A SPARSE APPROACH TO RIGHT NODE RAISING
BRADLEY LARSON
This article proposes a different means of deriving right node raising (RNR) sen-
tences in English. The three current mainstream analyses will first be shown to each
inescapably succumb to some particular sort of data, prompting a new analysis. The
proposed alternative will be argued to succeed in avoiding the pitfalls that the others
run into while at the same time proving easily amenable to the interface concerns
such as linearization and interpretation. This alternative is dubbed a “concatenative"
approach to RNR. Concatenation is the proposed sub-operation of merge proposed
by Hornstein (2009) and Hornstein & Nunes (2008) for adjunct and exploited by
Larson (2010) for coordination. It is applied in this instance to relate the shared
material in RNR to each conjunct without any c-command relation between much
of the first conjunct and the shared material. This provides a way of accounting for
Vehicle Change effects and asymmetrical NPI licensing effects as well providing an
intuitive application of the LCA for purposes of linearization.
I will first investigate the shortcomings of the current analyses and show the short-
comings to be inherent to those approaches. Second I will introduce the decomposed
merge concept proposed by Hornstein and extend it to RNR, showing how it suc-
ceeds where the others fail. Third I propose a synthesis of decomposed merge with
Pesetsky and Torrego’s (2004) feature sharing analysis to derive some structural con-
straints of RNR. Following this I propose a means of interpreting the gap position
1
2 BRADLEY LARSON
in the first conjunct via extra-grammatical inference. Finally I show that this new
approach to RNR is, as opposed to previous accounts, extremely simple to linearize.
1. Introduction
The study of right node raising (RNR) finds itself at an impasse. There are cur-
rently three main avenues of analysis concerning the construction, each of which
is compelling and comes with its advantages. Unfortunately, each is inherently in-
capable of accounting for certain data points. The same features that grant them
certain successes preclude them from extending to recalcitrant cases. Fiddling with
the various approaches will not skirt this. This first section argues that none of the
current accounts can handle all of the basic empirical facts of RNR. It is not simply
particular formulations of these accounts that fall short. In their most fundamental
form and without ad hoc machinery, the accounts simply predict the opposite of
what the facts provide.
The three accounts are Deletion, Movement, and Multidominance, and each enjoys
strong support in the literature. In what follows I will discuss the merits of each as
a prelude to the stubborn facts that are unaccounted for.
1.1. Deletion. Proposed most notably by Wexler and Culicover (1980) (see also
Kayne 1994; Wilder 1997; Hartmann 2000; Ha, 2006; An, 2007; Ince 2009) the
Deletion account holds that an element in the first conjunct is phonologically deleted
under some form of identity with an element in the second conjunct, just like in VP-
deletion. Portrayed graphically, this operation takes a string like (1) and generates
a string like (4).
(1) Ivan bought the short stories and Ivy read the short stories.
A SPARSE APPROACH TO RIGHT NODE RAISING 3
(2) Ivan bought [the short stories] and Ivy read [the short stories].Identity
(3) Ivan bought [the short stories] and Ivy read [the short stories].
Deletion
(4) Ivan bought and Ivy read the short stories.
1.1.1. Advantages. The analysis straightforwardly captures the fact that there always
exists in RNR an element in the second conjunct that is also interpreted in the first
despite not appearing there overtly. It also predicts some surprising facts about
RNR: island insensitivity and vehicle change capability.
Islands are violated when an element subextracts from inside one. Under the
Deletion analysis no movement occurs in deriving RNR, only deletion. We should
thus expect that when an element is right node “raised" (i.e. deleted) out of an island
there is no violation. And this is indeed the case, as shown in the complex NP island
below
(5) John knows a man who sells, and Fred knows a man who repairs, washing
machines.
derived from:
(6) John knows [island a man who sells washing machines] and Fred knows a man
who repairs washing machines.
4 BRADLEY LARSON
In addition to this, the Deletion (qua deletion of the same sort as VP-deletion as is
uniformly assumed) account correctly predicts that certain binding principle viola-
tions can be avoided via vehicle change (Fiengo and May, 1994). Example (7) in its
un-deleted form represents a Principle C violation. If deleted, the offending bindee
R-expression can undergo vehicle change into a more suitable guise. That is, at LF
the R-expression can be represented as its pronominal correlate. The same goes for
the RNR example in (8).
(7) a. *Mary [loves Johni] and hei thinks Sally does [love Johni] tooIdentity
b. *Mary [loves Johni] and hei thinks Sally does [love Johni] too
Deletion
c. XMary loves Johni and hei thinks Sally does [love himi] too.Vehicle change
(8) a. *Hei hopes that Susan won’t [fire Johni], but the secretary knows that she
will [fire Johni]Identity
b. *Hei hopes that Susan won’t [fire Johni], but the secretary knows that she
will [fire Johni]
Deletion
c. XHei hopes that Susan won’t [fire himi], but the secretary knows that
she will fire Johni Vehicle Change
A SPARSE APPROACH TO RIGHT NODE RAISING 5
These are strong predictions, correctly made under a Deletion account, but there
are other data points that such accounts cannot easily handle. In the next subsection
we lay these out.
1.1.2. Disadvantages. 1
Because the Deletion account does not posit any movement to derive RNR, the
account is able to predict that it is island insensitive. But the fact that no overt
movement has taken place also will essentially preclude scope ambiguities from aris-
ing. To see why this is precluded, consider the following. As noted by Boskovic and
Franks (2000), there is no scope ambiguity in (9), only surface scope is available
(9) Some delegate represented every candidate and nominated every candidate.
This contrasts with the monoclausal example in (10) in which there is no obvious
overt movement, yet there is indeed scope ambiguity.
(10) Some delegate represented every candidate.
Boskovic and Franks take the above data point to suggest that there is no across-
the-board covert movement operation. Whereas we can grant a covert Quantifier
Raising operation to get the scope facts in (10), granting the same ability in an
across-the-board fashion in (9) would falsely predict two available scope readings.
There are two universal quantifiers in (9) and only one apparent left-periphery for
1Howard Lasnik (pc) notes a very striking disadvantages to deletions accounts of RNR: deletion, inevery other instance, requires an antecedent preceding the ellipsis site. This seems to be a simplereason to doubt that RNR is derived via deletion.
6 BRADLEY LARSON
them to QR to. If QR were going to happen, then it would need to be across-the-
board (from two positions to one). We can’t get the reading that this would effect
and as such we do not want to posit this covert ATB movement as a possibility.
In more agnostic terms, the data above at least shows that elements that are in
some sense identical in coordinated structures cannot scope high. Taking this to be
the case, the Deletion analysis would predict that RNR sentences also fail to display
scope ambiguity. Though covert at PF, the deleted elements in (11) still exist at LF.
(11) Some policeman arrested [every teenager who was near the crime scene] but
ended up releasing every teenager who was near the crime scene
However, there is indeed scopal ambiguity in (11) despite the fact that at LF, (11)
would have two universally quantified elements just like the unambiguous (10). In
sum, In the case of island sensitivity, Deletion doesn’t involve movement and the
right prediction is made; in the case of scope ambiguity, Deletion doesn’t involve
movement and the wrong prediction is made.
As with the advantages and disadvantages stemming from the lack of movement in
such accounts, so too does the deletion aspect of the account bring with it good and
bad. While phonological deletion predicts the vehicle change facts, it sets the account
up for failure in dealing with relational modifiers. Conjunction reduction accounts
of coordination traditionally cannot handle sentences where reciprocal anaphors are
deleted:
(12) Ivan and Ivy saw each other
cannot be derived from:
A SPARSE APPROACH TO RIGHT NODE RAISING 7
(13) *Ivan saw each other and Ivy saw each other.
Also, the interpretation of (14) differs from its un-reduced form in (15). That is,
in (14) the books that Ivan read are similar to those that Ivy read yet in (15) the
books that Ivan read were all of a type, similar to each other and independent of
Ivy’s books.
(14) Ivan and Ivy read similar books.
(15) Ivan read similar books and Ivy read similar books.
Similar facts are problematic for Deletion accounts of RNR. The sentence in (16)
differs from its un-reduced form, shown in (17). The interpretation of similar below
tracks those of the pair above.
(16) Ivan wrote, and Ivy read, similar books.
(17) Ivan wrote similar books and Ivy read similar books.
Again, the Deletion account presents a double-edged sword. We want deletion
for the vehicle change aspect, but cannot have it for relational modifiers. In turn,
movement would seem to allow the attested scope ambiguities, but movement would
create island violations where they don’t exist. By their very nature, Deletion ac-
counts cannot avoid this tension.
1.2. Movement. The same advantages and disadvantages above can be found in
reverse with Movement accounts (as in Ross, 1967; Postal, 1974; Gazdar, 1981;
Williams, 1981; and Sabbagh 2007, 2008). Shown below, Movement analyses of RNR
8 BRADLEY LARSON
posit a form of across-the-board movement of like constituents to right-preripheral
position.
(18)
XP
book
XP
book
book
Such analyses fail to predict island insensitivity of the sort that we have encoun-
tered above. But they do allow for the scope ambiguity by giving the target for
movement a landing site to take scope from. Also, they do not straightforwardly
account for the vehicle change facts, but allow for the correct readings of relational
modifiers (so long as they scope high).
Movement accounts also fall short empirically with respect to c-command phe-
nomena. One example of such is negative polarity item (NPI) licensing. There must
be a c-command relation between a negative element and an NPI and it must apply
in overt syntax.
(19) No newspapers annoyed anyone
(20) *Any newspapers annoyed no one
(21) *Any newspapers didn’t annoy Ivan
When the targeted element is moved to the right-periphery like in (18), it is no
longer c-commanded by the relevant elements in the clause, but rather c-commands
A SPARSE APPROACH TO RIGHT NODE RAISING 9
them itself. We thus might expect NPIs not to be licensed in such a position when
there is negation in the clause that they moved from.2But this is not the case, as the
sentence in (22) is fine. On the other hand, were the negative element to be raised
to the right periphery, we would expect NPIs left in-situ to be licensed, but (23) is
ungrammatical.3
(22) Ivan didn’t buy and Ivy didn’t read any newspapers.
(23) *Any newspapers annoyed and any books enraged no students
One final disadvantage to this approach is that preposition stranding in RNR is
licit in languages that otherwise do not allow it, say German. In (24) it is clear
that German prohibits movement-derived preposition stranding. In (25) the same
preposition is stranded in the first conjunct, yet the sentence is fine.
(24) *Wem
whomsass
satdie
theKatze
catauf?
onWho did the cat sit on?
2 This differs interestingly from the scope ambiguity case. In the scope ambiguity case, the movedelement could be interpreted in its base position or in its final derived position. Here it seems thatonly the final derived position matters. This seems to be an additional restriction on NPI licensing.Perhaps there is a negative aspect to NPI licensing: a negative element cannot be within in thescope of an NPI at LF.3 It could be the case that NPI licensing requires not only c-command but precedence (as proposedin Klima, 1964) as well. This is not the case. In left-branching languages like Turkish and Japanese,NPIs are licensed by negation that c-commands from the right. In the following example fromTurkish (Ilknur Oded, p.c.) the NPI hic is c-commanded, but not preceded, by the verbal negation
(i) Ben hic elma yemedim.I any apple eat.Neg.Past.1sgI didn’t eat any apples.
To require both c-command and precedence would need to be a quirk of English for the argumentabove not to hold.
10 BRADLEY LARSON
(25) Die
TheKatze
catsass
satauf,
onund
andder
theHund
dogsass
satunter,
underdem
thedicken
fatMann.
manThe cat sat on, and the dog sat under, the fat man.
Furthermore, in English preposition stranding is disallowed with rightward move-
ment, see the extraposition example in (26). But such stranding is fine in English
RNR (27).
(26) *The cat sat on yesterday the fat man.
(27) The cat sat on, and the dog sat near, the fat man.
We again find an approach that falls short empirically due directly to what is its
saving grace with respect to some other construction. Movement gets the scope and
relational modifier facts correct, but in turn gets the island and NPI facts wrong.
There is no apparent escape from this tension.
1.3. Multidominance. The last existing account presented here is Multidominance.
Originally proposed by McCawley (1982) (see also Wilder, 1999; de Vos and Vicente,
2005; Gracanin-Yuksek, 2007; Bachrach and Katzir, 2009; Grosz, 2009; Larson, 2009)
Multidominance approaches argue that the element that is interpreted in both con-
juncts is actually simultaneously, token-identically in both conjuncts and only overtly
realized in the second one.
A SPARSE APPROACH TO RIGHT NODE RAISING 11
(28) CP
TP TP
Ivan VP
bought
Ivy VP
read
DP
the book
This sort of tree is made possible via the operation External Remerge proposed by
Citko (2005) whereby an element X is first externally merged with an element Y (29)
and then externally merged with subsequent element Z (30).4
(29) YP
X Y
(30)
ZP YP
X YZ X
This approach has the obvious advantages of being consistent with the island effects
discussed above. The shared element simply does not move. Also, it is able to account
for the scope facts since any covert movement would not be across-the-board. That
4 Whether this happens in the stepwise fashion as presented here or simultaneously is not importanthere. Though it is interesting to note that if it is stepwise, the operation is difficult to distinguishfrom Sideward Movement
12 BRADLEY LARSON
is, with covert quantifier raising of the shared element there would only be a single
instance of movement from one position, nothing across-the-board.
Multidominance approaches enjoy the added bonus of accounting for the preposi-
tion stranding effects mentioned earlier. In languages that do not allow preposition
stranding, like German, apparent preposition stranding is fine in RNR.
(31) Die
TheKatze
catsass
satauf,
onund
andder
theHund
dogsass
satunter,
underdem
thedicken
fatMann.
manThe cat sat on, and the dog sat under, the fat man.
This is readily explained under a multidominance account. The object of the
preposition dem Tisch is literally “still there" in both conjuncts. It is only pho-
netically realized in one (see Johnson (2007), Wilder, and Bachrach and Katzir for
linearization schemes).
But again here, there are data points that Multidominance cannot handle without
unmotivated stipulation. As mentioned above, NPIs are licit as the target of RNR
sentences and Multidominance approaches handle this fact easily. But the simple
symmetry found in (28) above makes the wrong predictions with respect to sentences
like (32) where only one conjunct contains a negative elements.
(32) a. Ivan bought, but Ivy didn’t read, any books
b. *Ivan didn’t buy, but Ivy read, any books
The multidominance approach does not make differential predictions concerning the
above sentences. The tree above shows the same relation between the conjunct with
the negation as the conjunct without it, regardless of order, and as such should either
A SPARSE APPROACH TO RIGHT NODE RAISING 13
predict the sentences in (32) to both be grammatical since the NPI is licensed in at
least one conjunct.
The Multidominance account does not involve any sort of deletion. There is only
one target token which, on pain of non-recoverability, cannot not be deleted. Given
this, we also do not expect the vehicle change effects evinced by Deletion accounts.
That is, there is no way to keep ‘hei’ from c-commanding ‘Johni’ in sentence (8c.)
repeated here as (33)
(33) Hei hopes that Susan won’t, but the secretary knows that she will fire Johni
In sum, thanks to the local advantages of the multidominance account, it fails to
fully account for the entirety of the facts.
1.4. Summary. We have seen that each approach to RNR, despite enjoying various
advantages, is incapable for principled reasons of handling some core data that the
others can. The power of any given approach only serves to undermine it in the end.
It would be a different matter if all the approaches failed in the face of the same sort
of data: everyone’s problem is nobody’s problem. But this is not the case here.
We can describe their failures in a chain of sorts:
1. Deletion can handle vehicle change and island facts but not scope facts.
2. Movement can handle scope facts but not c-command or island facts.
3. Multidominance can handle island facts but not NPI or vehicle change facts.
14 BRADLEY LARSON
Short of ignoring some of these problems, there is no obvious way of maintaining
any one of these three previous analyses.5 Instead I propose here a fourth way.
2. Decomposed merge
2.1. Introduction. In this section I lay out the concept of decomposed merge as
it applies to adjunction and coordination. In the section following, its coverage is
extended to RNR and is shown to account for the stubborn knot of facts discussed
above.
2.2. Basics. Hornstein (2009) as well as Hornstein & Nunes (2008), argue that the
operation Merge (Chomsky, 1995) should be decomposed into two sub-operations:
Concatenate and Label. Concatenate takes two elements and adjoins them to one
another like in (34). The result is not a targetable constituent in this account, but
rather a flat, string of syntactic elements.
(34) Concatenate XP and YP ⇒
XP YP
This object can undergo the second sub-operation, Label. This operation takes one
of the elements of the Concatenation and uses it to close off the string (that is,
preclude further operations on its internal elements), determine the type of string
that has been closed off, and atomize the erstwhile string into a syntactic constituent,
available to further operations.
5Barros and Vicente (2010) suggest what they deem an eclectic approach to RNR. That is, thereis no single way to derive RNR sentences. Rather, they suggest that Deletion accounts and Mul-tidominance accounts can work in concert to account for the facts. See Larson, in press for coun-terargumenation.
A SPARSE APPROACH TO RIGHT NODE RAISING 15
(35) Label the above object as an X ⇒ XP
X YP
This decomposition allows for an elegant account of the differential behavior of ad-
verbial modification. When an adverb Concatenates with a verb and does not project
(36), the verb+adverb complex is, in Hornstein’s words, “invisible" to the rest of the
structure. So when an operation like VP-deletion targets a VP with a Concatenated
adverb, the VP is deleted to the exclusion of adverb (37).
(36)
VPrun
Aquickly
(37) Ivan ran slowly and Ivy did run quickly.
When an adverb is both Concatenated and Labeled into the structure (38), VP-
deletion applies and deletes the adverb as well (39).
(38) VP
Vrun
Aquickly
(39) Ivan ran quickly and Ivy did run quickly too.
This approach provides an analysis of adjunction in Bare Phrase Structure (BPS)
terms that was not there before (see Hornstein’s discussion for details). Descriptively,
for the verb to be a potential target for deletion it must be a maximal projection.
Since the verb can be targeted for do substitution either inclusive of the adverb
or exclusive of it, two maximal projections are required. Given BPS, two maximal
16 BRADLEY LARSON
projections of the same head is an incoherent notion. To avoid this, the above
decomposition is proposed.
As shown above, adjunction can be wedded to BPS in an elegant fashion. But
more than just adverbs have been argued to be adjuncts. For instance, Munn (1993)
argues that coordination is also an adjunction structure (40).
(40) NP
NP &P
& NP
2.3. Adjunction of Likes. Larson (2010) argues that a similar Hornsteinian tack
must be taken with respect to coordination. The structural ambiguity that Hornstein
posits is also found in coordination. In (41) below, it is possible to target both the
topmost VP for do so substitution and also a lower one. And in (42), it seems that
the anaphor can be bound by either the topmost DP or by a lower one.
(41) a. Ivan [VP [VP ate an apple] and wrote a letter] in the park
b. . . . while Ivy [did so] in the library
c. . . . while Ivy [did so] and read a book in the park
(42) a. Ivan showed [DP the man] and the woman to [himself] and herself in
the mirror
b. Ivan showed [DP the man and the woman] to [themselves] in the mirror
If we only have one XP per projection, as in BPS, then we need some other way
to get these facts. The decomposition of merge discussed in the previous section
A SPARSE APPROACH TO RIGHT NODE RAISING 17
seems to fit the bill. We simply have structural ambiguity with coordination, just
like with traditional adjunction. That is, for example (42a), the coordination has
the structure shown in (43).
(43)
DPthe man
DPthe woman
(44) DP
Dthe man
Dthe woman
In the next section this approach will be applied to RNR and shown to overcome the
obstacles that plague the three other options.
3. Composing RNR
3.1. The Basics. When Hornstein states that the merely concatenated elements are
invisible to the rest of the structure he should not be understood as suggesting that
the individual elements (though as a complex invisible) are invisible to any possible
operation in isolation. There is no reason to suppose that they cannot undergo
subsequent external merge, since they remain maximal projections.
Say an element X concatenated with an element Q but did not label (45). The
Q spine could of course continue growing in the normal fashion (46). Likewise, the
adjoined X element can also undergo further structure building operations (47).
18 BRADLEY LARSON
(45)
Q X
(46) WP
UP W
XW Q
(47)
XP
YP X
X ZP
WP
UP W
W Q
Let us now see how this sort of structure building can be applied to RNR. Our
goal sentence is (48) below.
(48) Ivan bought, and Ivy read, the book.
First the direct object is constructed: “the" merges (concatenates and labels) with
“book"
(49) the
the book
A SPARSE APPROACH TO RIGHT NODE RAISING 19
Then, the verb “read" merges with the product of the pervious merge.
(50) read
read the
the book
In the same way that the DPs adjoin in (43) above, another verb (“bought") con-
catenates with “read" but does not undergo labeling.
(51)
bought read
read the
the book
There is no sense in which “bought" is now more of a root node than “read". Chom-
sky’s (1995) extension condition requires merge to apply to root nodes and as such
both “bought" and “read" are open to further merge operations. It should be noted
that, while a root node, “bought" is not the root of the rest of the tree, just its own
fledgling side-tree.
So both of these verbs can take subjects in their specifier positions and grow just as
normal trees would, the only difference being that at one point they contain elements
that have only been concatenated with each other.
20 BRADLEY LARSON
(52)
T
T read
Ivy read
T
T bought
Ivan bought
read the
the book
The two Ts (or eventual Cs) can be conjoined in the manner below using an &P
following Johannesson (1998).
(53) &
T
&
& T
T read
Ivy read
T bought
Ivan bought
read the
the book
The result is essentially Multidominance without the Dominance. There is no
hierarchically structured relation between the two verbs, just a flat concatenation.
A SPARSE APPROACH TO RIGHT NODE RAISING 21
Under either the Reinhart (1976) definition of c-command (or any really), there is
no command relationship between anything below T in the conjunct to the left and
the one to the right.
This approach can handle each of the otherwise intransigent data points discussed
above as we will see below.
3.2. Advantages. Recall the problems that we found with the previous three analy-
ses in the sections above. There was a list of empirical hurdles that need to be cleared:
1. The shared material can be island-internal
2. The shared material does not c-command from a peripheral position into theconjuncts.
3. The shared material can scope high
4. The shared material shows vehicle change effects with respect to the gap position
5. The shared material is not c-commanded by anything in the first conjunct
It turns out to be the case that none of the analyses in the literature can account
for all of them. Each account fails at one or the other and this was argued not to
stem not from the particulars of each theory but rather from its foundations.
The approach posited in the above section however succeeds in every instance
where the others fail.
3.2.1. Islands. First, the fact that the shared material does not show island effects is
captured in a similar manner to multidominance accounts: the shared material does
22 BRADLEY LARSON
not move. So unlike a movement analysis, this in-situ analysis predicts the lack of
island effects.
3.2.2. C-command. The shared material never moves and as such never moves to a
position from which it c-commands the relevant elements in the sentence. We then
predict the lack of NPI licensing by negation in the shared material.
In short, the problems plaguing movement analyses are not encountered here.
3.2.3. Scope. We have seen that the Deletion analysis cannot handle the scope facts.
It has been noted that there is no across-the-board covert movement, but we are
allowed covert movement in the proposed analysis as there is only one instance of
the shared material. Any QR need not work in an ATB fashion and thus is not easily
ruled out.
3.2.4. Vehicle change and ‘asymmetrical’ C-command. In this analysis the lack of
c-command effects between the first conjunct and the shared material are now ac-
counted for. Unlike the traditional multidominance approach, nothing in the first
conjunct c-commands the shared material. As such we correctly predict that it is
irrelevent to NPIs whether there is negation in the first conjunct. We saw this in
(32) repeated as (54) below.
(54) a. Ivan bought, but Ivy didn’t read, any books
b. *Ivan didn’t buy, but Ivy read, any books
The vehicle change effects are also predicted. There is nothing to bar the co-
indexation of ‘he’ and ‘John’ in (55) as there is no c-command relation between
A SPARSE APPROACH TO RIGHT NODE RAISING 23
them. The R-expression ‘Johni’ is, per the binding principles’ strictures, free. It is
not c-commanded by any co-indexed expression.
(55) Hei hopes that Susan won’t, but the secretary knows that she will fire Johni
If the co-indexed expression finds itself instead in the second conjunct, the sentence
is no longer acceptable (56).
(56) *The secretary hopes that Susan won’t, but hei knows that she will, fire Johni
The proposed lack of c-command between the first conjunct and the shared material
makes a wealth of predictions as we will see in the next subsection.
3.3. (Lack of) C-command Effects. We saw in the last section that the relevant
elements in first conjunct do not seem to c-command the shared material as far as
NPIs and principle C of the binding theory are concerned. In this section further
instances where c-command should play a role are investigated.
3.3.1. Binding Theory. With principle C accounted for, next on the docket is prin-
ciple B. Here a pronoun cannot be bound with a given minimal domain, like in (57).
But when there is no c-command relation, the co-indexation of the two is fine.
(57) a. *Johni likes himi
b. Johni’s mother likes himi
In example (58) adapted from Ha (2006) we find that co-indexation of the pronoun
and the antecedent is licit when the antecedent is in the first conjunct, but not when
it is in the second.
(58) a. Johni couldn’t, and Mary didn’t want to, shave himi.
24 BRADLEY LARSON
b. *Mary didn’t want to, and Johni couldn’t, shave himi.
Lastly, with principle A, we see that an anaphor must be c-commanded by a co-
indexed expression, as seen below.
(59) a. Johni likes himselfi
b. *Maryj likes himselfi
When there is a co-indexed expression in the second conjunct, the anaphor in the
shared material is bound and the sentence is fine, no matter what is in the first
conjunct. But when the co-indexed R-expression in second conjunct is not in a
position to license the reflexive, the sentence is bad.
(60) Ivanj thinks that Maryk likes, but Johni hates, himselfi.
(61) *Ivanj thinks that Maryk likes, but Johni thinks that Ivy hates, himselfi.
Also, when the co-indexed expression is in the first conjunct, it fails to license the
anaphor. Yet of course, flip the conjuncts and the acceptability reverses.
(62) *Johni hates, but Maryj likes, himselfi
(63) Maryj likes, but Johni hates, himselfi
Given the results concerning principle A, we should expect that reciprocals would
not be licit in RNR when the antecedents are shared by the two conjuncts. The
reciprocal anaphor, if it is only c-commanded by a single, singular expression, should
not be licensed. And this is what we find in (64) below.
(64) *Ivani saw, and Ivyj heard, each otheri+j
A SPARSE APPROACH TO RIGHT NODE RAISING 25
When there is a plural expression in the first conjunct and a singular one in the
second conjunct, the reciprocal is also not licensed.
(65) *Ivani and Igorj saw, and Ivyk heard, each otheri+j(+k).
However, if there is a plural expression in the second conjunct, we not only get
an acceptable sentence, more interestingly, we get a particular interpretation. The
sentence below is interpreted as Ivan and Igor seeing each other, Ivy hearing the both
of them, but no reciprocity in either seeing or hearing Ivy by the two males. That
Ivan and Igor can license the reciprocal is expected in this account as is the fact that
Ivy, not c-commanding the reciprocal, can play no part in the reciprocity.
(66) Ivy heard, and Ivan and Igor saw, each other
As will be shown in the an upcoming section, Ivy will still act as a thematic agent to
the reciprocal, which will explain that she can hear the two males, but not be seen
or heard by either of them.
3.3.2. Bound Variables. Another diagnostic of c-command is that of bound variables.
For example, the universally quantified noun phrase ‘every doctor’ c-commands the
pronoun in (67a) and as such the two can co-vary. When the quantifier phrase does
not c-command the pronoun, such covariance is not possible (67b).
(67) a. Every doctor thinks that he is a hero.
b. He thinks that every doctor is a hero.
When the quantified phrase is in the first conjunct of an RNR sentence, we predict
that the pronoun in the shared material will not be able to co-vary, and this is the
26 BRADLEY LARSON
case (68a): there is no-command relation. When the quantified phrase is in the
second conjunct, the bound variable reading is possible.
(68) a. Every doctor thinks, and heart patients hope, that he is a miracle
worker.
b. Heart patients hope, and every doctor thinks, that he is a miracle
worker.
Further in the vein of bound variable readings are crossover phenomenon. Crossover
effects have been categorized into weak- and strong-crossover.6 Strong crossover
effects are found when a certain element crosses over another c-commanding, co-
varying element like in (69) (or similarly (67b). In (69), there is assumed covert
movement of the quantified phrase ‘each boy’ to a left-peripheral position. In moving
to this position it has to cross over the c-commanding pronoun ‘he’ and cannot co-
vary with it.
(69) He likes each boy
Weak crossover effects are likes those in (69) except that the pronoun is further
embedded in the subject and does not c-command the moving element. In the
sentencs below, the prohibition of covariance is not nearly as strong. It is, like the
6 The crossover phenomena discussed here do not include examples of “traditional" strong crossoverlike in (i) below which shows that a wh-phrase cannot move above A c–commanding pronoun andstill co-vary with it.
(i) Who does he like?
To shoehorn the above type of sentence into an RNR-like configuration would result in an ATBwh-question and more on those later.
A SPARSE APPROACH TO RIGHT NODE RAISING 27
term suggests, weak. Sentence (70) can be a statement concerning each mother and
her own personal son.
(70) His mother loves each boy.
The sole difference between these two types of crossover is that there is a c-command
relationship between pronoun and object in (69) and a lack of such a relationship in
(70).
Given our take on RNR, we expect that with our non-c-commanding first conjunct
we should be able to make a weak-crossover effect out of a strong one. That is, take
a strong crossover string and split it across an RNR sentence such that pronoun is
in the first conjunct and the object is in the shared material. When we do this, we
find that the strong crossover sentences receive weak-crossover readings.
(71) possible co-varying reading
He selected, but Sally bought, each boy’s fathers day present.
Again, when the same potentially strong-crossover inducing pronoun is in the
second RNR conjunct, its strength is undiminished7:
(72) Sally bought, but he selected, each boy’s fathers day present.
7 It seems that going through this same procedure with weak-crossover strings creates even weakercrossover effects. That is, in (i) there is the still the slight twinge that weak-crossever exampleselicit. In (ii) I fail to detect any such feeling.
(i) Sally selected, but his mother bought, each boy’s fathers day present.
(ii) His mother selected, but Sally bought, each boy’s fathers day present.
In (i), the weak crossover effect is due to whatever it is that causes them generally: the quantifiedNP moves across the non-c-commanding bound pronoun at LF. In (ii) there is no such cross-overand there is thus no cross-over effect.
28 BRADLEY LARSON
As we have seen, asymmetries seem to be the rule in RNR. In the next section, we
will investigate deeper the engine driving this way of structure building and find
further asymmetries.
4. Featural Engine
4.1. Introduction. At first glance, it is not entirely apparent why the verbs in this
account concatenate with each other. The first conjunct’s verb requires an internal
argument. This internal argument exists, but it is in the second conjunct and it is
unclear how a relation is established between the two. There is no direct structural
relation between the two and the relation between the two verbs is relatively tenuous;
they have only been Concatenated and there is no c-command relation between them.
The first step in demystifying the selectional requirements of the first verb in this
account is to take issue with description of the Concatenation process as producing
a “tenuous" link between the two verbs. Hornstein takes the Concantenation sub-
operation to be the sole locus of featural relationships. Labeling serves to create a
compositional constituent, but it is Concatenation that mediates feature processes,
be they Agreement, valuation, checking, sharing, what-have-you.
4.2. Feature Sharing. The relation between the verbs is then a rather important
one, and will get us quite far in building a relation between object and first conjunct
verb. For the sake of argument, let us suppose that Merge is driven by some sort
of feature. That is, objects have an interpretable feature X and the verb bears an
uninterpretable feature of the same sort to be valued.8
8 It should be noted at the outset that these weakly hypothesized subcategorizing features are ofcourse not the sole features amenable to this analysis. The derivations investigated herein can beachieved via other features as well if one wished to posit them. Furthermore, depending on one’sconception of the grammar these features may not be at all necessary. That is, the features are
A SPARSE APPROACH TO RIGHT NODE RAISING 29
(73)
{see}uXval[_]
{John}iX[_]
In order to transmit the features, they must do as features do: Agree. Under this
analysis, all that is required for Agree would be Concatenation. So Concatenation
applies and Agree occurs as in Pesetsky and Torrego (2004) (PandT), namely the
feature ends up being shared by two positions (here, the ‘1’ represents each instance
of a single feature). Also, the heretofore uninterpretable feature on the verb is made
interpretable9.
(74)
{see}iXval[1]
{John}iX[1]
With the state of the derivation as in (74), the verb will undergo Label. Given
Chomsky’s inclusiveness condition and BPS, the resulting label must retain the fea-
ture it gained in the Concatenation step. It would violate inclusiveness to introduce
another version of the verb that does not have the feature in question. Though
adding the features adds content to the lexical items, one needs to assume this to be
acceptable if there is to be any checking of features at all. They may be affixed to
the lexical items pre-syntactically. The result of Labeling the verb is shown in (75).
posited with a more crash-proof grammar in mind: one that does not produce ill-formed objects(see Frampton and Gutman, 2002). These serve to prevent wild trees from arising. If one prefers agrammar in which wild trees will be filtered out by the grammar, this featural engine section canbe ignored.9 I am going to stray from the traditional notation here for clarity’s sake. When an uninterpretablefeature has Agreed with an interpretable one, I am going to change the u(ninterpretable) markingto an i(nterpretable)-marked feature. In (74) the uXval feature on the verb becomes iXval afterAgreeing with the internal argument.
30 BRADLEY LARSON
(75) {see}iXval[1]
{see}iXval[1]
{John}iX[1]
Pesetsky and Torrego note that the feature on see can “...now serve as the goal for
some later operation of Agree triggered by an unvalued, higher instance of F serving
as a new probe." (PandT page 4). In the case of verbs like see it is normally not
the case that the feature they gained from their internal argument will have any
reason to serve as a goal. As such it will eventually delete according to PandT and
Chomsky (2001). If however, there is another verb in the derivation that can serve
as the probe to this potential goal, there is nothing principled to prevent it.
Say there were another “uXval[_]" verb entered into the derivation: hear as in
(76). It could Concatenate with the Label of our mini tree in (75) and satisfy that
part of its featural requirements like in (77).
(76)
{hear}uXval[_] {see}iXval[1]
{see}iXval[1]
{John}iX[1]
(77)
{hear}iXval[1] {see}iXval[1]
{see}uXval[1]
{John}iX[1]
A SPARSE APPROACH TO RIGHT NODE RAISING 31
According to Pesetsky and Torrego, this will result in a single feature in three loca-
tions. With PandT’s feature sharing analysis we are indirectly able to link up the
verb in the first conjunct with the object in the second. This feat is achieved without
any c-command relationship whatsoever.
It should also be noted that while this approach may appear different in type
or technology than PandT’s account of feature sharing, there has been no added
machinery. In fact, given the lack of novel machinery, it would take a stipulation
to prevent this method of feature transmission from occurring, given the assumed
technology.
4.3. Adjunction Analogies. So we have seen that after one verb Merges (in the
sense of both Concatenate and Label) with its complement, another verb can come
and Concatenate with the result. Why is it not the case that Labeling ensues? It
certainly did when the first verb found its complement. Why not when the second
verb finds its complement equivalent? In this section we will compare this situation
with regular adjunction in PandT’s system.
The Concatenation of verb with verb suggested above is analogous to the situation
PandT describe when distinguishing (un)valued and (un)interpretable features. Take
the example phrase from Latin that they offer:
(78) puella
girl-Nom.Fem.SgRomana
Roman-Nom.Fem.Sg.Roman girls
In their analysis, ‘girls’ enters into the derivation valued for gender (feminine) and
number (plural). The adjective ‘Romana’ enters unvalued for gender and unvalued
for number. The adjective receives its feminine and plural ending from Agreeing
32 BRADLEY LARSON
with the noun, not from any inherent femininity or plurality. As such, the elements
in the derivation, in terms of valuation, begin as in (79).
(79)
{girls}Fem.val[_]
Num.val[_]
{Roman}Fem.[_]
Num.[_]
As for the interpretability of features, ‘girls’ enters the derivation with its gender and
number features interpretable. That is to say in PandT’s words, gender and number
make a particular semantic contribution to that item.10 The number and gender
features on ‘Roman’ do not however contribute to that lexical item’s interpretation
and as such enter the derivation unvalued. We then ultimately have the features
affixed to the lexical items in (80). The relevant features on ‘girls’ are valued and
interpretable, while the features on ‘Roman’ are unvalued and uninterpretable.
(80)
{girls}iFem.val[_]
iNum.val[_]
{Roman}uFem.[_]
uNum.[_]
These two lexical items have matching features (Fem. and Num.) and as such can
Concatenate and Agree, as in (81). Following Hornstein, with adjuncts there is no
projection required and we can leave the structure as is.
10 What the semantic contribution grammatical gender offers is a mystery to me. Maybe inter-pretable features need not always have semantic import.
A SPARSE APPROACH TO RIGHT NODE RAISING 33
(81)
{girls}iFem.val[1]
iNum.val[2]
{Roman}iFem.[1]
iNum.[2]
What would happen if there were no adjective to Agree with the noun? The deriva-
tion would function just fine. The noun had, for the relevant features, everything it
needed. Its features were already valued and interpretable. We want the same to be
the case with our Concatenation of verb with verb. It should not be the case that
this sort of Concatenation of verb to verb is necessary for the derivation to converge.
More often than not, sentences work out just fine without verbs Concatenating with
other verbs.
We thus want our verb-verb Concatenation to work as similarly to the above case
as possible. It should be the case that one verb functions like the noun in having its
relevant features valued and interpretable. The other verb should be able to enter
into the derivation afresh like the adjective.
In the particular case with the verbs, the initial one will be valued (it started out
that way) but its feature will also be interpretable having Agreed with its internal
object. This is to say that the verb’s complement is now sharing its interpretable
feature with the verb. The second verb will enter the derivation just like any verb in
this system: with a feature that is uninterpretable. Thus we have roughly analogous
sets of relative features in the (80) case above as with (82) below: one lexical item has
a feature that is interpretable while the other has a feature that is uninterpretable.
34 BRADLEY LARSON
(82)
{hear}uXval[_] {see}iXval[1]
{see}iXval[1]
{John}iX[1]
This can be generalized. We can now say that optional (adjunct-like) Concatena-
tions happen between elements that are fully ‘satiated’ for a given feature and those
that are not. This restricts adjunction in two important ways. One, it keeps adjuncts
(in this case adjectives and our proposed verb Concatenation) from being crucial to
the success of the derivation and maintains their optionality. Two, it requires that
they adjoin to relevant elements (those that they share features with). This will
preclude a number of wild structures in which irrelevant elements concatenate with
each other.
5. Theta Roles
5.1. Issue. At a glance, it is not entirely apparent how theta roles are to be handled
under this account. For the three other accounts, theta role assignment works by
transparent analogy to any otherwise normal, non-coordinated sentence. In this
concantenative account however, it is unclear how any thematic relation is established
between the verb in the initial conjunct and object in the shared material. There is
no direct structural relation between the two and the relation between the two verbs
is relatively tenuous, they have only been Concatenated and there is no c-command
relation between them.
In this section I will present new argumentation that arguments (internal argu-
ments in particular) are semantically independent of the verb (i.e., full thematic
A SPARSE APPROACH TO RIGHT NODE RAISING 35
separation) and can thus be syntactically independent of it as well. As opposed to
function application (FA) accounts of semantic composition, sentences with miss-
ing arguments are still well-formed in Neo-Davidsonian predicate conjunction (PC)
approaches. Sentences missing arguments can be acceptable so long as the missing
argument can be inferred, this inference depending on the number of event variables.
5.2. RNR Accounts. For Movement and Deletion accounts each verb initially (and
as far as LF is concerned always) has an argument as its complement and as such
thematic role assignment can work in the traditional configuration fashion. With
Multidominance it is superficially more difficult to thematically relate verbs and
arguments. But there is really no important difference here either: each verb has its
complement. They just happen to share the complement.
It is easy to see how thematic roles are assigned in the three other analyses. It is
not so clear with this new approach. One conjunct is missing an argument (or more
than one) in the syntax. What does this mean and how can we account for the theta
assignment?
There is no way that the verb in the first conjunct can structurally “see" the object
in the second conjunct. This verb has access to the features of the second conjunct
verb via its Concatenation with it. The first conjunct verb has not Concatenated
with the shared object, nor is it in any meaningful structural position with respect
to it (like c-command). The shared object is in no position to receive its thematic
role from the first verb.
But it nevertheless does receive a thematic role from that verb. We will investigate
how.
36 BRADLEY LARSON
5.3. Ways of Semantic Composition.
5.3.1. Function Application. One mainstream way of conceiving of semantic compo-
sition is Fregean Function Application (FA) in the sense of Heim and Kratzer (1998)
among many others. In this view verbs are construed as functions, functions that
necessarily take arguments. Nouns (or prepositional phrases or clauses) are those
arguments. For these verbs qua functions to be well-formed for interpretation, they
must be satisfied by arguments in the guise of nouns.
Given a simplistic example sentence like (83), why is it unacceptable according to
FA?
(83) *ran
Under the FA regime the verb run takes the (simplified) form in (84). It is a function
that takes an argument x and returns the proposition that x is a runner.
(84) [[run]] = λx.RUN(x)
In the FA approach, a verb on its own like above is unsaturated and as such cannot
be interpreted as true or false of anything and is in a word ill-formed. In a case
more apropos of our discussion, the sentence in (85) below is unacceptable for the
same reason. Its logical form is given in (86) and as is apparent, the verb is not
entirely satisfied and the result is something that is not formally ready for semantic
interpretation.11
(85) *Ivan hit
(86) [[Ivan hit]] = λyλx.HIT(x,y) x=Ivan
11 I’m abstracting away from the external-argument introducing little v head here.
A SPARSE APPROACH TO RIGHT NODE RAISING 37
This should not distract from the fact that there is indeed some thematic information
in (86). Ivan is the x of the ordered pair x,y and as such is the one who did the
hitting. The hittee however is unknown. That the agent is known is essentially for
naught because the sentence as a whole is going to be ill-formed.
5.3.2. Event Conjunction. Another way of semantic composition is so-called Predi-
cation Conjunction (PC) in the vein of Pietroski (2005) among others. In this view
verbs do not serve as functions nor nouns arguments. Rather, each is a predicate
of an event variable and is formally much more self-sufficient. That is, a verb is
fully-formed and interpretable independent of any noun.
However, a caveat of sorts: this analysis takes a much different route in determining
that (83) is unacceptable. The logical form of (83) is (simplistically) like that in (87).
(87) [[run]] = ∃e{running(e)}
Crucially, (87) is entirely well-formed and nothing is going unsatisfied. The above
straightforwardly denotes a running event. The reason why (83) is unacceptable
is not to be found in the syntax or the semantics but rather the lexico-pragmatic
conceptual weirdness of a running event missing a runner. The same goes for our
Ivan hitting scenario. Its PC logical form looks something like (88).
(88) [[Ivan hit]] = ∃e{hitting(e) & Agent(Ivan, e)}
The above again is formally sound and denotes a hitting event whose agent is Ivan.
The fact that Ivan hit is unacceptable stems from the conceptual oddity of someone
hitting without anything getting hit.
38 BRADLEY LARSON
5.3.3. So What? The fact that there is no such thing as a formally missing argument
in the PC approach is important in addressing the thematic role problem of the RNR
analysis presented here. Under the FA regime, a verb missing an argument is beyond
hope. The sentence will be ungrammatical and there is no way to salvage it. With
PC, the same sentence missing an argument is well-formed semantically and is open
to extra-semantic means of rescue.
I have argued that in RNR there is indeed a missing argument of sorts, syntactic
one, in the first conjunct. The extent to which that analysis is correct is the extent
to which we have an argument for PC as a means of semantic composition and not
FA.
Or vice versa. If PC turns out to be the fundamental means of semantic com-
position, we are not inherently prohibited from investigating RNR along the lines
suggested here.
All told, under an FA approach this RNR analysis would face an asymptotically
steep uphill battle against the formal problems of missing arguments.12 With PC
this is not so and we only need to fight against conceptual oddness. With the rest
of the sentence still to come and all of our human inferential power on our side, this
seems like a more tractable problem. Let’s explore how this might work.
12 I suppose it could be the case that an ungrammatical sentence could nevertheless seemacceptable. There are so-called “grammatical illusions" like that in (i) as noted in Phillips et. al.(2009). The sentence below is ungrammatical yet at first pass, seems fine.
(i) More people have been to Russia than I have.
This however contrasts with RNR sentences in that no matter how much rumination one devotesto RNR sentences, they continue to make sense. Furthermore, there seems to be little speakervariability in judging RNR sentence. This is something one might expect with an ungrammaticalyet not terribly unacceptable sentence.
A SPARSE APPROACH TO RIGHT NODE RAISING 39
5.4. Edo. A test case of this “missing argument" situation in PC-style to semantic
composition can be found in Edo double verb constructions. The sentences, like that
in (89) and analyzed by Baker (1989) as (90), involving null coordination and a null
pro.
(89) Ozo
Ozogha
willle
cookevbare
foodre
eatOzo will cook food and eat it
(90) Ozo will cook foodi and eat proi
The sentence above has some interesting and severe restrictions on its interpretation.
For one, sentences in this mould can only have the interpretation in which the cooking
was done with an eye towards the eating. As Pietroski (2002) puts it, the sentence
must describe a single event that begins with a cooking and ends with an eating.
Second, the food that is to be cooked must be the self-same food that is eaten.
Pietroski develops a PC analysis of how this pro is interpreted given the fact that
the sentence describes only one single event. In short, the pro needs to somehow be
interpreted as necessarily co-indexed with food. But Pietroski wonders how this is to
be done given that there is no c-command relation between the two. Even if there
were c-command between the two, it is unclear how this would be guaranteed.
Take it as a (entirely intuitive) premise that an event can only have a single Theme.
The sentence has one clear Theme: the complement of cook. That complement is
then locked in as the Theme of the sentence and any other argument in a Theme
position, say pro, must be interpreted as the same thing. And this is the reading we
40 BRADLEY LARSON
get. Thanks to the fact that there is only one event described in this sentence, the
pro that is eaten must be the food that is cooked.
However, Baker notes that there is no independent evidence for pro in Edo. It’s a
relatively ad hoc entitiy to posit in this position. Suppose we were to eschew ad hoc
entities. We might see what goes wrong by dropping pro, like in (91).
(91) Ozo will cook foodi and eat
As we have seen, there is going to be nothing formally wrong with the semantics of
(91) under a PC account. We merely have the conceptual oddity of eat not having
a complement. An eating event without anything being eaten.
But the same mechanism that gives pro its interpretation when we had it can
save eat ’s conceptual requirement of a Theme. This sort of inferential mechanism
presumably resides outside semantics proper and its underlying logic will serve as a
guide through the discussion of RNR in the next section.13
5.5. RNR. Much like in the Edo case, I argue that there is a missing argument in
RNR. This time however there is nothing to force the interpretation per se. Unlike
the Edo case, the sentence in (92) need not necessarily begin with cooking and end
in eating.
(92) Ivan cooked, and Ivy ate, a lot of food.
13Not just any sort of inference is possible. I maintain that the inference here relies on the existenceof an LF entity that serves as the basis for the inference. That is, a mentioned or implied objectfrom some previous discourse would no count. If an apple had previously been mentioned or pointedat, it would not exist in the LF of the sentence in question and thus not count a potential basis forthe inference.
A SPARSE APPROACH TO RIGHT NODE RAISING 41
It could be the case that Ivy ate a lot of food on Sunday and Ivan cooked a lot of food
on Monday. It follows that the food need not be the self-same food as in Ozo’s case.
Ivan could have cooked a lot of food completely unaware that Ivy was concurrently
eating a lot of food cross-state. The interpretations of (92) are much freer than in
Ozo’s sentence.
This suggests multiple events and what’s more, a total separation of them. Re-
member, that there was a mere one event in the Edo case allowed for syllogistic
guidance in determining the missing argument’s interpretation: There is one Theme
per event, food is the Theme of this event, eat conceptually requires a Theme, food
is that Theme. Instead, sentence (92) has a PC LF something like that in (93). That
is to say: There was a cooking event with Ivan as its agent and there was an eating
event with Ivy as its agent and a lot of food as its theme.
(93) ∃e∃e’{Agent(e, Ivan) & cooking(e) & Agent(e’, Ivy) & eating(e’) & Theme(e’,
a lot of food)}
With at least two events now, we are now missing that guidance. It is heartening
that the fact that there are multiple events correlates with freer interpretation, but
it means that we are going to have to determine the missing RNR argument in a
different fashion.14
5.6. Intonation. A signature aspect of RNR has been ignored in this essay so far.
RNR sentences universally display a unique, marked intonation pattern. Fery and
14 An interesting sidenote, it seems that if the coordination in a potentially RNR sentence is lowenough, its interpretation mirrors that of the Edo double verb construction, see (94).
(i) Ivan cooked and ate a lot of food.
Though it could be argued that this is mere verb coordination and not an instance of RNR.
42 BRADLEY LARSON
Hartmann (2005) among others explore how RNR is licensed in terms of prosody.
In this subsection I will posit that the particulars of this marked prosody assist in
overcoming the effects of the missing argument in the first conjunct.
Fery and Hartmann note that in a sentence like (94) there are rising accents on
hummed and sang.
(94) Hanna hummed, and Erika sang, a melody
In particular, the rising accent on hummed is characterized by them as “extremely
prominent". They also argue that two accented elements are interpreted as con-
trasting with one another. As we saw in the previous sections, to save our PC-style
account of RNR sentences, we require something to help fill in the missing argument.
In contrasting hummed and sang, one thing that distinguishes them is that the
former is missing an argument and the latter is not. What would happen if we were
to posit that the second conjunct accent picks out what is to follow the first accent?
This would not be to say that the element following sang should be inserted into the
position following hummed in any syntactic or formal sense, but rather as a means to
repair the conceptual weirdness of the first conjunct. There is not much formal to this.
I suggest that this sort of inference is essentially extra-grammatical and a heuristic
way to salvage an otherwise odd sentence. Recall, this odd sentence is universally
cross-lingusitically expected given the operation Concatenation as presented in the
previous sections.
Phillips (1996) notes that it is tempting but disadvantageous to consider RNR a
stylistic quirk or some sort of flourish not solely derivable from the grammar. This
A SPARSE APPROACH TO RIGHT NODE RAISING 43
accent-based inference account bites that bullet and indeed treats RNR as in part
dependent on extra-grammatical operations to produce an acceptable sentence.
But this does not seem like the worst thing. For one, RNR is, as far as is known,
found in every language. As such, the crucial steps that make an RNR sentence what
it is should lie outside the components of grammar that are subject to variation. This
model provides for this. The sentences are allowed to be missing an argument because
of the language-invariant Concatenation operation described above and the missing
argument is inferred via the inferential ability, common to all humans. In essence,
a core operation independent of any specific language gives rise to sentences that
require an operation independent of language in toto.
This model of RNR derivation straddles the variable inner workings of syntax
and predicts the construction to be universal. I consider this an advantage that
other accounts of RNR lack. They all require some operations or features that are
particular to RNR and it is not explained any deeper than as a construction. In
the following section I will investigate the repercussions of this approach to thematic
roles.
5.7. Predictions. Implicit in the above section is the idea that the thematic role of
the missing argument in an RNR sentence is independent of the thematic roles in the
second conjunct. The structure will be be built with regard only to subcategorical
requirements, not to thematic ones. That is, we could begin building a sentence like
in (95) below.
(95) [V P arrived a tall man]
44 BRADLEY LARSON
To this VP another verb that requires an internal argument can Concatenate (take
‘∧’ to signify Concatenation).
(96) [V P expect]∧[V P arrived a tall man]
The sentence hums along as normal and the result is (97), an unacceptable sentence.
(97) *Ivan expected and a tall man arrived
Why is this bad? Inferring a tall man into the gap after expected should be very
straightforward, there is no competing argument and expected desperately needs one.
The problem here is that the sentence is not fertile ground for the correct prosody.
It lacks the contrasting accents followed by gaps and arguments respectively, as
discussed above.
If we alter the sentence minimally so as to allow for the contrastive, rising accents,
the sentence is fine despite the differing thematic roles
(98) Ivan expected, and soon enough there arrived, a tall man.
Given the correct prosody, examples abound in which the overt and inferred argu-
ments differ in thematic role. In (99) the shared element is Agent in one conjunct
and Patient in another. In (100) Patient and Instrument are swapped. And so on.
(99) Ivan saw, and Ivy was seen by, the police
(100) Ivan bought, and Ivy ate her peas with, a fork
(101) Ivy climbed, and Ivan walked towards, the old oak tree.
A SPARSE APPROACH TO RIGHT NODE RAISING 45
Also, by modulating prosody it is possible to give an otherwise fine sentence a
novel interpretation. Take (102). As it stands the sentence is fine. With normal
prosody we understand it as meaning that Ivan was eating a meal during the time
in which Ivy was trying to sell some chestnuts. When RNR prosody is imposed, the
chestnuts are inferred as the complement of eating in the first conjunct, see (103).
(102) Ivan was eating while Ivy tried to sell some chestnuts.
(103) Ivan was EATING, while Ivy tried to SELL, some chestnuts.
This inference can work on more than just arguments. As shown below we can
infer phrases with functional material in them as well. Such is the case with the
preposition in the examples below. Interestingly, the functional material can be
inferred such that it serves a different purpose in each conjunct. Sentence (104) has
the understandable, if slightly painted by non sequitur, interpretation in which Ivan
was hired by someplace or other and Ivy happens to live by the bank. The sentence
with the RNR-type prosody is interpreted in such way that Ivy lives by the bank
that Ivan was hired by.
(104) Ivan was hired and Ivy lives by the bank.
(105) Ivan was HIRED, and Ivy LIVES, by the bank.
5.8. Conclusion. All told, this means of inferring the missing argument makes the
right predictions. Though it is indeed a novel, non-canonical way of doing so, it is
in a sense forced upon us given the previous discussion.
46 BRADLEY LARSON
6. Linearization
Much like with thematic roles, the functioning of linearization under the concate-
native account of RNR is initially less obvious than competing accounts. Also like
with thematic roles, the solution is fairly simple and intuitive. First however, the
linearizations schemes for the competing analyses will be analyzed and shown to be
wanting.
6.1. Deletion. In Deletion accounts of RNR, it is generally stipulated that the
shared material appears at the right edge of the right conjunct, as opposed to the
right edge of the left conjunct. This stipulation alone should be cause for concern.
There is however in Ha (2006) a coding of this stipulation via the ellipsis feature
he uses to account for the apparent elliptical properties of RNR. He posits a feature,
ERNR, that determines what gets deleted (among other things). Under the ERNR
account, the shared material in each conjunct is affixed with this feature, like in
(106). The C head then Agrees with the nearest ERNR feature and this Agreement
causes the deletion of string dominated by the feature, like in (107).
(106) [CP C [TP Ivan bought [ERNR the newspaper] [&P and [TP Ivy read [ERNR
the newspaper]]]]]
(107) [CP C [TP Ivan bought [ERNR the newspaper] [&P and [TP Ivy read [ERNR
the newspaper]]]]]
Agree
This approach runs into problems when there is more than one conjunct. An RNR
sentence with three conjuncts can only have its first instance of shared material
A SPARSE APPROACH TO RIGHT NODE RAISING 47
elided. A sentence initially like (108) would undergo Agree like in (107) and the
result would be (109). Under this approach, it is not possible to derive an RNR
sentence with all but one instance of the shared material elided.
(108) [CP C [TP Iris saw [ERNR the newspaper] [&P and [TP Ivan bought [ERNR
the newspaper] [&P and [TP Ivy read [ERNR the newspaper]]]]]
(109) Iris saw, and Ivan bought the newspaper, and Ivy read the newspaper.
This approach, though in the end untenable, is in the author’s opinion the least
stipulative of the Deletion accounts in terms of linearization. Another ostensibly
straightforward approach to linearization can be found in movement analyses.
6.2. Movement. Most movement analyses of RNR pre-date the LCA (where c-
command determines precedence) and as such they are free to posit movement to
the periphery that translates to movement to the right for purposes of externalization.
These, like the stipulative directionality of Deletion accounts, are not very convincing.
Ignoring this LCA problem, there is a recent movement account of linearization that
is much more interesting than the early ones (like Ross 1967 and Postal 1974).
Sabbagh (2007) presents an elegant method of deriving the linear order of RNR
sentences via movement. Summed up: so long as the movement does not contradict
the linear order as already spelled out (Fox and Pesetsky, 2005), the sentence should
be fine in that regard. Additionally, for Sabbagh rightward movement is argued to
be unbounded and apparently free.
48 BRADLEY LARSON
Given two (simplified) sub-trees not yet coordinated, like in (110) below, the two
instances of John are ordered like in Fox and Pesetsky (in a phase-based, multiple
spell out system) as (111a) and (111b) respectively (where “>" denotes “precedes").
(110) sees
Ivy sees
sees John
hears
Ivan hears
hears John
(111) a. Ivan > hears > John
b. Ivy > sees > John
Spellout having occurred, this ordering cannot be contradicted, and under Sabbagh’s
account, it won’t. To continue on in our simplified manner, the two trees can be
coordinated, resulting in (112). The two instances of John are then free to move out
of the CP to the right, adjoining to CP1 in an across-the-board fashion, like in (113).
(112) &P
hears
Ivan hears
hears John
&′
& sees
Ivy sees
sees John
A SPARSE APPROACH TO RIGHT NODE RAISING 49
(113) &P
&P
hears
Ivan hears
hears ti
&’
& sees
Ivy sees
sees ti
Johni
This movement is licit according to Sabbagh. The ordering when (113) is spelled out
does not contradict that in (111). The string John follows both sees and hears and
Sabbagh asserts that it also follows the larger &P complex. Since the intial ordering
is not contradicted, the final ordering is licit.
The problem with this account is that although it has the trappings of a non-
stipulative account, it falls prey to the same issues that traditional movement ac-
counts of RNR do. The fact that the moved element in (113) adjoins to the right of
the &P complex simply begs the question. There is no motivation to assume that the
moved elements will not adjoin to the left and disrupt the pre-established ordering.
There could potentially be some global constraint on this, but it is not stated. For
instance, it could be the grammar allows the shared material to move to either the
left- or right-preiphery and the left-peripheral version are filtered out. What’s more,
there is no motivation, featural or otherwise, to prompt the movement in the first
place. Without any explicit motivation for this movement we should rule this sort of
movement out. It is true that one could posit any number of features and filters to
50 BRADLEY LARSON
motivate and constrain the movement, but these would amount to ad hoc entities.
If they can be avoided, they should be.
In sum, movement accounts as they stand cannot avoid linear ordering stipulations.
Something which in the Minimalist Program is anathema.
6.3. Multidominance. Given the LCA, Multidominance accounts also have trouble
when it comes to linearization. Take an example like (114), The LCA would produce
orderings like those in (115) and (116).
(114) &
read &
& bought
Ivan bought
bought
Ivy read
read
the
the book
The right TP c-commands the left one and as such precedes it. This precedence
relation extends to everything dominated by the respective TPs. This produces the
following orderings (of many)
(115) Ivy > bought
(116) the book > the book
A SPARSE APPROACH TO RIGHT NODE RAISING 51
The ordering in (116) violates the LCA’s reflexivity restriction and Multidominance
requires alterations to the LCA to avoid this15. We do not want to alter the LCA
for a single construction. It is rather elegant as it stands and the following analysis
of RNR allows it to remain that way.16
6.4. The Concatenative Account. Unlike the previous analyses, the Concatena-
tive account requires neither unmotivated stipulation nor alterations to conceptual
machinery to produce linear orderings. Kayne’s LCA provides a straightforward
mapping between c-command relations and precedence and I assume it here.
Given the Concatenative RNR tree in (117), it is at first not entirely clear how
the Concatenation (dashed line) will affect or play a role in the linear order. One
could, for instance, suppose that the Concatenation of the two verbs introduces some
linear order ambiguity to the tree and as such requires something additional to be
linearizable.
15 See Wilder (1999) for a reformulation of the LCA to allow for Multidominance as well as Bachrachand Katzirs Multidominance linearization scheme for another alternative. Though also see An(2007) and Ha (2006) for problems for each of these alternatives respectively.16In the next section I do adopt an alteration to the LCA, but one that is, if not uncontroversial,less construction-specific than alterations for Multidominance.
52 BRADLEY LARSON
(117) &
T
&
and T
T read
Ivy read
T bought
Ivan bought
read the
the book
However, it is crucial that Concatenation does not create any c-command relations,
only Labeling can do that. That is, c-command relies on a parent shared between
two nodes and without Labeling there is no shared parent to construe c-command
over. Since the LCA is only concerned with c-command relations, the dashed line
is effectively invisible with respect to linearization. As far as the LCA is concerned
(117) appears like (118). In other words, the initial conjunct is essentially a complex
specifier for the purposes of linearization.17
17 This tree shares a resemblance to Phillips (1996) account of RNR in which, building left to right,the first conjunct sans the shared material forms a constituent. Here it also forms a constituent asfar as linearization is concerned.
A SPARSE APPROACH TO RIGHT NODE RAISING 53
(118) &
T
&
and T
T read
Ivy read
T bought
Ivan bought
read the
the book
Example (118) spelled out (with the initial conjunct acting like a run-of-the-mill
complex specifier) gives us the grammatical RNR sentence that we would expect.
And it does so without recourse to any stipulations aside from those inherent to the
assumed frameworks of Kayne and Hornstein.
6.5. Specifiers. Why is it that the following sentences (from Johnson, 2007) are
unacceptable?
(119) a. *Mary talked to about Jim, and Sally talked to Paul about, Marilyn
Manson.
b. *Mary talked to about Jim, and Sally talked about to Paul, Marilyn
Manson.
c. *Sally mailed to Bill, and Mary faxed, the contract to Sam.
How can the means of structure building presented above account for the fact that
the gap in the first conjunct of an RNR sentence must appear on its right edge (call
54 BRADLEY LARSON
it the right-edge restriction)? That is, the shared material does not seem to be able
to be in a specifier position in the first conjunct. In the sentences above there is
a preposition in the specifier position of a lower VP projection (following Larson,
1988).
To understand why this is the case, let us look at the state of the derived object
at a point in time before the first conjunct’s relevant specifier is formed (120).
(120)
Q X
Q can be taken as a complement without any obvious negative repercussions.
Also, from this perspective it is in principle possible for Q to also serve as a specifier.
But assuming a multiple spell-out framework like that of Uriagereka (1999) we can
preclude this. Under Uriagereka’s analysis, syntactic objects are spelled-out prior to
their having merged into a specifier position. Spell-out in his terms essentially renders
the object a giant word, its innards impervious to subsequent syntactic operations.
That this happens allows for a simpler conception of the LCA. Prior to this, an an
element within a specifier would precede, say, the verb without c-commanding it. It
would certainly be dominated by something that c-commanded the verb, but it itself
wouldn’t and an inductive step was required to achieve the effect of c-command.
If the specifier were made into a word-like element, this problem no longer arises.
What’s more it captured many other empirical facts.
A SPARSE APPROACH TO RIGHT NODE RAISING 55
Relevant for us is that the spell-out of the incipient specifier required that the
object be a command unit. A command unit would be any tree without any complex
specifiers, an object whose linear ordering could be read off the structure without
induction (121).
(121) X
Y X
X Z
The LCA determines the order of the elements in the above tree and then the tree
is flattened. When we look at the fledgling RNR structure above and consider how
it would work in this situation, it becomes clear what will go wrong. The tree is
presented again in (122).
(122)
Q X
Were the above object to be made word-like, there would no longer be a node X that
could serve to build the second conjunct. This would preclude the RNR sentence
from getting anywhere near completion. This then rules out the sentence in which
the shared element finds itself in a specifier position and this will in turn predict the
right-edge restriction in RNR sentences.
56 BRADLEY LARSON
It could also be the case that the above object does not even constitute a command
unit. As such it is not a viable candidate to be spelled out and inserted as a specifier
to begin with. There is no way to derive the right-edge restriction violating sentences
under this view either.18
7. Conclusion
I hope to have convinced the reader of two things. One, the current ways of
analyzing Right-node Raising are untenable without substantial revision. Two, such
revision is unnecessary. The approach presented here accounts for a broader swath
of facts than any of its competitors and does so in a simple, principled manner.
One such avenue for further investigation concerns ATB wh-questions. Much like
the asymmetrical c-command effects found in RNR from left to right, ATB wh-
questions seem to shows asymmetrical c-command effects from right to left. As a
taste of further inquiry, I present (123). Wh-island effects arise only in the first
conjunct, not the second.
(123) a. *What did Ivy say who buys t and Jack say Ivan sells t?
b. What did Jack say Ivan buys t and Ivy say who sells t?
18 It would of course need to be the case that the structure could be flattened when the coordinationenters the picture. Though if the entirety of the structure were flattened at this point, nothingwould go awry. The massive &P would simply be spelled out as a word. That is, we can spell-outobjects like (122) so long as the Q is merging as the specifier to something in the spine thatcontains X. This has the repercussion of ruling in sentences like (i)
(i) The man who likes, met the woman who hates, armadillos.
In this sentence the Q analogue likes eventually finds itself as a subpart of a specifier. The nodethat it is a specifier to happens to dominate the X analogue hates and the sentence is ruled in, tomy ears correctly so.
A SPARSE APPROACH TO RIGHT NODE RAISING 57
8. Bibliography
An, D. 2007. Syntax at the PF interface: Prosodic mapping, linear order, anddeletion. Ph.D. dissertation. University of Connecticut, Storrs.
Bachrach, A. and R. Katzir. 2009. Right Node Raising and Delayed Spellout.In Interphases: Phase-Theoretic Investigations of Linguistic Interfaces, ed. byGrohmann, Kleanthes K. Oxford: Oxford University Press.
Baker, M. 1989. Object sharing and projection in serial verb constructions. Linguis-
tic Inquiry 20:513-553.Barros, M., and L. Vicente. 2010. ÒRight Node Raising requires both ellipsis and
multidominationÓ. The 34th Penn Linguistics Colloquium. March 19-21.Boskovic, Z., and S. Franks. 2000. Across-the-board movement and LF. Syntax
3.2 :107-128.Chomsky, N. 1995. The Minimalist Program. Cambridge, MA: MIT Press.Chomsky, N. 2001. Derivation by phase. in Ken Hale: a life in language 1-52, (ed.)
Kenstowicz, M. Cambridge, Mass.: MIT Press.Citko, B. 2005. On the Nature of Merge: External Merge, Internal Merge, and
Parallel Merge. Linguistic Inquiry 36: 475-496.Fery, C. and K. Hartmann. 2005. Focus and prosodic structures of German gapping
and right node raising. The Linguistic Review 22. 79-115.Fiengo, R. and R. May. 1994. Indices and Identity, MIT Press, Cambridge, Mass.Fox, D., and D. Pesetsky. 2005. Cyclic linearization of syntactic structure. Theoret-
ical Linguistics 31: 1-46.Frampton, J. and S. Gutman. 2002. Crash-Proof Syntax. In S. Epstein and D. Seely,
eds., Derivation and Explanation in the Minimalist Program. Blackwell, Oxford,90-105
Gazdar, G. 1981. “Unbounded Dependencies and Coordinate Structure," Linguistic
Inquiry 12, 155-184.Gracanin-Yuksek, M. 2007. On sharing. Doctoral Dissertation, MIT.Grosz, P. 2009. Movement and agreement in Right Node Raising constructions. Ms.,
MIT.Ha, S. 2006. Multiple dominance CAN’T, but PF-deletion CAN account for Right
Node Raising. Paper presented at the 42nd Annual Meeting of the CLS.Hartmann, K. 2000. Right node raising and gapping: Interface conditions on prosodic
deletion. Amsterdam: John Benjamins.Heim, I. and A. Kratzer. 1998. Semantics in Generative Grammar. Oxford: Black-
well.Hornstein, N. 2009. A Theory of Syntax : minimal operations and universal gram-
mar. Cambridge: Cambridge University Press.
58 BRADLEY LARSON
Hornstein, N., and J. Nunes. 2008. Adjunction, labeling, and bare phrase structure.Biolinguistics 2:57-86.
Ince, A. 2009. Dimensions of Ellipsis: Investigations in Turkish. Ph.D. dissertation.University of Maryland, College Park.
Johannessen J. B. 1998. Coordination. Oxford: Basil Blackwell.Johnson, K. 2007. LCA+Alignment=RNR, talk presented at the Workshop on Co-
ordination, Subordination and Ellipsis, Tubingen, June 2007.Kayne, R. 1993. Toward a modular theory of auxiliary selection. Studia Linguistica
47:3-31.Klima, E. 1964. Negation in English. In J. Fodor and J. Katz (eds.), The structure
of language: Readings in the philosophy of language, 246-323. Prentice Hall.Larson, B. 2009. “Ellipsis Does, But Right-node Raising Doesn’t, Involve Deletion".
GLOW 32. Nantes, France.Larson, B. 2010. “Bare Phrase Coordination". GLOW 33. Wroclaw, Poland.Larson, B. In Press. A Dilemma with Accounts of Right-node raising. Linguistic
Inquiry
Larson, R. 1988. On the Double Object Construction. Linguistic Inquiry. 19:335-391.
McCawley, J. 1982. “Parentheticals and Discontinuous Constituent Structure," Lin-
guistic Inquiry 13, 91-106.Pesetsky, D. and E. Torrego 2004. The Syntax of Valuation and the Interpretability
of Features, Ms. MIT/U.Mass Boston.Phillips, C. 1996. Order and Structure. Ph.D. dissertation. MIT.Phillips, C., Wagers, M., and Lau, E. 2009. Grammatical illusions and selective falli-
bility in real-time language comprehension. Language and Linguistics Compass.Pietroski, P. 2002. Function and concatenation. In Gerhard Preyer and Georg Peter,
eds., Logical Form and Language. 91-117. Oxford: Clarendon Press.Pietroski, P. 2005. Events and Semantic Architecture. Oxford: OUP.Postal, P. 1974. On Raising, MIT Press, Cambridge, Massachusetts.Reinhart, T. 1976. The syntactic domain of anaphora. Ph.D. dissertation. MIT.Ross, J. 1967. Constraints on variables in syntax. Doctoral Dissertation, MIT.Sabbagh, J. 2007. Ordering and linearizing rightward movement. Natural Language
and Linguistic Theory 25:349-401.Sabbagh, J. 2008. Right Node Raising and extraction in Tagalog. Linguistic Inquiry
39:501-511.Uriagereka, J. 1999. Multiple Spell-Out. in S.D. Epstein and N. Hornstein (eds.)
Working Minimalism. Cambridge, MA: MIT Press, 251-82.de Vos, M., and L. Vicente. 2005. Coordination under Right Node Raising. In
Proceedings of WCCFL 24, ed. Alderete et al, 97-104. Somerville: CascadillaPress.
A SPARSE APPROACH TO RIGHT NODE RAISING 59
Wexler, K. and P. Culicover. 1980. Formal Principles of Language Acquisition, MITPress, Cambridge, Massachusetts.
Wilder, C. 1997. Some Properties of Ellipsis in Coordination. Studies on UniversalGrammar and Typological Variation, ed. A. Alexiadou and T. A. Hall, 59-107.Amsterdam: John Benjamins.
Wilder, C. 1999. Right Node Raising and the LCA. In Proceedings of WCCFL 18,ed. Bird et al, 586-598. Somerville: Cascadilla Press.
Williams, E. 1981. “Transformationless Grammar," Linguistic Inquiry 12, 645-654.