
the morphological and phonological components at the end of phases.
We look at the new system in detail, uncover some problems with it, and
suggest ways in which the Agree operation can be modied in order to overcome
these problems.
1. The Problems
Icelandic past participles are inected for case and number. They are not inected
for person, a fact which the DBP theory exploits, as we shall see. Typical glosses
would be:
(1) a. Max(nom) was killed(nom,sg)
b. There was killed(nom,sg) someone(nom)
c. Max expected someone(acc) to be killed(acc,sg)
d. Max expected there to be killed(acc,sg) someone(acc)
The valued/unvalued feature mechanism introduced in DBP was, in part,
motivated by diculties in the MI system in correctly analyzing examples like
(1d). We begin with a detailed examination of the derivation of (1d) in the DBP
system. It is a good illustration of how the system is intended to work. At an early
stage, (2) is built:
(2) [ Prt
Num[ ]
Case[ ]
[ kill someone
Num[sg]
Per[3]
Case[ ]
]]
Only the syntactically relevant features are shown. The nominal
someone
enters the derivation with a full set of
-features (whichwe represent as a person
and number feature), and a (structural) case feature, whichis\unvalued", its
value to be determined when it agrees with a case-assigning head. After
kill
selects
someone
,
kill
is selected by a participial functional head Prt, with unvalued
number and case features. Note that Prt does not have a full set of
-features, but
does (at least in Icelandic) haveanunvalued case feature. A feature is unvalued
when it enters the derivation if and only if it is uninterpretable (on the head that
contains it).
Prt in (2) is in a position to probe for matching features. There are matching
features on
someone
, where \matching" is taken to mean \of the same type",
not necessarily the same with respect to value. Since both Prt and
someone
are \active", which means they contain unvalued features, agreement takes place
between them. The eect of agreementis to value a feature whichisunvalued if
it matches a valued feature. It assumes the value of the valued feature it matches.
Thus, after agreement, (2) becomes:
(3) [ Prt
Num[sg]
Case[ ]
[ kill someone
Num[sg]
Per[3]
Case[ ]
]]
The derivation proceeds, eventually reaching:
(4) [ T
r
Per[ ]
EPP[ ]
[be [ Prt
Num[sg]
Case[ ]
[ kill someone
Num[sg]
Per[3]
Case[ ]
]]]]
Here, T
r
is raising Tense, which is assumed to be defectivein having only a person
feature, as opposed to nite Tense, which has full
-features. The
-features of
Tense, of anyvariety, are initially unvalued. Likeevery Tense, T
r
does contain
an EPP feature, which can only be satised by an extra merge operation|extra
above the required semantic merge(s), which in this instance was merger with VP.
NowT
r
is the probe, and it matches
someone
and apparently could agree with
it. If, however, the numeration contains an expletive
there
, the preference for the
simpler Merge operation, as opposed to the more complex operation Move, dictates
that
there
be merged. The expletiveenters with a single unvalued person feature
which matches and agrees with the current probe T
r
. The EPP feature of T
r
is satised by the merger; the person features (b oth unvalued) remain unvalued.
While
-features are valued by matching, EPP-features are valued simply by
attracting speciers. This gives:
(5) [ there
Per[ ]
T
r
Per[ ]
EPP[
p
]
[be [ Prt
Num[sg]
Case[ ]
[ kill someone
Num[sg]
Per[3]
Case[ ]
]]]]
Wehave used
p
above as the value, and only p ossible value, of the EPP-feature.
Since T
r
still has an unvalued person feature, it continues to probe, and agrees
with
someone
, which can value the person feature of T
r
, assigning it the value 3.
The ECM verb
expect
then next selects T
r
. Next, the functional head v
, which
2
has unvalued
-features, selects
expect
and
Max
.At this point, the representation
is:
(6) [ Max
Per[3]
Num[sg]
Case[ ]
v
Per[ ]
Num[ ]
[ expect [ there
Per[ ]
T
r
Per[ ]
EPP[
p
]
[be [ Prt
Num[sg]
Case[ ]
[ kill someone
Per[3]
Num[sg]
Case[ ]
]]]]]]
Now v* is the probe, and it agrees with
there
, and then with T
r
, the agreement
involving unvalued person features, which remain unvalued. A probe continues
to probe if it contains unvalued features, so v* next agrees with Prt. Only heads
with full sets of
-features can value features on agreeing heads. Thus Prt cannot
value features of v
, but v
can value features of Prt. The only feature of Prt that
v
can value is Case, which becomes Case[acc]. Likeunvalued EPP-features, and
unlike
-features, case features are not valued by matching.
-agreement with a
case-assigning head can value a case feature as a side eect. Of course, v
is an
accusative case assigner. This gives:
(7) [ Max
Per[3]
Num[sg]
Case[ ]
v
Per[ ]
Num[ ]
[ expect [ there
Per[ ]
T
r
Per[3]
EPP[
p
]
[be [ Prt
Num[sg]
Case[acc]
[ kill someone
Per[3]
Num[sg]
Case[ ]
]]]]]]
The head v
still has unvalued features, so it continues to probe. It next agrees
with
someone
, yielding:
(8) [ Max
Per[3]
Num[sg]
Case[ ]
v
Per[3]
Num[sg]
[ expect [ there
Per[ ]
T
r
Per[3]
EPP[
p
]
[be [ Prt
Num[sg]
Case[acc]
[ kill someone
Per[3]
Num[sg]
Case[acc]
]]]]]]
This marks the end of the v
phase. (In this note we use the term \phase" to
refer to v
and C phases, which are actually called \strong phases" in DBP.) The
derivation concludes with merger of the matrix T, which has a full set of unvalued
-features and selects v
, and then agreement with the subject
Max
. The latter
moves to satisfy the EPP of the matrix T. The unvalued
-features of the matrix
T are valued via agreement with
Max
, and the latter's case feature receives a
nominativevalue as a side eect of the agreement.
In the DBP theory, the valuation of a feature whichentered the derivation
without a value plays the role of marking the feature for deletion; but the deletion
will actually only take place at the end of a phase. The derivation must \remember"
until the end of the phase which features havechanged from unvalued to valued.
DBP does not specify any mechanism to accomplish this.
The accountaboveisintended to summarize the mechanics of feature checking
as given in (the revised version of) DBP. A residual problem is that the p erson
feature of
there
is still unvalued in (8). We return in Section 3 to discuss this in
more detail, consider a suggestion of Chomsky, and suggest a possible alternative.
3
Consider now the derivation of (1c), repeated here as (9). A second problem
will appear.
(9) Max expected someone(acc) to be killed(acc,sg)
The derivation is identical to the derivation detailed ab ove up to the point where
T
r
is the probe. At this point, however, there is no
there
in the numeration and
and the EPP feature of T
r
must be satised bymovement. This gives:
(10) [ someone
j
Per[3]
Num[sg]
Case[ ]
T
r
Per[3]
EPP[
p
]
[be [ Prt
Num[sg]
Case[ ]
[ kill
t
j
]]]]
After
expect
selects T
r
, and v
selects
expect
, the relevant structure is:
(11) [ v
Per[ ]
Num[ ]
[ expect [ someone
j
Per[3]
Num[sg]
Case[ ]
T
r
Per[3]
EPP[
p
]
[be [ Prt
Num[sg]
Case[ ]
[ kill
t
j
]]]]]]
Nowv
is the probe. We discuss the Minimum Link Condition in detail below,
but according to the formulation in DBP (and other published versions that we
knowof),v
must agree rst with
someone
. When v* agrees with
someone
, all
features of both heads are valued, exactly as in (8) above. The result is:
(12) [ v
Per[3]
Num[sg]
[ expect [ someone
j
Per[3]
Num[sg]
Case[acc]
T
r
Per[3]
EPP[
p
]
[be [ Prt
Num[sg]
Case[ ]
[ kill
t
j
]]]]]]
At this point, v
has no unvalued features and hence is \inactive"; it stops
probing. The case feature of Prt, which in fact surfaces with morphological
accusative case, remains unvalued in this derivation.
Wenow turn to possible solutions to the two problems whichhave been
uncovered: the unvalued person feature of
there
in (1d) and (8), and the unvalued
case feature of Prt in (1c), (9), and (12).
2. Transitivity
Above, we identied two apparent diculties in the DBP checking system. We
repeat them here.
(13) a. There was killed(nom,sg) someone(nom)
b. Max expected someone(acc) to be killed(acc,sg)
First, in (13a), the person feature on
there
remains unvalued. Although
there
agrees with the matrix T, which has full
-features and is therefore capable of
valuing other
-features under
-agreement, provided that its own
-features are
valued, the agreement with
there
takes place before those
-features are indeed
valued. In fact, the p erson feature of
there
never gets valued in any DBP derivation.
4
Second, in (13b), the case feature of Prt remains unvalued. The problem, on the
face of it, appears to be of a dierent sort, an intervention problem. The raised
nominal
someone
intervenes between Prt and v
, which could otherwise
-agree
with Prt and value its case feature.
In spite of their dierent appearances, both problems would be overcome if
there was some principle of transitivityof
-agreement in force; informally,if
-agrees with
, and
-agrees with
, then
-agrees with
.Ifwe supp ose that
the
-agreement induced by transitivity automatically induces feature valuation
which aligns the values of case and
-features of elements which
-agree, then
the valued person feature of
someone
in (13a) induces valuation of the person
feature of the expletive as a side eect of T/
someone
agreement since T and the
expletivehave previously agreed. In (13b), Prt rst
-agrees with
someone
, then
the latter
-agrees with v
. Under transitivity, the second agreement operation
induces indirect
-agreementbetween v
and Prt, whichvalues the case feature of
Prt.
Incorporating a principle along these lines into the derivational framework re-
quires inventing new mechanisms. One must specify how the derivation remembers
or records the list of pairs of heads that haveentered into the Agree relation. And
one must specify how this record is used at each stage of the derivation to induce
the valuation of features that is implied by transitivity. Exactly how is the person
feature of
there
in (13a), for example, indirectly valued when the person feature of
T (which previously agreed with
there
)isvalued? Frampton and Gutmann (2000)
propose that the mechanism which both records agreement and performs indirect
valuation is \feature coalescence": features that have agreed, whether valued or
not, become
the same
entity.
Belowwe examine two alternate solutions to the problem of valuing the case
feature of Prt in (13b). One capitalizes on Chomsky's notion of \o ccurrences". The
second explores dierent formulations of the Minimal Link Condition. In spite of
the suggestion above that the two problems have a unitary solution, there is some
justication for independent solutions to the two problems wehave identied. It
is not entirely clear that the person feature of the expletivemust be syntactically
valued. The person feature of the expletivemust be valued only for theory-internal
reasons: uninterpretable features must be deleted before they reach the interfaces.
In DBP terms, this means unvalued features must be valued before they reach the
interfaces. But now that the mechanism for deletion is valuation, it seems natural
to explore the possibility that unvalued features cause a derivation to crash if they
reach the morphological interface, as opposed to the LF interface, since unvalued
features presumably cannot be spelled out by the morphology. Since
there
type
expletives are devoid of overt morphology, apparently as a cross-linguistic universal,
it may be that morphology does not even look at the p erson feature of the expletive.
Alternatively, morphology could universally default the value to 3rd p erson.
5
3. Occurrences
Chomsky (p.c.) suggested that the problem of valuing Prt in (13b) could b e
handled by adopting the suggestion at the end of DBP that syntactic movement
might be replaced by the recording of \occurrences". Consider a typical case of
agreement, b etween T and a nominal:
(14) T
Per[ ]
Num[ ]
EPP[ ]
be Prt
Num[sg]
Case[ ]
kill someone
Per[3]
Num[sg]
Case[ ]
T, as always, contains an EPP feature. In the o ccurrences alternativetomovement,
after T agrees with
someone
, the presence of the EPP feature causes T to be marked
as an \occurrence" of
someone
. Some mechanism is required for recording this.
The simplest is to assume that the EPP feature is valued byapointer to the head
which it is the occurrence of. But no movementof
someone
to the specier of
T takes place until the end of a phase, when material is sent to the phonological
component. In the case of successive cyclic nominal raising, then, agreement will
always be with the
tail
of the chain, rather than with the head of the chain.
In the occurrences version of the DBP theory, the stage of the derivation of
(13b) that was previously represented as (11) is instead:
(15) [ v
Per[ ]
Num[ ]
[ expect [ T
r
Per[3]
EPP[
j
]
[be [ Prt
Num[sg]
Case[ ]
[ kill someone
j
Per[3]
Num[sg]
Case[ ]
]]]]]]
Here, T
r
is an occurrence of
someone
. Nowv
agrees sequentially with T
r
and
then Prt, valuing the case feature of Prt as accusative, though no valuation of
the
-features of v
takes place, since neither T
r
nor Prt has full
-features. v
continues to probe, and agrees with
someone
, whichvalues the
-features of v
and
the case feature of
someone
. Occurrences are described at the end of DBP; for the
interaction of occurrences and the order of feature checking, as indicated here, see
Chomsky (forthcoming).
4. Lo cality Conditions on Probing
The Minimal Link Condition is in essence a locality condition on probing. It limits
the \space" in which a probe is permitted to search for a goal. In this section,
wewant to explore the possibility that the locality condition can be formulated in
suchaway as to allowv
/Prt
-agreement in (16), while at the same time blocking
-agreement in the contexts in which agreementmust b e prevented:
(16) [ Max
Per[3]
Num[sg]
Case[ ]
v
Per[ ]
Num[ ]
[ expect [ someone
j
Per[3]
Num[sg]
Case[ ]
T
r
Per[3]
EPP[
p
]
[be [ Prt
Num[sg]
Case[ ]
[ kill
t
j
]]]]]]
The relevant conguration for a discussion of intervention eects is (17), where
Probe c-commands Goal
1
, which c-commands Goal
2
.
6
(17) Probe
:::
Goal
1
:::
Goal
2
:::
The question is: Under what circumstances does the presence of Goal
1
prevent
Probe from agreeing with Goal
2
? As far as we can see, the intervention condi-
tion (18) is sucient to rule out agreement in those instances in which the DBP
analysis requires it to be ruled out, but does allow agreement in (17).
(18) Goal
1
blocks Probe/Goal
2
-agreement if (and only if):
a. Goal
1
has full
-features; and
b. Goal
2
is in a lower phase than Goal
1
.
Under (18), v
/Prt agreement in (16) can take place prior to v
/
someone
agreement
because the two goals are in the same phase.
The key examples in whichintervention is required involve dative experiencers
in Icelandic. Consider (19), for example.
(19) a. *Max expected Jon(dat,sg) to like horses(acc,pl)
b. [ v
Per[ ]
Num[ ]
[ expect [ Jon
j
Per[3]
Num[sg]
Case[dat]
T
r
Per[3]
EPP[
p
]
[
t
j
v
exp
[ like horses
Per[3]
Num[pl]
Case[ ]
]]]]]
If we assume that v
exp
establishes a phase, as Chomsky does,
-agreement with the
object is blocked in (19), as desired.
2
The subject experiencer has full
-features
and is in a higher phase than
horses
. Replacing the DBP formulation of the MLC
by Condition (18) simplies the DBP account of:
(20) there were believed to have been caught several sh
At the point in the derivation that the matrix T becomes the probe, and putting
aside the question of whether or not the person feature of the expletiveisvalued
or not, the relevant heads (with c-command from left to right) are:
(21) T
Per[ ]
Num[ ]
EPP[ ]
:::
Prt
1
Num[pl]
Case[ ]
:::
there
Per
:::
Prt
2
Num[pl]
Case[ ]
:::
sh
Per[3]
Num[pl]
Under Condition (18), neither Prt
1
nor Prt
2
is an obstacle to T/
sh
agreement.
Finally, note that replacing DBP version of the MLC by Condition (18) also
simplies the account of subject raising after ob ject shift. In (22), following the
account of object shift in DBP, the object is assumed to have raised to a second
specier of v
. There is no Condition (18) intervention since the object and sub ject
are in the same phase.
(22) a. Max expected Jon(acc,sg) to us(acc,pl) see (with object shift)
b. [ T
r
Per[ ]
EPP[ ]
[ us
j
Per[1]
Num[pl]
Case[acc]
Jon
Per[3]
Num[sg]
Case[ ]
v
Per[1]
Num[pl]
EPP[
p
]
[saw
t
j
]]]
7
The considerations in this section are tentative, but sucient to make the point
that the problem of valuing the case feature of Prt in (16) can perhaps b e dealt
with by a reconsideration of the locality conditions on probing.
5. Conclusion
One innovation introduced in DBP is that the deletion of features should be
precisely understood as the consequence of the valuation of features that entered
the derivation unvalued. This proposal has clear conceptual strengths. For one, it
means that case features on nominals start out unvalued, and become valued (i.e.
are assigned case) via
-agreement with a case-assigning head. This desideratum
had seemed hard to achieve, because it was not clear how past participles could be
case marked under this assumption. But the MI/DBP system succeeds.
In this note wehave exhibited some related apparentaws in the DBP system,
and discussed some possible solutions, which themselves seem to op en the wayto
further fruitful work on the precise nature of agreement and movement.
Notes
1. Shortly after the rst version of Chomsky's \Derivation by Phase" (DBP) was circulated, the
four authors formed a study group to help us understand the new proposals. This paper tries to
summarize some of our discussion, with the intention of helping others who may also be wrestling
with some of the same questions. Weintend it as a (partial) \reading guide" to DBP, not to be
read independently of that work.
We thank Noam Chomsky for his consideration of the questions we raised and his discussion of
possible solutions.
2. The p oint here is only to show that Condition (18) does all the work done by the formulation
of the MLC given in DBP.We do not explore the question of whether the account given in DBP
of (19) is adequate. This would require a discussion of whether the traces of movement trigger
-intervention. Chomsky assumes that they do, in order to draw the contrast between Icelandic:
(
i
) *there T(pl) me(dat) v
exp
seem p eople(nom,pl) to be in the ro om
(
ii
) me
j
(dat) T(pl)
t
j
v
exp
seem people(nom,pl) to be in the room
The account relies on movementin (
ii
)obviating intervention in (
i
). But if movement can obviate
intervention, an account is needed of why object shift in (19), which apparently is possible, cannot
obviate intervention.
See Frampton and Gutmann (2000) for a dierent approach to the (
i
)/(
ii
) contrast.
References
Chomsky, Noam. 1999. Derivation by Phase. ms., MIT.
Chomsky, Noam. 2000. Minimalist Inquiries: The Framework. In
Step by Step:
Essays on Minimalist Syntax in Honor of Howard Lasnik
, eds. Roger Martin,
David Michaels, and Juan Uriagereka. MIT Press, Cambridge.
Frampton, John and Sam Gutmann. 2000. Agreementis Feature Sharing. ms.,
Northeastern Univesity. (Available at
http://www.math.neu.edu/ling/
.)
8
