Nat Lang Linguist Theory (2013) 31:1015–1066
DOI 10.1007/s11049-013-9206-8
Explaining the structure of case paradigms
by the mechanisms of Nanosyntax
The Classical Armenian nominal declension
Pavel Caha
Received: 12 January 2010 / Accepted: 1 August 2012 / Published online: 26 October 2013
© Springer Science+Business Media Dordrecht 2013
Abstract This paper looks in detail at the Classical Armenian nominal declension.
I highlight several generalizations that can be read off the surface paradigms, including
restrictions on syncretism, fusional vs. agglutinative expression of categories and
the emergence of unexpected thematic vowels. Subsequently, I explain these generalizations
within the framework of Nanosyntax (Starke 2009, 2011).
The defining features of the account are fine-grained syntactic representation
(a single feature per syntactic terminal) and phrasal spell-out. I argue that these two
tools allow us to replace a separate level of morphological (paradigm specific) structure
by a syntactic tree.
Keywords Case · Case syncretism · Classical Armenian · Paradigm · Phrasal
spell-out · Nanosyntax · Agglutination · Fusion
1 Introduction
There seems to be a broad consensus in the literature that in order to describe the fine
details of morphological paradigms, one needs something in addition to the syntactic
structure and the lexicon (for some recent work in the generative tradition, see,
e.g., Williams 1981, 1994; McCreight and Chvany 1991; Halle and Marantz 1993;
Ackema and Neeleman 2007). The current paper focuses in detail on case declension
of Classical Armenian in order to (re-)evaluate the type of evidence usually brought in
favor of separate morphological (or more narrowly paradigm) structure. I argue that
once syntax is made sufficiently fine-grained, with individual features corresponding
to syntactic heads, paradigm structure may be dispensed with.
P. Caha (B)
CASTL, University of Tromsø, 9037 Tromsø, Norway
e-mail: pavel.caha@uit.no
1016 P. Caha
I start by looking at case syncretism. The term refers to a situation in which a
particular noun does not morphologically distinguish two (or more) distinct cases.
An example is given below. In (1), we can observe that the noun ‘God’ has the same
form in the ablative and the genitive. In (2), we see a noun that distinguishes the two
environments.1
(1) Syncretism of GEN-ABL
a. ordi-k‘
son-NOM.PL
Astuc-oy
God-GEN
‘the children of God’
b. i
by
Astuc-oy
God-ABL
‘by [the benevolent] God’
(2) No syncretism of GEN-ABL
a. merjawor
acquaintance
Daniêl-i
Daniel-GEN
‘an acquaintance of Daniel’
b. i
from
Daniêl-ê
Daniel-ABL
‘from Daniel’
Syncretism is a pervasive feature of natural languages (Baerman et al. 2005), and the
declension of Classical Armenian is a good example of this. In Sects. 2–3, I show
that in Classical Armenian, syncretism is subject to a restriction such that in a linear
sequence of cases, only adjacent cases may enter this relationship. The existence
of such a type of constraint has led McCreight and Chvany (1991:91) to argue that
“syntactic features are inappropriate to the modeling of paradigms, and that geometric
representations [...] can provide more insight, particularly in the description of
syncretism.”
This paper argues that no specific paradigm structure is needed in order to account
for the data, including the observed restrictions. In Sect. 4, I demonstrate that the
facts may be captured using a particular case decomposition, such that when we go
left to right in the linear sequence uncovered, cases grow in complexity with each
case adding a feature compared to the previous one (cf. Starke 2009; Bobaljik 2012).
Section 5 compares the proposal to some common alternative approaches, and shows
its advantages. In sum, the current theory is reductionist: a specific module dedicated
to paradigm structure is not necessary in order to account for the facts (cf. Bobaljik
2002 and further work in that spirit).
At the same time, I argue in Sect. 6 that each case feature corresponds to a syntactic
head (cf. Caha 2009; Caha 2011b).2 The argument is based on the observation
that case attraction in Classical Armenian is sensitive to the linear sequence that re-
1Examples from Krause and Slocum (no date).
2To put the idea in a context: within Cartography (Cinque and Rizzi 2010), it is generally agreed that each
feature is hosted by a separate head. A related idea is presented in Kayne (2005). Starke (2009, 2011)
develops these ideas in a specific framework, Nanosyntax, whose research program I follow closely.
Explaining the structure of case paradigms by the mechanisms 1017
stricts syncretism. Thus, despite its reductionist nature, the theory does not deny the
existence of paradigm structure. Paradigm structure is real, but not autonomous: it
corresponds to the syntactic structure.
Section 7 is devoted to a detailed presentation of phrasal spell-out, a device that
may map both terminals as well as complex phrasal nodes directly on their pronunciation
(drawing on Starke 2009). With the proposal in place, Sect. 8 turns to its
consequences for the splits between agglutination and fusion (cf. Plank 1999). I use
the term fusion in order to describe the fact that sometimes, case is expressed together
with other inflectional categories. For example, in (3a), it is expressed together with
number. The term agglutination, on the other hand, refers to instances where such
categories are expressed separately, as in (3b). The separation of case and number
is seen in comparison with the singular form, which is just like the plural minus the
plural marker -k‘. (No such segmentation is possible in (3a)).
(3) a. azg-i
nation-LOC.SG
— azg-s
nation-LOC.PL
b. azg-aw
nation-INS.SG
— azg-aw-k‘
nation-INS-PL
In Sect. 8 I observe that agglutination and fusion are not distributed randomly across
the paradigm. Specifically, agglutination in a particular case K entails agglutination
in all cases that can be found to the left of K in the linear sequence relevant for
syncretism—apparently yet another paradigm structure effect. The theoretical challenge
is to model the varying number of exponents in the two distinct cases shown
in (3a, b) (as well as capturing the regularity such splits are subject to): what blocks
*azg-i-k‘ as the plural of azg-i in (3a)?
Williams (1994:22) says concerning a similar example that “in general, if there
are two ways of filling a slot in the paradigm [azg-i-k‘ vs. azg-s], only one may be
used. [...] This reveals that there is a target slot to fill, which is independent of the
rules for filling it, and that slot is given by the paradigm.”
Working in a Distributed Morphology model, Halle and Vaux (1998) answer the
challenge differently. They propose that there is only a single terminal in syntax for
both case and number (in effect blocking azg-i-k‘ by a proposal about the syntax of
Classical Armenian). In order to account for the instrumental plural, they propose a
morphology specific rule (Fission) that turns the single underlying terminal into two
surface positions.3 Thus, despite important differences in execution, both solutions
entail the existence of a morphology specific level of representation (corresponding
to the paradigm structure, with a particular number of slots to fill).
In Sect. 9, I offer an alternative solution in terms of phrasal spell-out. I show
that under this view, the number of exponents that are needed to spell out each particular
case naturally emerges from the way lexical insertion interacts with syntactic
structure. There is no need for any morphological rules (like Fission) producing
3Had there been two positions in syntax, a separate rule (presumably Fusion) would be needed to handle
those cases where only a single marker appears.
1018 P. Caha
paradigm representations with a particular number of slots, a conclusion I strengthen
in Sect. 10. Here, I focus on two consonantal declensions in the language, and show
that apparently unexpected thematic vowels show up in this type of declension.
I demonstrate, however, that no additional rules are needed to explain their existence;
the vowels appear as a consequence of the proposed spell-out procedure interacting
with the independently established syntactic structure.
The result is a theory of paradigm structure that eschews both syntax independent
representations and syntax independent rules, leading to the claim that paradigm
structure is in fact best accounted for in terms of syntactic structure, using established
notions such as binary branching, structural intervention and others.
2 A linear restriction on syncretism
I open the empirical discussion by introducing a strong restriction on syncretism in
Classical Armenian. In establishing the generalization, I rely on the description offered
in Schmitt (1981), and I also partly draw on aspects of the analysis of Classical
Armenian presented in Halle and Vaux (1998).
According to standard descriptions, Classical Armenian has seven cases: nominative,
accusative, locative, genitive, dative, ablative and instrumental. I will stick to
this description, noting that the INS can also be used as a comitative (so potentially
there can be eight cases), and that the GEN and DAT are not formally distinguished
in the nominal declension (so potentially there can be six cases, with the caveat that
GEN/DAT are distinct in pronouns, including demonstratives). Putting aside for now
one apparent exception (to be discussed later), syncretism in case is restricted to contiguous
regions in a linear sequence, which I give in (4b).4
(4) Case Contiguity (Armenian):
a. Non-accidental case syncretism is restricted to contiguous regions in the
following sequence:
b. NOM – ACC – LOC – GEN – DAT – ABL – INS
The statement (4) rules out a number of logically possible syncretisms. For instance,
NOM and LOC cannot be syncretic to the exclusion of ACC, because they are not
contiguous in (4b). This restrictive prediction of (4) is borne out and the syncretism
is unattested. I show that schematically in the first column of the table (5) (syncretic
cells highlighted by shading, N.A. means ‘not attested’).
4The phrase “non-accidental” in (4) is intended to exclude two sources of homophony: (i) phonological
conflation, whereby two distinct underlying forms end up homophonous due to a regular phonological
process, and (ii) accidental homophony. Accidental homophony plays no role in this paper; however, one
instance of homophony will be analyzed as an example of phonological conflation in the sense described
above (see Sect. 10.3).
Explaining the structure of case paradigms by the mechanisms 1019
(5) The restrictions on syncretism in Classical Armenian
It is, however, not the case that NOM and LOC are never the same; see the second
column of the table (5). Crucially, when this happens, then ACC is syncretic with
the NOM–LOC pair, leading to a syncretism which targets a contiguous region of
(4b). Similarly, as shown in the first column, GEN and ABL are never syncretic to
the exclusion of other cases. But again, such a syncretism becomes possible once it
covers DAT as well, applying to a region contiguous in (4b) (see the dark shade in the
second column).
A similar prediction is that, for instance, LOC–DAT or LOC–ABL syncretisms are
unattested, but LOC–GEN–DAT or LOC–GEN–DAT–ABL can be found. This is shown
in the remainder of the table (5).5
To evaluate the restrictive power of such a linear constraint abstractly, consider
some numbers. In a system with seven cases (Classical Armenian), there are 120 logically
possible syncretisms. The statement (4) predicts that 99 of these are unattested,
and allows only 21.
It has been argued in the literature that Classical Armenian is not unique in possessing
a linear contiguity constraint, and that such a situation is quite general; see
McCreight and Chvany (1991); Johnston (1996); Caha (2009) and references there.6
If this is so, such a large scale descriptive reduction of possibilities in language after
language raises a challenge for the theory of syncretism. Clearly, the theory should
be able to derive this, and, preferably, offer an account that extends to other aspects
of case than syncretism.
In this paper, I offer a way to understand the linear constraint (4) as a consequence
of the proposal that case features are syntactic heads, ordered in a functional
sequence. I also show that the same linear sequence is relevant for other phenomena
5DAT–LOC syncretism across a distinct GEN is attested in the pronominal declension (e.g., mez/jez ‘we,
you, ACC/LOC/DAT,’ mer/jer ‘my, our’). The reason for this apparent exception is that the ‘genitive’ pronoun
is in fact a possessive pronoun, which formally does not belong to the paradigm, and hence, disturbs
the picture.
There are several pieces of evidence to support this claim: (i) unlike regular genitives, the pronominal
possessive forms take additional agreement suffixes; (ii) unlike regular genitives, the possessive forms
show no syncretism with any other case; this is expected if they are not case forms, but unexplained
otherwise; (iii) the ‘genitive’ ending -r (me-r/je-r ‘my, our’) shares diachronic origin with morphemes
routinely classified as possessive markers and not as case markers, e.g., in Latin forms such as nos-ter
‘our.’
6Similar observations have been made in other domains than case, see Starke (2005, 2009); Bobaljik
(2007, 2012); Pantcheva (2010, 2011) and Vangsnes (2013) for approaches directly related to the one
pursued here.
1020 P. Caha
in the grammar, and work out a proposal how this follows from the initial proposal
developed for syncretism.
2.1 Establishing the order
Let me begin by introducing a different aspect of the constraint (4), which is positively
predictive. In particular, if a language does not show any syncretism, then any
linear ordering of cases yields (trivially) a correct generalization. Evidence for any
particular ordering is thus more convincing if supported by as many attested examples
as possible.
This aspect of the linear constraint is illustrated on a representative sample of
paradigms shown in the table (6). The table is organized in such a way that the cases
are ordered top-down according to the sequence given in (4b), i.e., NOM–ACC–LOC
and so on. The cells in darker shade show pair-wise syncretisms of adjacent cases,
and move gradually one notch down as we go in the table from left to right. Lighter
shade marks cells where syncretism (irrelevantly for the main point) extends beyond
a simple pair.
(6) Attested syncretisms in Classical Armenian
The linear order is then established as follows. In the singular of the noun ‘word,’
NOM and ACC show syncretism to the exclusion of all other cases. From the perspective
of linear ordering, this means that they must be neighbors in the linear order
relevant for syncretism: NOM–ACC. ACC and LOC are the same in the plural, see the
shading in the plural of ‘nation.’ This leads to NOM–ACC–LOC. LOC in turn must be
adjacent to GEN and DAT on the basis of the syncretism in the singular of ‘nation’:
NOM–ACC–LOC–GEN/DAT.
In all the paradigms above, and in the nominal system in general, GEN and DAT
are always the same. This means that their order cannot be determined internally to
the nominal declension of Classical Armenian. Thus, the reason I state the order as
GEN–DAT rather than DAT–GEN is not motivated by Classical Armenian alone. It is,
however, justified by cross-linguistic data, which I discuss in the next section.
While the mutual order of GEN and DAT cannot be decided internally to Classical
Armenian, it can be established that the ABL comes after these two cases, due to
the syncretism in the plural (see ‘river’). This leads to the ordering NOM–ACC–LOC–
GEN/DAT–ABL. INS shows no syncretisms in Classical Armenian; it then comes ei-
Explaining the structure of case paradigms by the mechanisms 1021
ther last or first.7 (It cannot come in the middle, because then it would disturb the
needed adjacency between other cases.)
We are thus left with four possible orderings, and hence, four possible ways to
state a linear constraint on syncretism in Classical Armenian. I give them in (7):
(7) Four possible sequences with syncretisms contiguous
a. NOM – ACC – LOC – GEN – DAT – ABL – INS
b. NOM – ACC – LOC – DAT – GEN – ABL – INS
c. INS – NOM – ACC – LOC – GEN – DAT – ABL
d. INS – NOM – ACC – LOC – DAT – GEN – ABL
2.2 Summing up
As we have seen, the syncretisms of Classical Armenian occupy contiguous regions
in the linear sequence given in (4). While correct in the sense that the constraint
shows minimal violations (with one exception to be presented and explained away
in Sect. 10.3), the sequence in (4) is underdetermined by the actual data. First, there
is no evidence for mutual ordering of GEN and DAT; second, since INS shows no
syncretisms, it can either come last or first.
3 What is the Case sequence of Armenian?
In this section, I present some considerations which favor the statement (4) over possible
alternatives.
3.1 Universal Contiguity
The first argument for the ordering (7a) comes from cross-linguistic comparison.
Caha (2009) proposes a hypothesis, Universal Contiguity, which says that across languages,
there is a fixed sequence of cases in which only contiguous regions show
syncretism:8
(8) Universal (Case) Contiguity:
a. Non-accidental case syncretism targets contiguous regions in a sequence
invariant across languages.
b. The Case sequence: NOM – ACC – GEN – DAT – INS – COM
In the sequence (8b), GEN precedes DAT, and INS comes after these two. In Classical
Armenian, only the order (7a) is consistent with this cross-linguistic pattern. Hence,
it is reasonable to assume that it is the correct one.
7There is actually one homophony of INS with DAT, but this is due to a phonological conflation (see Halle
and Vaux 1998:n.7). I come back to this later on.
8It has to be mentioned that Plank (1991) argues that a linear contiguity constraint is too strong for a
number of languages. See Johnston (1996) and Caha (2009) for a linear analysis of some of the languages
identified as problematic in Plank (1991).
1022 P. Caha
Consider a couple of examples from three different branches of Indo-European
that illustrate (8). In Russian (McCreight and Chvany 1991), the ordering is unambiguously
NOM–ACC–GEN–PREP–DAT–INS, see (9). Relevantly, GEN precedes DAT,
and DAT precedes INS.
(9) Syncretism in Russian (McCreight and Chvany 1991)
The order GEN–DAT–INS appears also in Old English (Plank 1991; Caha 2009):
(10) Syncretism in Old English (Plank 1991)
Finally, syncretism in Sanskrit (Plank 1991; Johnston 1996) is only consistent with
an order which includes GEN–DAT–INS, where the position of DAT between GEN and
INS is relevant for our present concerns.
(11) Sanskrit (paradigms from Baerman 2008)
These observations allow us to understand the pattern found in Classical Armenian
as a special instance of the general scenario (8). If that is so, we have to order GEN
before DAT, and INS must be last rather than first.9
9Needless to say, there are a number of questions that arise once cross-linguistic perspective is taken. Two
issues deserve mentioning: (i) local cases show up at places that are not directly comparable; (ii) particular
languages present challenges to the view expressed in (8) (the reviewers mention specifically Latin). I cannot
possibly address these issues within the space of an article, and I can only provide references to places
where these challenges are addressed. For local cases, see Caha (2009:Chap. 3.4.3–4) and Caha (2011a).
For cross-linguistic issues (including the discussion of Latin), see Caha (2009:Chaps. 3 & 8).
Explaining the structure of case paradigms by the mechanisms 1023
3.2 Case attraction
There is additional evidence that INS should come last (and not first) in the sequence:
case attraction (Plank 1995:43; Blake 1994). In Classical Armenian, the complement
of a noun is ‘normally’ expressed by the GEN, as schematically depicted in (12a).
However, if the head is in ABL or INS, the dependent GEN can be ‘attracted.’ Attraction
consists in replacing GEN by the case which is carried by the head noun, see
(12b, c). However, if the head noun is in another case (NOM, ACC, or LOC), attraction
is unattested, see (12d):10
(12) Case attraction in Classical Armenian
a. [ N [ N-GEN ] ]
b. [ N-ABL [ N-GEN ⇒ N-ABL ] ]
c. [ N-INS [ N-GEN ⇒ N-INS ] ]
d. *[ N-NOM/ACC/LOC [ N-GEN ⇒ N-NOM/ACC/LOC ] ]
A real language example of attraction is in (13).
(13) [ bazmut‘-eamb
crowd-INS.SG
[ zawr-awk‘-n
force-INS.PL-DEF
Hay-oc‘
Armenian-GEN.PL
]]
‘with a crowd of the Armenian forces’ (Plank 1995:43)
In (13), the head noun ‘crowd’ bears INS. The head has a complement, ‘of the Armenian
forces.’ The head of the complement, ‘forces,’ would ‘normally’ occur in GEN.
However, as a result of attraction, it bears INS (see the boldfaced affix), inherited from
the head noun.
Importantly, if we choose a sequence with INS adjacent to ABL, we capture not
only the restrictions on case syncretism, but also the restrictions on case attraction, as
seen in (12b–d). Specifically, in the syncretism sequence in (4), cases to the right of
GEN have the power to attract it, cases to its left cannot.
3.3 Summing up
To sum up the two preceding sections: syncretism in the Classical Armenian nominal
declension is restricted to contiguous regions in a linear sequence of cases. Out of four
possible orderings, we have chosen the one which (i) captures additional restrictions
in the language (case attraction), and (ii) is consistent with a larger cross-linguistic
pattern. In the next section, I provide a way to understand the case sequence theoret-
ically.
4 Deriving Contiguity
In this section, I show that (4) follows from a particular feature decomposition, interacting
with a spell-out procedure based on the Elsewhere Condition. Relevant for
10It cannot be decided whether attraction does or does not occur in the dative. Since the dative is always
the same as the genitive, case attraction, if active, applies vacuously.
1024 P. Caha
the issues highlighted in the introduction is going to be the conclusion that linear
contiguity is expressible in terms of feature decomposition, and does not need to be
encoded through paradigm-specific notational devices.
4.1 Cumulative decomposition
The first thing we need is that the number of features characteristic for each case has
to grow monotonically as we move along the contiguity sequence in (4). Thus, what
we need is that, for instance, NOM = [A], ACC = [A, B], LOC = [A, B, C] and so on,
see (14). I will call such a decomposition ‘cumulative.’
(14) Cumulative case decomposition
a. NOM = [A]
b. ACC = [A, B]
c. LOC = [A, B, C]
d. GEN = [A, B, C, D]
e. DAT = [A, B, C, D, E]
f. ABL = [A, B, C, D, E, F]
g. INS = [A, B, C, D, E, F, G]
What is crucial to derive Case Contiguity is the formal character of the decomposition,
and not the content of the features. Consequently, the proposal is compatible
with a number of interpretations, not crucial to the argument (though see Sect. 6.2
for a suggestion). Let me also remark that even though (14) may suggest that case
features are privative, this is not necessary. For example, NOM can be [+case], ACC
can be [+case; −independent], etc. As long as we keep adding features, the resulting
system can be used to derive contiguity.
4.2 Two spell-out rules
I now proceed to show how this representation contributes to understanding the contiguity
requirement on syncretism. The gist of the proposal is the idea, shared among
various frameworks, that lexical entries are not tailor-made for one representation
only, but they can be associated to a larger number of representations. Linear Contiguity
then follows if an entry may only target a set of cases that form a contiguous
region on the scale in (4).
With the decomposition (14) in place, such a goal may be achieved in two ways.
Either we propose that if a given lexical entry α applies to a case K, it applies to all
cases that contain K. For example, if α is specified for GEN ([A,B,C,D]), it applies
automatically also in DAT, ABL, INS. This strategy is known as the Subset Principle
(see, e.g., Halle 1997 among many others).11
Alternatively, we may propose that if a lexical entry α applies to a case K, it applies
also to all cases contained in K. Thus, if α is specified for GEN, it applies also in LOC,
ACC, NOM. This assumption is called the Superset Principle (see Starke 2009). I state
the Superset Principle informally below:
(15) The Superset Principle (preliminary version): A phonological exponent is
inserted into a node if its lexical entry contains all features of that node.
11The Subset Principle, as Halle (1997) states it, is in fact a combination of a ‘pure’ Subset Principle and
the Elsewhere Condition. Here, I ignore the Elsewhere part of the Subset Principle, and treat it as a separate
rule (cf. Harley 2008).
Explaining the structure of case paradigms by the mechanisms 1025
It can be shown that if one adopts the representations in (14), both approaches derive
(4) as a theorem. In what follows, I demonstrate this only for the Superset Principle
(see Bobaljik 2012 for a related proposal based on the Subset Principle). This choice
is made on the basis of the observation—to be justified in Sect. 9.3—that for certain
facts of Classical Armenian, the Superset Principle offers a better solution than its
alternative.
The theory sketched so far constrains syncretism to contiguous regions in (4), but
it is as yet incapable of dealing with syncretisms that do not include NOM. Thus, recall
that the entry for a given case automatically applies to all cases contained in it,
leading only to syncretisms that stretch from the particular case to NOM. This apparent
problem disappears once competition among entries is taken into consideration,
and the Elsewhere Condition is adopted to regulate it. The reasoning follows.
I start from the observation that there are situations where more than one entry
is applicable in a given case. Suppose, for example, that there are the two following
entries, α and β:
(16) a. /α/ ⇔ [C,B,A]
b. /β/ ⇔ [A]
In (16a), α is specified for the features of LOC. By the Superset Principle, it may then
also be inserted in ACC and NOM. β is specified for NOM. According to the Superset
Principle, this exhausts its applicability. The result is that in NOM, both rules may
apply:
(17) The range of applicable environments of α and β
The two entries ‘compete,’ and the result is determined by the Elsewhere Condition.
This condition says that the most specific entry wins over the others (see Kiparsky
1973, my formulation draws on Neeleman and Szendr˝oi 2007).
(18) The Elsewhere Condition: In case two rules (R1 and R2) can apply in an
environment E, R1 takes precedence over R2 if it applies in a proper subset
of environments compared to R2.
In our example, the entry for β (R1), takes precedence over α (R2) in NOM, because
it applies in a proper subset of environments. (It does not apply in ACC, LOC.) The
result of the competition is that α surfaces in LOC and ACC only, a contiguous region
that does not include NOM. Hence, once we introduce the Elsewhere Condition, the
generative capacity of the system is increased to yield also contiguous syncretisms
without NOM.
Consider now how the proposed system derives the Contiguity constraint. To see
that it does, suppose that we want to encode a syncretism which would violate it, e.g.,
NOM and LOC are the same to the exclusion of ACC, as in the hypothetical paradigm
(19). If it turns out that such a syncretism cannot be encoded by the spell-out system
1026 P. Caha
operating on the proposed decomposition, we prove that the system derives Contiguity,
as manifested in Classical Armenian.
(19) An offending paradigm
To generate the offending paradigm, we have to come up with an entry that can
appear both in the locative and the nominative. Such an entry is (20).
(20) /α/ ⇔ [ C B A ]
Now we need an entry which can spell out ACC (B, A) but not LOC (C, B, A), see (21).
Such an entry provides a perfect match for ACC, and due to competition, removes it
from the set of cases where the entry (20) surfaces.
(21) /β/ ⇔ [ B A ]
However, the entry (21) can apply in NOM (A) as well. Hence, the entries (20) and
(21) compete not only for ACC, but also for NOM, see (22). In such a situation, the
rule introducing β takes precedence over α in NOM as well, because it is a better
match; see (23).
(22) The applicability of α and β
(23) The paradigm generated
In words: whenever we get α in LOC and β in ACC, we fail to get α in NOM.
This means that the system derives Contiguity, because it cannot generate offending
paradigms. We thus leave this section with the conclusion that Contiguity follows
from the interaction of two proposals: (i) cumulative decomposition and (ii) the Superset
Principle and the Elsewhere Condition.
5 Case Contiguity and alternative approaches
This section considers two common approaches to case morphology—one based on
cross-classification, and one on feature geometries. I show that these alternatives fail
Explaining the structure of case paradigms by the mechanisms 1027
to capture linear contiguity—in fact, they provide virtually no constraints on possible
vs. impossible syncretism.
5.1 Jakobsonian cross-classification and its descendants
In Jakobson (1962), cases are represented as collections of equipollent features (like
[+/− marginal], [+/− ascriptive]), each of which cross-classifies the full set of cases
into two groups. This system has the immediate advantage that natural classes of
cases can be referred to with the help of such features, and syncretism can be restricted
to these classes.12
From the perspective of Case Contiguity, two formal properties of such systems
appear problematic (see also McCreight and Chvany 1991 and Johnston 1996). First,
the expressive power of cross-classification goes beyond Contiguity. To see that, consider
the cross-classification of 4 cases—NOM, ACC, LOC and GEN—by two features,
X and Y (24). The particular decomposition is proposed so as to match the facts of
Classical Armenian as close as possible. The natural classes definable by such a decomposition
are given in (25).
(24) Cross-classification
+Y −Y
+X NOM ACC
−X GEN LOC
(25) a. [+X]: {NOM, ACC}
b. [−Y]: {ACC, LOC}
c. [−X]: {LOC, GEN}
d. [+Y]: {NOM, GEN}
e. [Ø]: {NOM, ACC, LOC, GEN}
The natural classes (25a–c) are those that we need to capture the observation that
syncretism targets contiguous regions in the sequence NOM–ACC–LOC–GEN. However,
the syncretism in (25d) should be disallowed. Thus, the conclusion is that a
system of cross-classification predicts a linear contiguity constraint to be false. Equivalently,
the existence of a linear contiguity constraint proves such a model wrong.
The second problem is that the system also under-generates (compared to Contiguity):
as things stand, there is no way to define syncretisms of three terms (contiguous
or not). However, there is no known constraint to the effect that syncretism can target
sets of two and four cases, but not three, or any other similar restriction (see (5) for
examples of three term syncretisms).
To increase the generative capacity of the system in the right direction, crossclassification
is usually accompanied by a mechanism which incorporates the Elsewhere
Condition (18).
12Relevant recent references include (among many others) Halle and Vaux (1998); Bobaljik (2002); Müller
(2003); Embick and Noyer (2007); Calabrese (2008); Harley (2008).
1028 P. Caha
With the Elsewhere Condition in place, we get a three term syncretism by the
interaction of two rules, (26a, b).13
(26) a. [+X,+Y]: {NOM} → /phon A/
b. [Ø]: {NOM, ACC, GEN, DAT} → /phon B/
These rules (by themselves) pick out overlapping natural classes of cases: (26a) says
that NOM is realized by /phon A/, and (26b) requires that any member of {NOM, ACC,
GEN, DAT} is pronounced as /phon B/. The two rules clash for NOM. The winner
is determined by the Elsewhere Condition (18). Since /phon A/ applies in a proper
subset of cases compared to /phon B/, /phon A/ takes precedence for NOM. As a
result, NOM = /phon A/, ACC, GEN and DAT = /phon B/.
However, this is a deadly fix. The result is that any syncretism becomes possible.
The reasoning is this: the entry (26b) can, in principle, insert /phon B/ to any of NOM,
ACC, GEN, DAT. Consequently, we can get any triplet of these by assuming that any
one of the cases is spelled out by an entry similar to (26a). And we can also get any
pair by assuming (26b) and that two cases of our choice are spelled out by two rules
similar to (26a).
In conclusion, if Case Contiguity is correct, Jakobsonian cross-classification is not
an adequate model of case representation. Its shortcoming is purely formal, and holds
regardless of the content of features, and how the pluses and minuses distribute over
individual cases.
5.2 Feature geometries
A number of proposals add structure to the proposed features, and go thus in the
direction of the present account.14 For instance, adapting the proposal by Williams
(1981), we could device the following feature geometry for a part of the Armenian
system:
(27)
I will not comment on the content of the features, and note only a formal difference
between such a geometry and Jakobsonian cross-classification (24). In particular, the
feature which distinguishes among the direct cases ([independent]) is different from
the feature which distinguishes among the indirect cases ([local]). This is in contrast
with the cross-classification (24), where the feature which distinguishes NOM from
ACC [+/−Y] is the same that distinguishes between LOC and GEN. This difference
13The rule in (30a) reads as follows: the feature matrix [+X, +Y], corresponding to NOM, is realized by
the phonology /phon A/.
14Relevant references include Williams (1981, 1994); Wiese (2004); McFadden (2007).
Explaining the structure of case paradigms by the mechanisms 1029
represents a step in the right direction, because it was the feature [+Y] which allowed
for a non-contiguous syncretism to be singled out; see (25d). Thus, in (27), each
feature defines a set that forms a contiguous region in the syncretism sequence.
Looking into Williams’ system in more detail, we can think of lexical entries as
specified for non-terminal nodes of the feature tree above. A particular case is then
spelled out by the entry that is closest to it in the tree hierarchy. For example, NOMACC
syncretism is accounted for by an entry that targets the node [+direct]. Starting
from either of NOM/ACC, and going upwards, this is the entry that we encounter first.
In paradigms where NOM and ACC are not syncretic, lexical items are fully specified
and target corresponding terminal nodes.
Then, in order to get ACC-LOC syncretism, we must postulate an entry specified
for the root node [case], applicable in all four cases. If we then further postulate
tailor-made entries for NOM and GEN, ACC-LOC syncretism emerges.
With such a mechanism, however, any syncretism becomes possible once again.
To show that, let me keep the entry that targets the [case] node; this entry may apply
in any of the cases. With this entry in place, we get any triplet of cases by inserting a
distinct marker in one terminal node of our choice. In a similar way, if we have two
competing entries targeting the terminal nodes, any pair of cases can be syncretic.15
5.3 Summary
The previous two sub-sections have argued that two common feature systems intended
to account for syncretism fail to derive linear contiguity. The reason for this
are the formal properties of the feature representation employed. In particular, in these
systems, each case has an ‘exclusive address:’ the feature combination that characterizes
a given case neither contains, or is contained by any other case. In such systems,
‘default’ entries can be restricted in arbitrary ways by competing lexical entries targeting
individual cases, with the result that any syncretism may be generated.
The current system differs in precisely this aspect: there is no ‘exclusive address.’
For example, if we have an entry apparently tailor-made for GEN (A,B,C,D), this
entry also targets LOC (A,B,C), ACC (A,B), and NOM (A). Hence, the drawback of
the systems discussed is not replicated.
However, the current system is not the only response Contiguity. Pursuing a similar
agenda, other approaches have drifted away from features altogether (McCreight
and Chvany 1991; Wiese 2003), and proposed that paradigms are (spatial) objects
in their own right, corresponding to a linearly ordered set of cells. Markers are then
specified for contiguous sub-spaces of such paradigms.
15Formally, the system makes some non-trivial predictions concerning syncretism. For example, if NOM
and GEN are syncretic to the exclusion of other cases, then ACC and LOC cannot be. That is because
both NOM-GEN and ACC-LOC syncretisms can be only obtained by inserting an exponent under the [case]
node in the tree (27). Since only one marker can be inserted at a given node, the combination of the two
syncretisms is impossible.
However, as far as I am aware, there is no known constraint on syncretism that has such an implicational
format (if one type of syncretism is attested in a paradigm, then another type of syncretism cannot be, even
though it is independently attested). If that is so, and such predictions find no empirical support, this in
fact adds up to the drawbacks.
1030 P. Caha
The main reason I do not adopt such an approach is theoretical simplicity. As
we will see shortly, I argue that cumulative sub-classification can be implemented
in syntax, using mechanisms that are independently needed. To the extent that such
an approach turns out to be successful, it has the a priori advantage of making an
independent concept of a paradigm space theoretically unnecessary. That does not
mean giving up on the idea that paradigms are structured objects (as opposed to a
set of cells); what it means is that paradigm structure reduces to (or derives from)
syntactic structure.
6 The layered structure of case features
The general conclusion that I would like to build on in this section is that Case Contiguity,
if correct, helps us establish the formal properties that a case decomposition
should comply with. The crucial question is: Why should a case decomposition have
such properties?
6.1 The tree structure
The answer I suggest here is based on the observation that cumulative classification
is equivalent to the decomposition shown in (28), where individual features head
projections in a binary branching syntactic tree:
(28) [ins G [abl F [dat E [gen D [loc C [acc B [nom A DP ] ] ] ] ] ] ]
The tree encodes the proposal that a nominative DP is a type of syntactic constituent,
in which the DP is the complement of the feature [A]. The accusative is a similar
constituent, one which is built on top of the nominative ([A]) by the addition of [B],
so that ACC = [A, B]. And similarly for the other cases: LOC = [A, B, C], GEN =
[A, B, C, D], and so on. Thus, the proposal in essence says that the features A, B, C
etc., needed for syncretism, are the primitives of the syntactic structure, and they are
ordered in a functional sequence.
Note that the feature [B] is not ‘accusative.’ Accusative is the name of a constituent
which arises as the result of merging [A] and [B] on the top of the DP in this order. To
make this clear, I avoid calling the terminals ‘accusative’ but reserve that label only
for the non-terminal projections. The labels of the non-terminal nodes, such as acc
or gen are chosen for clarity of presentation, and they do not imply that the label is
qualitatively different from the head. I assume that the ‘true’ label of the accusative
constituent is BP, but I avoid referring to it in that way because such a label is quite
opaque.
Finally, in order to terminologically distinguish the representations (14) and (28)
(both are instances of cumulative decomposition), I call (14) flat cumulative decomposition,
and (28) layered cumulative decomposition.
6.2 The content of the features
In the preceding sections, I have stressed that what is at stake here are the formal
properties of the decomposition, rather than the actual content of the features. How-
Explaining the structure of case paradigms by the mechanisms 1031
ever, for the sake of completeness, I add a brief speculation as to what the content
might be.
I will be assuming a view on feature meaning such that each feature operates on
the semantics of its complement. In the prototypical case, each feature restricts the
denotation of its complement to a smaller subset (even though I make use of other
types of semantic contribution as well). Under this view, nominative will be the least
specific case semantically, with other cases getting increasingly more specific. With
this general background, I would like to distinguish four ‘zones’ in the proposed
functional sequence:
(29) a. Zone 1: NOM-ACC (structural cases)
b. Zone 2: LOC-GEN (stative cases)
c. Zone 3: DAT (goal case)
d. Zone 4: ABL-INS (source cases)
As indicated, I understand NOM-ACC as pure structural cases, compatible with a wide
range of interpretations. To keep things as simple as possible, I assume that the interpretation
of the feature [A] is ‘case.’ The feature [B] is then perhaps best understood
as ‘dependent,’ reflecting the hypothesis that only NOM may surface as the so-called
‘default case.’16 A potential problem for this simple view is that ACC in Classical
Armenian occurs in goal PPs. To make this fact compatible with the ‘structural’ approach
to NOM-ACC, I follow Abraham (2003) in his suggestion that the ACC in goal
PPs is actually verb-governed.
The second zone comprises two stative cases. LOC denotes static location, and
GEN a static possession (broadly understood). The reason why these two should be
the simplest of the semantic cases is connected to a consensus in the literature that
states are basic, and that change of state meanings are derived from them (see, among
many others, Kratzer 2000 and Ramchand 2008 for verbal semantics and Jackendoff
1983 for spatial expressions).
The idea for dative is then that the feature ‘E’ in (28) adds a change of state meaning
to the two stative cases, leading to a general goal reading. This subsumes the
prototypical dative use as a recipient case (i.e., change of possession case); see Caha
(2009:Chap. 5) for a defense of such an idea.17
Finally, ABL and INS are understood as source cases. The lower ABL covers not
only spatial source of motion reading, but also a more abstract ‘source of event’ reading,
corresponding to the passive agent. INS may then be understood along similar
16Schütze (2001) argues that English has a default accusative case, which in English is manifested only
on pronouns. However, all the examples that are used to establish this claim (coordination, left dislocation,
gapping, short replies) are contexts that exclude weak pronouns and require strong pronouns (see
Cardinaletti and Starke 1999). This makes it more likely that the data rather suggest that ‘he/she’ are weak
pronouns, whereas ‘him’ a strong pronoun, ambiguous for NOM/ACC.
An anonymous reviewer reminds me that he/she are available in (some) coordination structures
(whereas weak pronouns generally are not, although some counterexamples have been reported)—I leave
this for future research. The reviewer also notes that he/she can be contrasted and stressed; but these facts
are compatible with being a weak pronoun (Cardinaletti and Starke 1999:153–154).
17In a number of languages, the goal zone also comprises the allative (DAT=ALL is frequent). But recall
that the goal case in Classical Armenian is ACC. The suggestion is that the ACC is verb governed, and
Armenian directionals thus correspond to applicatives.
1032 P. Caha
lines (i.e., as a ‘source’ of the event), the distinction between agent and instrument
proper being both animacy and the controlled-by-agent restriction on instruments.
The reason why the source zone should be derived by the addition of features on
top of the goal zone is not immediately obvious, but finds an analogue in work by
Pantcheva (2010, 2011). On the basis of a cross-linguistic sample, Pantcheva discovered
that if a language shows morphological containment between spatial source and
goal, then source always contains goal. She explains that by proposing that source
paths are constructed by reversing the orientation of goal paths. If that is so, her proposal
explains why the source zone is more complex than the goal zone, with the
feature ‘F’ plausibly corresponding to the type of reversative semantics she proposes.
6.3 Case attraction as evidence for a layered structure
Leaving the content of the features aside, the general interest of the proposal is
twofold. Theoretically, the proposal entails that the organization of features inside
morphemes is governed by the same principles as the organization of phrases inside
sentences (i.e., a binary branching tree). The descriptive restriction on syncretism,
namely Case Contiguity, is then just a reflex of this deeper hypothesis concerning the
architecture of grammar (see Kayne 2005; Starke 2009; Cinque and Rizzi 2010).
In addition, we gain a number of empirical predictions. For instance, Caha (2011b)
shows that both across and within languages, movement may target various positions
in between such case features. This can only be so if (28) (as opposed to (14)) is the
correct representation.18
In this section, I provide Armenian-internal evidence from case attraction that favors
the layered structure (28) over the flat representations in (14). Specifically, as
noted in Sect. 3.2, case attraction in Armenian is related to the system of syncretism
in an intriguing way. I repeat the observation below:
(30) Case Attraction (Armenian):
a. In the ‘syncretism sequence’ below, cases to the right of the GEN have
the power to attract it, case to its the left cannot do so.
b. NOM – ACC – LOC – GEN – DAT – ABL – INS
In order to show how case attraction forces us to abandon flat structures, I first
present an analysis of this phenomenon as the combination of (i) agreement and
(ii) subsequent ellipsis. I portray this schematically in (31):
(31) Case attraction as agreement plus ellipsis
head noun-casei [[ dependent noun-GEN] AGR=casei ]
18In languages where the noun lands in between the case features, the features are split into two groups
and get realized as two separate morphemes (one preceding, one following the noun). Hence, such languages
also provide evidence for the proposed containment relations; see Caha (2011b) for a number of
illustrations for each containment proposed.
Explaining the structure of case paradigms by the mechanisms 1033
In words: GEN marking is elided because of the presence of a structurally higher
agreement marker, which bears identical case features as the head noun.19
The first reason for adopting (31) is that an apparently bizarre phenomenon is decomposed
into two well-known and independently attested processes. To see agreement
marking of possessors apply independently, consider the example in (32). It
comes from Guugu Jalanji, and shows that the possessor ‘Dick’ is marked both by
GEN (-ndamun), and by an AGR marker (-du) tracking the case of the head noun
(‘dog’):
(32) Dicki-ndamun-du
Dick-GEN-ERG
kaya-ngka
dog-ERG
‘Dick’s dog’ (Guugu Jalanji, Plank 1995)
Turning to Armenian itself, it can be shown that case attraction shares two characteristics
with such a process. First of all, attraction is local: it does not apply recursively
to the possessor of the attracted phrase. To see that, consider (13) again (repeated
below for convenience).
(33) [ bazmut‘-eamb
crowd-INS.SG
[ zawr-awk‘-n
force-INS.PL-DEF
Hay-oc‘
Armenian-GEN.PL
]]
‘with a crowd of the Armenian forces’ (Plank 1995:43)
In the example, the attracted genitive (zawr-awk‘-n) has itself a genitive dependent
(Hay-oc‘), which escapes attraction.
The same restriction has been observed for possessor agreement in Lardil by
Richards (2007). The relevant data are given below, the ACC -i spreads from the head
noun (‘boomerang’) to the head of its GEN dependent (‘older brother’), but not on the
dependent of the dependent (‘boy’):
(34) [ marun-ngan
boy-GEN
thabuji-kan-i
older.brother-GEN-ACC
] wangalk-i
boomerang-ACC
]]
‘(You broke) the boy’s older brother’s boomerang.’
Secondly, attraction appears to be optional. This is shown in (35), where both nonattraction
(35a) and attraction (35b) are available:
(35) a. i
by
knoˇj-ê
wife-ABL
t’agawor-I-n
king-GEN.SG-DEF
b. i
by
knoˇj-ê
wife-ABL
t’agawor-Ê-n
king-ABL.SG-DEF
both: ‘by the wife of the king’ (Plank 1995:20)
19I assume that agreement involves a separate base-generation of case features on top of the agreeing
constituent, GEN in our case. At LF, these features must match with the case features base-generated on
the head noun.
1034 P. Caha
This may now be seen in parallel to the fact that agreement is likewise ‘optional.’ In
Old Georgian, for example, the possessor does not need to agree with the head noun,
as in (36a), but it may do so, as shown in (36b):20
(36) a. saxl-is
house-GEN
pat.ron-ma
mistress-ERG
‘the mistress of the house’ (Plank 1995:7)
b. gwam-isa
body-GEN
krist-es-isa
Christ-GEN-GEN
‘of the body of Christ’ (Plank 1995:4)
Summing up, there are at least two properties that case attraction shares with clear
cases of possessor agreement (locality and optionality); this follows if case attraction
involves agreement as a crucial component.
Adopting such a view brings us to the question why the genitive marker does
not co-occur with the agreement marker (as it does in (34) and (36)). There are two
logical options: (i) the genitive has never been there, or (ii) it has been elided on the
surface. I argue here for the latter option.
Let me begin by showing that such an ellipsis process is independently needed. An
example is given in (37). To set the stage, (37a) shows that in Amharic, the possessor
is preceded by the possessive marker yä.21 This marker gets deleted when the whole
DP is preceded by the dative marker lä in (37b). (It gets deleted after other oblique
markers as well.)
(37) a. yä-Girma-n
of-Girma-ACC
wändimm
brother
‘Girma’s brother-ACC’
b. lä-yä-Girma
to-Girma
wändimm
brother
‘to Girma’s brother’ (Amharic, Baker and Kramer 2010)
Note that in Amharic, ellipsis is case sensitive: it does not apply in (37a), i.e., in
ACC (it does not apply in NOM either). The reason Armenian should be analyzed
along the same lines comes from the fact that attraction is also case sensitive (30).
The explanation for this emanates from an independent condition on ellipsis, which I
state in (38):
(38) Recoverability of ellipsis (case attraction):
The feature content of the elided morpheme (genitive) has to be recoverable
from the antecedent (the agreement marker).
20In Old Georgian, agreement is in fact not optional, but depends on the position of the dependent noun.
Simplifying things slightly, in post-nominal position, agreement applies; in pre-nominal position, it does
not. I assume that in Classical Armenian, the GEN noun also has access to two distinct positions, one
where agreement is obligatory and one where it is disallowed. The existence of such two distinct positions
is masked by the fact that the noun moves even higher up.
21The example also has an accusative marker -n, which marks the whole DP, and surfaces following the
first constituent of the DP.
Explaining the structure of case paradigms by the mechanisms 1035
The following table spells out the solution in detail. The first column reflects the proposal
that both the genitive case and the agreement marker are underlyingly present.
Combining this proposal with cumulative case decomposition, we arrive at the representations
in the middle column. Ellipsis then applies to the genitive, and produces
the output in the last column:
(39) An ellipsis account of the facts
As is evident, the condition (38) derives the observation that attraction does not happen
in NOM, ACC, LOC (the cases above the double line). That is because in these
cases, the proposed ellipsis cannot be recovered: the feature D finds no antecedent in
any of the cases.22
This approach thus correctly restricts the application of case attraction in nontrivial
ways, accounting for its locality, optionality, and case sensitivity. Importantly,
for the account to work, we need the initial proposal that case features are cumulative:
GEN must be contained in all cases to its right, and in no case to its left.
Further, the account leads to the conclusion that case features are layered. That is
because of an independently established conjecture that the antecedent of an ellipsis
must be a constituent (for textbook treatments see, e.g., Haegeman 1994 and Carnie
2008; see also Pesetsky 1995; Phillips 2003 and Baltin 2006 for relevant discussion).
If that is so, then the features of GEN must not only be contained inside those cases
which license its ellipsis, but they must form a constituent to the exclusion of other
features. This in turn forces us to abandon the initial non-layered decomposition (14)
as insufficient. In contrast, the facts are immediately compatible with (28).23
7 Phrasal spell-out
As a starting point, let me repeat that we have taken a step beyond cumulative decomposition
and added the proposal that each case feature corresponds to a separate
head in the tree (cf. Kayne 2005; Starke 2009; Cinque and Rizzi 2010). Theoretically,
this explains the formal properties of cumulative decomposition; empirically, we find
supporting evidence in phenomena such as case attraction.
22I assume that in NOM, ACC, LOC, Classical Armenian is thus forced to make use of the possessor position
where agreement does not apply.
23Admittedly, the facts do not provide evidence for such a rich structure as (28) posits. (We only have
evidence for GEN being a constituent inside DAT/ABL/INS.) But the point is that whenever we can find
evidence for layering, we do.
1036 P. Caha
But the model also raises questions concerning spell-out. In particular, since case
morphemes spell out more than one feature (Sect. 4.2), and each feature is a head
(Sect. 6.1), the model virtually necessitates a theory where spell-out may target more
than one head. In this section, I present a phrasal spell-out algorithm with such an
effect, drawing mainly on Starke (2009).24
The algorithm I describe takes the syntactic structure as its input, and provides the
phonological representation as its output. This simplifies the architecture of grammar
in the sense relevant for this paper: there is no intermediate morphological structure.
In later sections, I show that such a theory is not only simpler concerning its architecture;
it also captures subtle details of paradigm structure.
7.1 Three core assumptions
The proposal to be developed understands spell-out to be a cyclic bottom-up translation
of syntactic structure onto phonological representation. At each node, a PF is
constructed, and stored in memory until the next cycle. The intermediate PF representations
do not necessarily correspond to a part of what is pronounced in the end.
Intermediate representations may be scraped off, and replaced by a completely different
representation in case the algorithm requires that.25
Coming to the actual algorithm, the first assumption I make is that it scans the
structure bottom up, with lexicalization attempted at every node.
(40) Assumption 1: Spell-out proceeds bottom up.
Thus, in (41), spell-out targets Y and Z before YP, and it targets X and YP before
XP.26 It is not crucial whether Y is spelled before Z, or X before YP.
(41)
In Starke (2009), bottom-up spell-out is part and parcel of the cyclic spell-out hypothesis
(cf. Starke 2010; Caha 2011b; Pantcheva 2011). In particular, spell-out applies
after each step of Merge. And since Merge is standardly assumed to proceed bottom
up (Chomsky 1995), so does spell-out. In this paper, I simplify the model and ignore
the interleaving of Merge and lexical access for ease of presentation: I use a model
24Cf. McCawley (1968); Weerman and Evers-Vermeul (2002); Neeleman and Szendr˝oi (2007); Ramchand
(2008); Radkevich (2009); Taraldsen (2010); Pantcheva (2010).
25These stages thus do not correspond to unchangeable phases, a notion that may be close to what Chomsky
(2008:143) has in mind: “For minimal computation, as soon as the information is transferred it will
be forgotten, not accessed in subsequent stages of derivation [...] Working that out, we try to formulate a
phase-impenetrability condition PIC [...] PIC holds [...] for the mappings to the interface.”
The view assumed here comes closer to the one described in Šurkalovi´c (2011). Šurkalovi´c (2011:97)
proposes that “spell-out does not proceed in chunks but in concentric circles,” with the consequence that
“the input to phonology at each phase is cumulative, consisting of the spell-out of the current phase together
with the spell-out of the previous phases.” (Šurkalovi´c 2011:84).
26See Bobaljik (2002) for a related proposal.
Explaining the structure of case paradigms by the mechanisms 1037
where syntax produces a full-fledged output representation, which is then step-wise
translated onto phonological form.
Thus, for each tree derived by syntax, we go through the structure bottom up, and
construct a PF at every node. The PF consists of material eligible for insertion at
a given node. The principle which determines what entries are applicable is given
below; it is a refined version of the Superset Principle (15).
(42) Assumption 2: The Superset Principle (cf. Starke 2009):
A lexically stored tree matches a syntactic node iff
a. the lexically stored tree contains the syntactic node including the material
dominated by that node OR
b. the lexically stored tree matches all daughter nodes (ignoring traces).
Sometimes, an item matches both because of (42a, b), but it is enough that only one
condition obtains. Because of that, I save space and mention only one reason for
obtaining match in such cases. When an item matches a node, it represents the whole
tree dominated by that node at PF, and overrides any material that may have been
inserted earlier on.
I further assume that a successful lexical search sends to PF a set of items that
match the node. (If the set has more than one member, competition arises, with the
winner determined by the Elsewhere Condition.)
To see the workings of the system in an example, assume that in addition to the
structure (41), we have the lexical entries (lexically stored trees) in (43):27
(43) a. Z ⇔ /α/
b. YP ⇔ /β/
c. Y ⇔ /γ /
d. X ⇔ /δ/
How is the structure (41) spelled out using the lexical entries (43)? We start from
the node Z, and determine the set of entries applicable for that node. This set comprises
of (43a, b), since each contains the node Z. Lexical search thus produces an
intermediate stage of PF derivation: the set {α, β}.
As a next step, spell-out targets Y. Y is contained in the entries (43b, c), and
hence, we obtain the set {β, γ }. At this point, we have the PF representation in
27In general, I assume that lexical entries are triplets phonology, syntax, concept . Since the conceptual
contribution of case affixes is negligible (if present), I ignore it here.
Further, I assume that lexical entries are links between phonological, syntactic and conceptual representations,
each representation a well-formed object in the relevant module. This answers some frequent
questions about why there should be syntactic trees ‘in the lexicon’ (see (43b)), and related worries about
the role of syntax and the lexicon in constructing expressions. From the linking conception, it follows that
the lexicon has no means to create structures of its own. Consequently, it follows that the only possible
target of lexicalization (when it comes to the syntax part of the lexical entry) is a well-formed syntactic
structure. Thus, for example, [X, Y, Z] cannot be a part of a lexical entry, because syntax cannot construct
ternary branching trees (cf. Kayne 1994; Chomsky 1995). Thus, despite the fact that the lexicon stores
outputs of syntax, it does not duplicate it; the lexicon only links the outputs of syntax to representations in
other modules.
1038 P. Caha
(44a) (two unordered sets). For reasons of clarity, I will be using the representation
in (44b) to represent such a translation. The advantage of (44b) is that it shows how
the representation (44a) was derived.
(44) a. {β, γ }, {α, β}
b.
Spell-out then targets YP. Since the entry for {β} matches both daughters of YP
(see (44b)), YP is also matched by {β}, yielding the representation (45a), which may
be by convention represented as in (45b):
(45) a. {β}
b.
This shows that when lexical search produces an item that matches a phrasal node,
the whole tree is represented by this item at PF, and all material previously inserted
below that node is deleted from the PF representation. Consequently, in the course
of spell-out, β won as the spell-out of YP over a combination of two items, say the
sequence γ > α. Starke (2009) calls this effect “The Biggest Wins Theorem,” since
the entry with the biggest tree in the lexicon wins. Empirically, we get the effect that
portmanteau morphemes win in competition over agglutinative forms. To mention
just a couple of examples: off wins over *from on; went over *go-ed; in Armenian,
as discussed in more detail below, the ABL.PL c‘ wins over ABL+PL ê-k.
Spell-out then targets X, with the only match {γ } introduced by (43d). At this
point, we have the two unordered sets in (46a), which corresponds to the conventional
representation (46b):
(46) a. {γ }, {β}
b.
There is no match for the whole XP; hence, there is no ‘overriding.’ Rather, the
translation procedure at XP introduces the mutual ordering of {γ }, {β}. Since the
feature spelled out by {γ } c-commands those spelled out by {β}, the phonological
representation orders the sets of (46a) as γ > β. Rewriting of previously inserted
material and ordering of sets are nicely visible in the tree representations, and I am
going to use such representations exclusively.
My final assumption concerns situations where the spell-out algorithm does not
provide a unique candidate for spell-out. Thus, coming back to our initial example in
(41), suppose that we have two additional entries:
Explaining the structure of case paradigms by the mechanisms 1039
(47) a. ⇔ / /
b. ⇔ /ζ/
The change this brings into the picture is the following. Since each of the terminals
in (41) is contained in the entries in (47), we get two more competitors for each node.
Consequently, we end up with (48), instead of (46b):
(48)
When spell-out targets XP, we obtain the set { , ζ} (each matches both daughter
nodes). The unique spell-out of XP is thus not determined by the principles we have
introduced so far. What we need is the Elsewhere Condition:
(49) Assumption 3: The Elsewhere Condition: In case two rules, R1 and R2, can
apply in an environment E, R1 takes precedence over R2 if it applies in a
proper subset of environments compared to R2.
In our case, the rule introducing applies in a proper subset of cases compared to
the rule introducing ζ (ζ can apply wherever does, but is inapplicable for W,
and WP). Hence, wins in competition for the spell-out of XP. Importantly, I assume
that Elsewhere computation applies only where indispensable, i.e., in cases where the
‘Biggest Wins Theorem’ has not produced a unique candidate.
Recall that the Elsewhere Condition and the Superset Principle were introduced
in Sect. 4.2, and they are necessary to derive Contiguity. Incorporating these assumptions
replicates the result in a phrasal spell-out system.
7.2 Two consequences
Let me now turn to two consequences of the proposed model that will turn out to be
important in accounting for various interesting phenomena in Classical Armenian.
(50) Consequences
a. If a phrase XP is spelled out by a single morpheme in a language L, any
sub-structure of XP is spelled out by a single morpheme as well (in that
language).
b. Insertion may only target features that form a constituent.
To see the consequence (50a), consider the representation in (51). Suppose that in
this tree, XP is spelled out by α:
1040 P. Caha
(51)
In order for that to be possible, either (52a) or (52b) has to hold, because of the
Superset Principle, see (42):
(52) a. The lexical entry for α contains the whole tree dominated by XP, and
hence, it contains also (the tree dominated by) Y and Z.
b. The lexical entry for α matches both Y and Z.
From (52), it is easy to see that for every XP, the item which matches that XP matches
also its immediate sub-constituents (Y and Z), and will be used to spell out these
nodes (unless there is a better match due to the Elsewhere Condition). We may now
repeat this reasoning recursively for Y and Z: since α matches these nodes, it also
matches their immediate sub-constituents, and so on, until we reach the terminal
nodes. In conclusion, if XP may be spelled out by a single morpheme α in a particular
language, then any substructure will correspond to a single morpheme, because α is
always a match. This holds regardless of whether α actually surfaces (there may be a
better match).
To see the corollary (50b), suppose that we modify the example structure (41) by
introducing a complement to Z, pronounced as phon X:
(53)
(54)
In (53), Z may be replaced by the set {α, β}, and Y by {β, γ }, using the entries
in (43). The result is shown in (54). Importantly, YP in (54) cannot be replaced by
{β}, despite the fact that {β} matches both Y and Z. That is because in YP, {β}
matches only one daughter (Y), but not ZP. This in turn is caused by a combination
of two factors. First, the entry for β (see (43b)) does not contain a tree identical
to the ZP of (54), and second, not all daughters of ZP are matched by {β} (QP is
not). Hence, even though Z and Y are both contained in a single entry, they cannot
be spelled out by that entry if they do not form a constituent to the exclusion of
other features. I turn to this prediction in Sect. 10, and show that it explains the puzzling
phenomenon of morpheme splitting. The general conclusion from such facts is
Explaining the structure of case paradigms by the mechanisms 1041
that since the constituency of individual features that make up morphemes matters
for spell-out, we gain evidence for positing such structures inside individual mor-
phemes.
Finally, I note that if QP had extracted out of YP in (54), then β would match YP
once again. The reasoning is this: since β matches Z, it also matches ZP where Z is
the only non-trace daughter (recall that QP extracts). Further, since β now matches
both ZP and Y, it also matches YP. This is once again related to the statement that Y
and Z can be spelled out as one morpheme just in case they form a constituent. After
QP extracts, Y and Z form a constituent to the exclusion of other features, and hence,
they may be lexicalized together.
The fact that insertion ignores traces has a consequence that will become relevant.
In particular, insertion at a node that dominates a trace produces the same output as
insertion at its non-trace daughter. Recall the example: when ZP immediately dominates
a trace, insertion at ZP produces the same results as insertion at Z. Hence,
whether ZP is present in the tree has no effect at PF. I use this corollary (55) later on
to simplify complex representations when focusing on their spell-out.
(55) The tree simplification convention: Nodes that dominate a trace may be
skipped in spell-out representations.
In the remainder of the paper, I turn to the consequences in (50), and show their
validity on empirical material.
8 The hierarchy of agglutination and fusion
I start with the prediction (50a): if an XP is expressed by a single morpheme, any
sub-structure corresponds to a single morpheme as well. By combining this prediction
with cumulative decomposition, we obtain an interesting testing ground. In particular,
since one case represents a substructure of another case, we predict that if a case X is
expressed together with some other inflectional category, then all cases contained in
X also fuse that category with case.
This prediction is nicely borne out in Armenian, as I argue below. Specifically,
I focus on the plural declension and argue that the INS.PL has separate markers for
class, plural and case. These categories are, however, subject to fusional expression in
other cases. Specifically, NOM, ACC and LOC show only a single marker, while GEN,
DAT and ABL are characterized by two markers: a class marker, and a number/case
portmanteau. I summarize these claims in the table (56), where shading indicates
synthetic expression of given categories:
(56) The template for Armenian plural declension
1042 P. Caha
The interest of (56) lies in the observation that the same sequence which restricts
syncretism and case attraction governs also the agglutinative vs. fusional expression
of categories. I state this in (57):
(57) Agglutinative vs. fusional spell-out (Armenian):
a. In the sequence below, if a given case is expressed together with some
other category, then all cases to its left are as well.
b. NOM – ACC – LOC – GEN – DAT – ABL – INS
The current section establishes the empirical generalization (57), and the next section
derives it from the three spell-out assumptions.
8.1 Case and number
I start by taking a closer look at the singular/plural distinction. In most cases, the
singular and plural endings are different. This can be understood if (most) case exponents
also spell out number. The exception to this is the INS.PL, which is built on
top of the INS.SG by the affixation of -k‘, a morpheme which also shows up in the
NOM.PL.
(58) The analytic nature of the instrumental plural
A reasonable hypothesis is that -k‘ marks plural. This finds confirmation in the verbal
paradigm, where -k‘ differentiates the 1.SG and 1.PL agreement (59a), and in the
composition of the 2nd person pronouns (59b):
(59) a. sirem
love.1.SG
–
–
sirem-k‘
love.1-PL
b. du
you.SG
–
–
du-k‘
you-PL
Turning to NOM.PL, a question arises whether it is expressed by one or two morphemes.
Under one possible hypothesis, -k‘ is just plural, and NOM is -ø. I call this
the agglutinative analysis. An alternative hypothesis (which I call the fusional analysis)
is that -k‘ spells out both number and case.28 These analyses are hard to tease
apart, but there is indirect evidence in favor of fusion.
28This analysis presupposes the Superset Principle; when -k‘ acts as a pure plural marker, it spells out a
sub-constituent of its specification.
Explaining the structure of case paradigms by the mechanisms 1043
First of all, the agglutinative analysis requires an additional (null) morpheme. How
serious a drawback that is depends on the details of implementation; but as a general
methodological guideline, we should favor an analysis that does not postulate an
additional (invisible) entity, if it can be dispensed with.
Secondly, if -k‘ is just plural, we end up with a kind of split between agglutination
(NOM, INS) and fusion (other) that does not seem to be attested anywhere else
(basing myself on the survey in Plank (1999) and other languages I have looked at).
Specifically, in languages with such splits, a particular marking strategy always targets
contiguous regions on a case hierarchy (usually grammatical vs. oblique). On the
agglutinative analysis, Classical Armenian would be the only counterexample to this
generalization. And since this unlikely typological feature is an artifact of an analysis,
it seems that an alternative analysis (which lacks this particular consequence) is
preferred.
Third, it is not only the case that a particular marking strategy (agglutination vs. fusion)
occupies a contiguous region on a case hierarchy, there is a virtually unexceptional
bias towards fusional expression in the NOM (see, once again, Plank 1999 for
an overview).29 As an example, consider the Mordvin definite declension, discussed
in McFadden (2004).
(60) Mordvin Definite Declension (Er´za dialect)
What we see here is the stem kudo- ‘house’ inflected for number, case and definiteness.
Almost all cases split case from definiteness/number morphemes (these are in
small caps). The exception to this is NOM.SG, where case is expressed together with
these categories as the portmanteau -´s.30
Going back to Armenian: assuming the fusional analysis of NOM.PL, case and
number are expressed separately in the most marked case (INS), and fuse in other
cases:
29Plank (1999) finds this pattern in Brahui, Finnish, Karelian, Chukchi, Yawelmani, Tunica, Romani,
Mordvin, Estonian, Tocharian A and Wakhi. I add the Prizren-Timok dialect of Serbian (Caha 2011b).
30I have to mention that Plank (1999) identifies an additional minority pattern where—as Plank argues—
there is agglutination in the grammatical cases, and fusion in the semantic cases. This pattern represents a
counterexample to the theory presented here; but there are reasons to doubt its existence.
According to Plank, this pattern arises in Chechen, Archi, Georgian and (interestingly) Classical Armenian.
Including Classical Armenian on the list of languages that exhibit agglutination in the grammatical
cases (specifically NOM) makes it obvious that Plank’s conclusions rest on the agglutinative analysis
(NOM=-ø+k‘). Once scrutinized, it turns out that 3 out of the 4 languages (Chechen, Archi, and Classical
Armenian) are listed as agglutinating on the basis of an analysis that assumes a zero morpheme. This is a
rather weak evidence.
Thus, what remains from the minority pattern under scrutiny is a single example: the so-called archaic
plurals in Modern Georgian. I have to leave this example for future research.
1044 P. Caha
(61) Case and Number
8.2 Class markers
Now compare the plural paradigm we have looked at with other plural paradigms:
(62) The plural declension in Classical Armenian
The comparison reveals a separate vocalic element between the root and the morphemes
-c‘ and -w. This suggests that we are looking at a morpheme, but what is this
morpheme? Starting from the observation that its quality is determined by the root,
whereas the quality of the plural -k‘ is not, it is attractive to analyze this morpheme
as originating locally to the root. As for its identity, Halle and Vaux (1998) take it to
be a theme marker, a type of a nominal classifier. I adopt this approach here as well.
But why is the class marker absent in NOM, ACC and LOC? I suggest that in these
cases, class is expressed together with number and case. In other words, -s and -k‘
spell out also class. This analysis is depicted in (63) in abstract terms, and a breakdown
of concrete paradigms is given in (64):
(63) The template for Armenian declension
(64) Case, Number and Class
31This form apparently lacks the instrumental -w/v. That is due to phonology: the class marker u and the
instrumental -w fuse into one segment. I argue for this later in this article.
Explaining the structure of case paradigms by the mechanisms 1045
To conclude: if the morphological analysis depicted in the table above is correct,
then the generalization (65) is established:
(65) Agglutinative vs. fusional spell-out (Armenian):
a. In the sequence below, if a given case is expressed together with some
other category, then all cases to its left are as well.
b. NOM – ACC – LOC – GEN – DAT – ABL – INS
The next section derives the generalization from the spell-out algorithm introduced
in Sect. 7.1 operating on the proposed functional sequence of cases.
9 Deriving the splits between agglutination and fusion
In Sect. 7.1, I introduced an algorithm that maps syntactic constituents onto their
pronunciation. Thus, when number and case are expressed as one morpheme, this is
encoded by spelling out a constituent containing these two categories. Knowing the
constituent structure is thus essential for carrying out the analysis; and I turn to this
issue now.
9.1 The order of morphemes
According to the view presented here, morphemes are either heads or phrases. If that
is so, we have two tools to take into consideration when we derive morpheme orders:
both phrasal and head movements may be involved in the derivation of the Armenian
inflected nominal.32
One way in which phrasal movement differs from the traditional head movement
(Travis 1984; Baker 1988) is that the head of the extended projection, the noun in our
case, can move across two heads without inverting their order, as in (66a), in violation
of the Head Movement Constraint. A derivation which obeys the Head Movement
Constraint is shown in (66b), and the movement of N across X inevitably leads to the
inversion of X and Y (excorporation aside).
32In the ideal case, head-movement is treated as a special instance of phrasal movement. The analysis
presented here does not allow such a reduction. The issue how to replace the type of head-movement I
employ in this paper is subject to ongoing research in Nanosyntax.
1046 P. Caha
(66) a. Phrasal movement:
b. Head movement:
In Classical Armenian, a theory based solely on head movement leads to a wrong
prediction. To see that, consider the base-generated order of the markers in question.
I follow the literature and adopt the hierarchy in (67), where the NP is dominated
by the projections of the Classifier (Cl), Number (Num, see Borer 2005), and finally
case (K, see Bittner and Hale 1996).33
(67) [ K [ Num [ Cl [ N ] ] ] ]
We know that the Noun has to move higher than K, since K is a suffix; head movement
then yields the sequence (68), which is empirically wrong: Num follows K in
Classical Armenian (recall -a-w-k‘ CL-INS-PL).
(68) *N-Cl-Num-K
What we need is a derivation involving phrasal movements (see also Koopman and
Szabolcsi 2000; Julien 2007; Muriungi 2008). The simplest derivation (granted the
general approach originating in Kayne (1994)) is shown in (69).
33A reviewer suggests that arbitrary thematic vowels are hard to account for ‘in syntax,’ and require the
existence of a specific morphology module—contrary to the architectural claims made here. I do not agree
with this conclusion for reasons I sketch below.
In particular, the selection between a particular root and an adequate thematic vowel can be treated as
idiomatic (in a technical sense). Space considerations prevent me from developing a full fledged account
of idioms, and I give only its bare bones. In Nanosyntax (cf. Starke 2010), idioms like kick the bucket are
treated as phrasal lexical entries that make reference to other entries. Thus, when a bottom-up translation
derives a constituent with its constituent parts spelled out as kick the bucket (as opposed to kick the pot),
that triggers the insertion of the concept DIE. Similarly, the Armenian lexicon contains a number of idioms
that refer to a constituent consisting of the root and a particular thematic vowel (cf. Marantz 1995). Noting
that there are idioms that contain words which do not make sense out of that idiom, roots on their own may
or may not have interpretation, depending on whether any conceptual information is associated to them
outside of the particular idiom. Assuming an account along these lines, the pairing of roots and thematic
vowels requires a specific morphological component no more than phrasal idioms do.
Explaining the structure of case paradigms by the mechanisms 1047
(69)
First NP and Cl invert. Then we add Num, which is crossed by its complement,
leading to N-Cl-Num. Num must end up last, and hence stays in situ. I put it in
bold. After we add K (specifically, the relevant number of heads), only ClP moves
across K, leading to N-Cl-K-Num. This is the correct order.34
I assume two additional movements that do not lead to surface re-ordering, but
change the constituency. First, I assume that case features undergo head-movement
from A to the highest K head, see (70). This creates a constituent that contains only
case features, but no number feature. This constituent is needed in the INS.PL, where
-w targets such a constituent.
(70)
Combining the movements in (69) and (70) produces graphical clutter. To remove
it, I use the tree simplification convention (55) and skip the case nodes whose head
evacuates by head movement, yielding (71). Recall that such a simplification has no
effect on the actual PF.
(71)
34I assume that the NP is head final, and may contain adjectival modifiers. Note as well that pre-nominal
adjectives do not carry agreement in Armenian.
1048 P. Caha
Note further that in (69), the class feature(s) do(es) not form a constituent with case
and number to the exclusion of the head noun. Since we know from (64) that class
may be spelled out together with case and number, we need this constituent. Thus,
I assume that the head NP sub-extracts from the final landing site of the ClP. Disregarding
traces, this will leave a series of non-branching projections in the Spec
of KxP. Due to the particular formulation of the Superset Principle I have offered,
anything that can spell out the Class head will also spell out ClassP, and hence, I abbreviate
the structure into (72). This structure will be used in the remainder of this
paper as a general template for insertion.
(72)
9.2 Packaging of categories by means of phrasal spell-out
Let me now turn to the splits between agglutination and fusion, as attested in Armenian.
In particular, we need to derive (73) (repeated from (64)). In the table, I abstract
over the quality of the theme, and abbreviate it as V.
(73) Case, Number and Class
In order to show how (73) is derived, I first write down the lexical entries of the
relevant markers. Then I go step by step through the structure of the individual cases.
Once we see how things work, I turn to some general conclusions.
(74a, c) give the entries for -k‘ and -s. These entries are able to spell out all of class,
case and number, and they differ in the number of case features they possess. Note
that in the entry of -s, I omit the nodes that dominate traces due to head-movement
of A to the highest K head, i.e., C. This is done in order to allow for easier parsing of
the structure, using the convention (55).
Explaining the structure of case paradigms by the mechanisms 1049
(74) a. -k‘ ⇔
b. -c‘ ⇔
c. -s ⇔
d. -V ⇔
e. -w ⇔
For GEN/DAT/ABL, we need two additional entries, introduced in (74b, d). The
marker -c‘ is specified for the case features A-F (representing the ABL), but unspecified
for class. In (74d), I use the generic V for class markers, and ignore their differences
(see fn. 33). Finally, -w is a pure case marker, specified for the complex INS
head (A-G), but no class and number, see (74e).
With the entries in place, I turn to the spell-out of the NOM.PL. Its structure is in
(75b). Starting from Num0 (representing plural), we find three matching entries: {-k‘,
-s, -c‘}. Since the algorithm ignores traces, the same set matches NumP, because it
matches its only daughter. Thus, I replace NumP in (75b) by the relevant set in (75a).
The order of elements in the set reflects their specificity; in case competition arises,
the leftmost element of the set wins.
1050 P. Caha
(75) a.
b.
A0 is matched by {-k‘, -s, -c‘, w}, because these entries contain A0. I also assume
that singular case markers in general may spell out A0, and I add them into the set
for completeness, even though—as we will see—they will never get inserted. Hence,
I replace A0 of (75b) by the set {SG, -k‘, -s, -c‘, w} in (75a), where SG stands for the
set of singular case markers.
Turning to the complex class node, we may note that it is contained in all the
class vowels, as well as in -k‘ and -s. The set comprising these elements replaces the
phrasal class node in (75b) to yield the last ingredient of (75a).
Now both daughters of Nom in (75a) are matched by {-k‘, -s, -c‘}, hence, this
set also matches Nom . The whole NomP is then matched by {-k‘, -s}, since these
elements match both daughters. With no more projections to spell out, competition
arises with the result that -k‘ is selected on the basis of the Elsewhere Condition. (It
applies in proper subset of cases compared to -s.)
The next case I discuss is the accusative (76b). The only difference in the spell-out
procedure (compared to (75)) concerns the case node. Here, once again, the set {SG,
-k‘, -s, -c‘, w} matches the A head (replacing it in (76a)). Proceeding to B0, the set
{SG, -s, -c‘, w} applies. As a result, the complex case head can be spelled out by {SG,
-s, -c‘, w}.
(76) a.
b.
Proceeding further upwards, {-s, -c‘} match Acc , since they match both daughters.
{-s} then comes out as the only candidate for the whole AccP.
Explaining the structure of case paradigms by the mechanisms 1051
In general, every exponent that matches the AccP in (76) matches also the NomP
in (75). This is the property of the system from which the generalization about agglutination
and fusion follows. Since (i) in the ACC, the constituent containing class,
case and number receives a single marker, and (ii) every marker that matches AccP
matches also NomP, it follows that there is such a portmanteau marker available in
the NOM as well. If the same marker surfaces, we get syncretism, if there is a better
match, we get a different marker. But no matter which of these two situations obtains,
the generalization is derived.
Let me now skip GEN/DAT and proceed directly to ABL, shown in (77b).
(77) a.
b.
There is no difference in matching for either ClP or NumP compared to the previous
scenarios. I will take a shortcut through the insertion at the complex case node,
and focus on the complex head as a whole. It is clear that neither -s or -k‘ can spell it
out (neither contains the features from D upwards). What we are thus left with is -c‘,
as well as -w, plus the singular ablative marker -ê (for which see (58)). Thus, I replace
the complex case head with the set {-ê, -c‘, -w}, see (77a). Spell-out then targets Abl .
The entry -c‘ matches both daughters, and it emerges as the sole possibility for the
spell-out of this node. Note that due to the way spell-out proceeds (recall the Biggest
Wins Theorem), we correctly exclude the possibility that ABL.PL gets spelled out as
the combination of the ABL.SG -ê with -k‘.35
Note as well that for any case smaller than ABL, -c‘ is in the competitor set for
that node. Since -c‘ is also a competitor for the spell-out of NumP, then in any case
smaller than ABL, -c‘ is going to be a candidate for a fusional spell-out of number
and case. From that, it follows that fusion of case and number in the ablative forces
their fusion for any smaller case.
Finally, spell-out reaches the AblP node. There is no match, and hence, the two
daughter sets are ordered: {V, -k‘, -s}> {-c‘}. In the first set, a thematic vowel wins
in competition. This correctly yields V>-c‘.
Turning now to the INS.PL (78b), there is only a single candidate for the spell-out
of the complex case node: -w. (78a) shows the result of matching:
35A virtually identical situation obtains in the GEN/DAT.PL, since -c‘ is the only candidate for the spell-out
of both the case and number branch of the bar node.
1052 P. Caha
(78) a.
b.
At Ins , there is no match. The algorithm thus orders the daughter sets as {-w}>{k‘,
-s, -c‘}. Due to competition, we arrive (correctly) at -w-k‘. At InsP, the ordering
between the class marker V (the most specific of the three eligible) and w-k‘ is introduced,
yielding V-w-k‘.
9.3 Agglutination and Fusion vs. the Subset Principle
This section turns to the following question: can we formulate an alternative analysis
using the Subset Principle? The claim of this section is that this is impossible; but
there are qualifications to be made along the way.
Most importantly, the analysis should deliver the splits between agglutination and
fusion “for free,” because this is what has been achieved in the previous section using
the Superset Principle. However, the standard way to deal with such splits in syntaxoriented
frameworks such as Distributed Morphology (see, e.g., Embick and Noyer
2007) is by recourse to rules which either fuse (Fusion) or split (Fission) positions
of exponence (syntactic terminals). For example, Halle and Vaux (1998) account for
some of the facts discussed above by saying that (i) number and case represent a
single node in syntax (that is why number and case are fused most of the time); and
(ii) a special rule (Fission) creates a separate number position in the INS.PL (that is
why the INS.PL is agglutinative). At a general level, it is assumed that the paradigm
structure is rigid: each case has the same number of morphological slots to fill. If
there are too few or too many markers, additional rules are invoked.
The approach suggested here is different. There is no definite number of slots to
fill; there are only features that need spell-out. If there are comparatively few of them
(as in the NOM), one marker may do for the whole lot. If there are comparatively
many features (as in the INS), three markers may be needed. The borders between
morphological slots are not cast in stone, they are fluid and emerge in the process
of insertion. As a consequence, there is no need for special rules to fuse or split the
assumed morphological positions. And, of course, once special morphological rules
are gone, the need for a special morphological representation evaporates.
The conclusion that the Subset Principle is inadequate as a tool in modeling
phrasal spell-out presupposes that the view entertained here is what the theory should
achieve. The claim is not that the Subset Principle is inadequate for the purposes it
Explaining the structure of case paradigms by the mechanisms 1053
is generally used; it is inadequate if one wants to extend its use beyond what it was
designed to handle.
In presenting the argument, I further take for granted that the contiguity restriction
(plus case attraction and agglutination/fusion) shows that cumulative classification is
on the right track. I now proceed to the reasoning.
The starting point is that under a Subset Principle approach, the marker -k‘ is just
plural, and only that. It cannot be specified for class, or case, because it shows up in
contexts where case and noun class are not present—recall sirem ‘I love’ vs. sirem-k‘
‘we love’ (59). In addition, there is the fact that in the INS.PL, -k‘ co-occurs with case
and class (so it apparently does not spell them out). Assuming this, then in order to
ensure that in the NOM.PL, we do not see the class marker -a co-occurring with -k‘,
-k‘ has to be allowed to ‘extend upwards’ in the tree (75), swallowing everything on
the way. This is consistent with the Subset Principle based reasoning: since Pl is a
subset of the NOM.PL tree, -k‘ may spell out the whole tree.
Why does -k‘ not surface in the ACC.PL in (76)? Since PL is a subset of ACC.PL, -k‘
is applicable. One is then led to propose that it is blocked by -s, specified as ACC.PL.
(Because of such a specification, -s is inapplicable in NOM.PL: ACC.PL is not a subset
of NOM.PL.) -s also spells out the locative, because ACC.PL specification represents
a subset of it. Problems arise when we focus on the spell-out of GEN/DAT/ABL.PL.
Taking the ablative as an example, see (77), I start by noting that -s must be a match
for this constituent, following the general logic of the Subset Principle. Since it does
not appear here, we must block it by c‘, which is then specified as GEN.PL, so that
-c‘ actually appears all the way from GEN to ABL. But what cannot be captured is
the fact that the class marker -a emerges when -c‘ is inserted. Somehow, this has to
be baked into the difference between -s and -c‘. The only way to do that seems to
be specifying -s for class, but leave -c‘ unspecified for class. (That is why the class
marker appears with -c‘ but not with -s.) But then -k‘ must also be specified for class
(since neither -s or -k‘ co-occur adjacent to class markers). However, this contradicts
the initial observation that -k‘ cannot be specified for class.
At a general level, the problem is that if the unmarked case is expressed by a single
marker, and that marker is allowed to spell out super-structures (the essence of the
Subset Principle), then all more complex cases should have fusion as well (because
the unmarked case corresponds to their subset). In other words, the generalization
comes out the other way round: if we get fusion in an unmarked case, we should get
it in all marked cases.
Because of this, the only way to get the facts right seems to be making INS the
least marked case (the feature A), and NOM the most marked case (features A-F). But
problems arise even on this view, as I argue below. Starting with the now unmarked
INS.PL (now actually corresponding to the structure (75)), we specify -k‘ for plural,
-w for the feature A (the sole feature of the INS) and the class markers spell out the
class node. The fact that -c‘ does not appear as a case-number portmanteau in the
INS.PL can then be encoded by specifying it as ABL.PL (now corresponding to the
features A+B+Pl, i.e., a part of the tree in (76)). If that is the specification, then -c‘
is not a subset of the unmarked INS. Further, because ABL is a subset of all cases safe
the INS, -c‘ will further be correctly eligible for insertion in all such cases, deriving
the observation that fusion of case/number in the ABL entails their fusion from ABL to
1054 P. Caha
NOM. The fact that -c‘ actually appears only in DAT/GEN.PL is then another instance
of blocking. In particular, the LOC/ACC.PL -s is going to be specified as LOC.PL
(features A-E+Pl), which makes it a better competitor in case both -c‘ and -s apply.
In addition, if class is added to the entry of -s, then no class marker will appear from
LOC.PL to NOM.PL, i.e., the range of cases where -s may apply.
The place where the system breaks down is NOM.PL. In order to prevent -s to appear
here, we have to block it by a specific NOM.PL marker (features A-F+Pl+Class).
But the marker which actually appears there is -k‘, which we already have listed in
the lexicon as just [Pl]. With such a specification, however, it cannot block -s from
occurring there. And if we change the specification of k‘, we cannot insert it under
Pl only, as required for INS.PL. We may then conclude that even under the reversal
of the markedness hierarchy, we cannot give a unified account of the facts using the
Subset Principle.36
9.4 Conclusions
This section has derived the observation that the ‘syncretism scale’ restricts fusional
vs. agglutinative expression of morphological categories (each strategy forms a contiguous
region with fusion at the unmarked end). This fact follows from the way cumulative
classification interacts with the consequence (79a). Importantly, the actual
splits between agglutination and fusion are successfully derived only with reference
to the underlying syntactic representation (and the actual lexical entries).
(79) Consequences
a. If a phrase XP is spelled out by a single morpheme in a language L, any
sub-structure of XP is spelled out by a single morpheme as well (in that
language).
b. Insertion may only target features that form a constituent.
10 The constituency of features
In this section, I turn to (79b). To see its teeth in a simple example, recall the structure
of the Armenian noun (72). In this structure, there is no constituent that contains
number and class to the exclusion of case. Hence, there can be no class/number portmanteau
to the exclusion of case. This is borne out: when number marking is separated
from case (the -k‘ in INS.SG), it is insensitive to class. When class is separated
from case (various vowels in INS.SG.), it is insensitive to number (the same vowel
emerges in INS.PL, see (58)).
36In addition, the explanation of Case Attraction in terms of ellipsis disappears under the markedness
reversal.
Explaining the structure of case paradigms by the mechanisms 1055
The logic of the prediction allows us to account for more subtle effects. Suppose,
for example, that two categories, α and β in (80a), can be spelled out by a single
portmanteau morpheme P. The entry of P is then as in (80b). Now if another category,
γ in (80c), structurally intervenes between α and β, α and β cannot be spelled out
by P. That is because α and β do not form a constituent in (80c).
(80) a. ⇒ P
b. /P/ ⇔
c. *⇒ P
What happens instead is that in (80c), each of α and β is spelled out on its own
(leaving aside an even more specific portmanteau that contains γ ). I call this effect
morpheme splitting: features that may correspond to a single morpheme in a given
language must be spelled out by two pieces when they are separated by an intervener.
I argue that morpheme splitting explains certain puzzling facts concerning consonantal
declensions in Armenian.
Morpheme splitting also has consequences for the architecture of the spell-out system.
In particular, if the effect is real, it demonstrates that the expression of syntactic
features by phonological material is not driven by some kind of a simple economy
condition, such that one merely tries to minimize the number of exponents. Instead,
morphemes are sensitive to structural configurations of their features. If that is so,
we gain additional justification for proposing that morphemes are structured enti-
ties.
10.1 Morpheme splitting in a-stems
Let me first illustrate morpheme splitting on the familiar declension of the a-stems.
(81) Classical Armenian, a-stem declension
Recall that in the INS.PL, we have identified a sequence of three morphemes: -a-wk‘.
I have suggested that -a- expresses class, which is spelled out together with other
inflectional categories in the unmarked cases, i.e., in NOM, ACC and LOC. I show that
on the example of the nominative, repeated from (75):
1056 P. Caha
(82) a.
b.
Under this analysis, the -k‘ in NOM.PL actually spells out the features which are
realized as -a- in the INS. Combining these statements, we realize that the reason why
-a- and -k‘ cannot fuse in the INS.PL is constituency: the features expressed as -wintervene
between the class marker and plural:
(83) a.
b.
The structural intervention of -w- thus forces a single morpheme (-k‘) to split
in two (-a- and -k‘), and reveals a hidden structure inside an apparently indivisible
morpheme: the nominative plural -k‘.
10.2 n-stems
N-stems in Classical Armenian offer another opportunity to observe the same effect.
I give two examples of this declension type below. The two types differ by vowel
quality in NOM/ACC/LOC.PL.
(84) Classical Armenian n-stems
Explaining the structure of case paradigms by the mechanisms 1057
The traditional approach (to be rejected) analyzes -n- and the preceding vowel
(if any) as a single morpheme, which combines with the root to form a stem. The
variation in the vocalic element of the theme marker is then seen as a variation of the
stem, distinct from affixation.
However, if we compare the vowel to the rest of the system, we notice that it is
identical to the vowel we observe as a case ending in a-stems. The following table
shows this, with parts identical across the paradigms boldfaced. For INS, I follow
Schmitt (1981:46) and equate -w and -b as two guises of the same morpheme, realized
as -b after consonants, and -w after vowels.37
(85) A-stems and n-stems decline the same (modulo n)
Taking the parallel seriously, we uncover two facts. The first one is a positional
asymmetry between the case marker -i- (precedes -n), and the case markers -ê and -w
(follow -n-). The way the asymmetry cuts across the paradigm (INS + ABL vs. the
rest) seems to be yet another instance of case hierarchy effects.38
The second observation is directly relevant for the present concerns: unlike in the
singular, where we find a perfect match between the a-stem and the n-stem (modulo
n), an unexpected vowel emerges between the root and -n- in NOM/ACC/LOC.PL. I put
it in small caps:
(86) A mysterious vowel in the plural of n-stems
The theory developed up to now allows us to understand the emergence of this
unexpected vowel as a predicted instance of morpheme splitting.39 In particular, the
37I also assume that m in INS.PL is a phonological variant of the stem marker n.
38In the present account, the ordering in LOC, GEN and DAT is due to phrasal movement. The root moves
above case, without pied-piping the stem marker along.
39Note that an analysis in terms of a purely phonological process is not viable, or at least not obviously
so. To be sure, Classical Armenian does exhibit vowel-zero alternations (Schmitt 1981:38–39); however,
1058 P. Caha
stem marker -n- structurally intervenes between the class node, and the constituent
containing case/number, as indicated by its surface position in GEN-INS.PL. Because
of this intervention, -k‘ in NOM.PL cannot spell out both class and case/number features.
As a consequence, the class features must be spelled out by a separate morpheme:
the extra vowel. The account is fleshed out in more detail below.
First, I analyze the ‘stem marker’ -n- as a derivational morpheme. This proposal
is based on the facts discussed in Olsen (1999). According to this study (p. 113), the
n-stem declension is one of the most populated ones in Classical Armenian (around
2000 items). Importantly, the vast majority of these nouns are abstract nouns derived
from adjectives, nouns and verbs; example derivations include ‘old → oldness,’
‘brother → brotherhood,’ and ‘kill → killing’ (p. 547). This fact is accounted for if
-n- attaches to a (potentially) complex base, see (87), and turns it into a noun. I further
assume that the resulting complex is further dominated by the regular nominal
sequence of projections, including Class, Number and Case:40
(87) [ K [ Num [ Cl [noun n [ derivational base (A,V,N) ] ] ] ]
The derivation of the surface order in the plural declension starts by a series of movements
of the base: first leading to [base-n], and then the base moves above the class
marker [base-Cl-n]. Once it lands in the Spec of the Class marker, it pied-pipes this
constituent to the left of Num: [base-Cl-n]-Num. Upon the merger of K (K [baseCl-n]-Num),
the constituent [base-Cl-n] moves across it without pied-piping Num,
leading to the structure (88).
(88)
The phrasal node below the K head in (88) reflects the proposals made earlier concerning
head movement of case features and tree simplification. As in the general
they follow a different pattern than the one we are dealing with here. Thus, in the prototypical instance, we
have a vowel in a closed syllable amis ‘month, NOM.SG’ and a zero in an open syllable amøs-u ‘month,
GEN.SG’. This regular pattern is different from the data at hand, where both zero mas-øn ‘part, NOM.SG’
and a vowel mas-in-k‘ ‘part, NOM.PL’ show up in a closed syllable. In other words, a theory based on
epenthesis (as observed elsewhere in the language) would wrongly predict NOM *mas-in. Cf. fn. 41.
40There is a number of nouns that are apparently underived. In fact, Olsen (1999:156) notes that the
declension type is so productive, that any noun whose ACC.SG accidentally ended in the sequence Cn were
re-analyzed as n-stems (i.e., C-n). I believe that these facts are compatible with the general view taken
here, since in general, there is no obstacle in nominalizing nouns (recall ‘brotherhood’). Once again, the
phrasal idiom view suggested in fn. 33 needs to be invoked.
Explaining the structure of case paradigms by the mechanisms 1059
case, the base further sub-extracts from Spec,ClP and need not concern us further.
The resulting structure (with traces omitted) is shown in (89):
(89)
(90) /-k‘/ ⇔
Now recall that -k‘ (and -s) spell out a constituent composed of the Class marker,
K heads and number. The entry for -k‘ is repeated in (90). This entry, however, cannot
spell out ClP node in (89), because of the presence of the nP node. As a result (89)
gets spelled out as (91), with {-k‘, -s, -c‘} replacing Nom , {V, -k‘, -s} replacing Cl0,
and {n} replacing nP:
(91)
Thus, the way insertion is set up predicts that once the class marker is structurally
separated from the case/number constituent by the intervening -n-, each must be
spelled out separately. This in turn explains the appearance of the additional vocalic
element between the base and the stem marker -n- in NOM, ACC and LOC plural. This
is a non-trivial result, recall that the emergence of the vowel disturbs the otherwise
perfect parallel (modulo n) between the a-stem and the n-stem declensions, shown
below:
(92) The mismatch between a- and n-stems in NOM., ACC., LOC. PL.
As things stand, however, this picture leads also to the prediction that the vowel
which appears between the noun and -n- is identical to the class marker which occurs
also in the oblique cases. As I show in a moment, this is correct for some nouns, but
1060 P. Caha
it is wrong for the paradigm above. We do not get -a-, but -u- (a class marker of the
u-stems).
I do not know what is the source of the difference. What is needed is a proposal
of how individual class markers break down into components, for example where
exactly -n- and -a- come in, and so forth. Alternatively, one can see the -u- as a result
of ablaut, which has been suggested in Schmitt (1981:44). Thus, we would have an
underlying a, subject to a morphologically triggered ablaut, yielding the surface -u(see
Guerssel and Lowenstamm 1996 for a relevant discussion). I leave this for future
research.
On the other hand, a non-negligible number of n-stems does in fact show the predicted
nominative -a-n-k‘ and the accusative/locative -a-n-s (Schmitt 1981:103 lists
about 30 such nouns). I am going to discuss one such example, namely the plural
forms of the noun ‘sister in law,’ shown below:41
(93) An alternating noun: ‘sister in law’
In the cells stretching from NOM to ABL plural, this particular noun inflects either
as a n-stem, or as an a-stem. The n-stem forms include the bracketed material, the
a-stem forms exclude it. This variable behavior provides us with a minimal pair. According
to the analysis, the presence of the additional n in NOM/ACC/LOC should lead
to a structure where an additional class marker has to show up. And this is exactly
what happens. Thus, compared to the a-stem forms of the same noun, the inclusion
of -n- in NOM, ACC or LOC automatically leads to the emergence of an additional
vowel. (Equivalently, the absence of -n- leads to the disappearance of this vowel.)
The paradigm shown above thus bears out the predictions in their strongest form.
10.3 On phonological conflation
In this section, I address an apparent syncretism of the instrumental with the dative
(and other cases) across the ablative in u-stems, shown in (94).
41This paradigm provides additional evidence against treating the vocalic marker as epenthetic. That is
because the root is vowel final, and hence, the vowel does not improve the phonotactics in any way.
Explaining the structure of case paradigms by the mechanisms 1061
(94) An INS–DAT syncretism in Classical Armenian?
The discussion of this piece of data will ultimately show that the syncretism is irrelevant
for the Case sequence (as given in (4)), because it is the product of a phonological
conflation. Crucially from the current perspective, the discussion is also going
to provide additional evidence for the analysis of morpheme splitting, and that is why
I discuss this paradigm here.
The highlighted syncretism goes against the order of cases in Classical Armenian,
because—as established on independent grounds—INS must be separated from DAT
by ABL. Thus, we apparently get a situation in which non-adjacent layers of case
show syncretism, and this cannot be accounted for by the present system. However,
I am going to argue that the syncretism is the result of a phonological process which
merges u and w into one segment, and thus the underlying morphology looks as
depicted in (95).
(95) a. Dat: u
b. Ins: u-w ⇒ -u
There are two aspects of the proposal. The first aspect is that the sequence uw is
simplified to u in Classical Armenian. The second aspect is the hypothesis that there
is an underlying -uw present in the instrumental. I provide evidence for these two
claims in turn.
The evidence for the existence of the relevant phonological process consists in
showing that the process is attested elsewhere in the language. The following example
set shows this. In (96a), we observe an instance of a productive suffix -wor, which
(in one of its uses) means ‘bearer’ (Olsen 1999:362). When this suffix attaches to the
stem zim-u ‘armour,’ we obtain the relevant sequence uw. The resulting form zimuor
then confirms that uw yields u quite generally (Olsen 1999:367, 615; cf., Schmitt
1981:46).
(96) a. ałełna
bow
-wor
-bearer
→ ałełnawor
‘archer’
b. zimu
armour
-wor
-bearer
→ zimuor
‘soldier’
Turning to the second point: what evidence is there for the underlying presence of
-uw in the instrumental plural of the u-stem ‘time’? The evidence is provided by the
declension of the noun ‘day’:
1062 P. Caha
(97) The r-stem ‘day’ with a mysterious -b
The declension of the r-stem ‘day’ shows a parallel to the n-stems in the following
sense: in the cells from NOM to ABL, the noun ‘day’ inflects just like the u-stem
‘time’ with an additional r that occupies a structural position analogous to -n. What
is to be noted is the presence of two markers (apart from r) in the INS of ‘day’ (u and
b). This contrasts with apparently only one marker in the paradigm ‘time’ (u). The
contrast sticks out when we realize that the homophonous DAT u does not split, but it
is ordered to the left of r. Hence, what is initially mysterious is the emergence of the
-b in aw-u-r-b; that’s why I have put it in small caps in (97).
The point is that the unexpected appearance of -b becomes completely regular
once we adopt the proposal that the INS -u is underlyingly u-w. This sequence is
merged into one segment in cases where they end up adjacent, but it is preserved
when they are separated by the consonantal stem marker:
(98) The r-stem ‘day’ explained
This provides the evidence needed for the underlying presence of u-w in the INS of
u-stems, and supports the conclusion that the syncretism of DAT and INS across a distinct
ABL is to be treated as accidental, and irrelevant for the issues under discussion.
The paradigm of ‘day’ is interesting also for the phenomenon of morpheme
splitting. Because of the parallel behavior of r and n, we now predict that in the
NOM/ACC/LOC.PL, a class marker should emerge between the root aw ‘day’ and the
r. The prediction is borne out. As the table below shows, the class marker -u- (in
small caps) appears where predicted, even though it has no counterpart in the (otherwise
parallel) declension of the u-stem noun ‘time’:
(99) Morpheme splitting in r-stems
Explaining the structure of case paradigms by the mechanisms 1063
The paradigms above show that the appearance of the vocalic class marker in the
relevant cells of the paradigms is not an effect of a particular class (the n-stems or the
r-stems). It is a structurally governed process, which appears every time the relevant
structural configuration obtains.
11 Conclusions
The main goal of this paper has been to show that if one adopts (i) fine-grained
syntax, and (ii) phrasal spell-out, an interesting analysis of the Classical Armenian
declension appears. Two aspects of the approach deserve final highlighting.
The first one is the fact that the observed ‘case sequence’ governs the behavior of
at least three logically independent empirical domains (syncretism, case attraction,
and agglutination vs. fusion). I believe it is a challenge for any theory to explain
why these phenomena should be subject to the same regularity. The current proposal
achieves this by ‘cumulative decomposition:’ the proposal that individual cases are
characterized by a growing number of features (cf. Caha 2009). Because of the way
the case sequence penetrates various aspects of Classical Armenian grammar, I think
its discovery and its particular implementation may well be the main empirical contribution
of this paper. If correct, the approach taken here suggests that there is only
one mode of grammatical organization of smaller units into bigger chunks (i.e., binary
branching tree), no matter how small (sub-morphemic features) or big (phrases)
the units are.
And secondly, it is also important to pay attention to the mechanisms which derive
these results on the basis of the proposed decomposition. In the current theoretical
context, phrasal spell-out based on the Superset Principle (due to Starke 2009) may
be the most important innovation described here. Without phrasal spell-out, syntactic
terminals are the only locus for insertion, and their number determines the number of
slots in the paradigm. But since the number of separate morphological slots for class
and gender varies (none, 1 or 2), and the number of syntactic terminals for number
and gender is constant, the mismatches have to be stated by brute force and lead to
the consequence that a separate morphological structure becomes indispensable. In a
phrasal spell-out system, the number of terminals may not correspond to the number
of morphemes. Consequently, the number of paradigm slots may vary, and appears as
an automatic consequence of the interaction between the syntactic structure and the
actual entries.
Acknowledgements A number of people gave me comments and suggestions concerning the material
presented here. I started working on the topic as a part of my dissertation-related research, and my supervisor
Michal Starke has substantially influenced the initial stage of work. Jonathan Bobaljik and Hilda
Koopman (in their capacity as committee members) have left their mark on this work as well, for which
I am very grateful. In addition, a very early version of the paper was presented at WOTM 4 in Leipzig
(2008), and I thank the audience there for feedback and interesting discussions.
Subsequently, Gillian Ramchand and Peter Svenonius have provided me with detailed comments on
a draft of this material as it appeared in Nordlyd (Tromsø working papers in linguistics, 2009). Major
changes have occurred in the first draft submitted to NLLT due to the comments from three anonymous
reviewers, accompanied by an additional review by the handling NLLT editor, Gereon Müller. For the final
round of comments, I am indebted to Marcel den Dikken. A very special thanks to Marina Pantcheva, who
has read and commented on all these various versions.
1064 P. Caha
I thank all these people for bringing up empirical challenges, pointing out problems of analysis, finding
better ways of putting things, removing typos, correcting my English, and asking some big picture
questions.
References
Abraham, Werner. 2003. The myth of doubly governing prepositions in German. In Motion, direction and
location in languages: In honor of Zygmunt Frajzyngier, eds. Erin Shay and Uwe Seibert, 19–38.
Amsterdam: John Benjamins.
Ackema, Peter, and Ad Neeleman. 2007. Morphology = syntax. In The Oxford handbook of linguistic
interfaces, eds. Gillian Ramchand and Charles Reiss, 325–353. Oxford: Oxford University Press.
Baerman, Matthew. 2008. Case syncretism. In The handbook of case, eds. Andrej Malchukov and Andrew
Spencer, 219–230. Oxford: Oxford University Press.
Baerman, Matthew, Dunstan Brown, and Greville G. Corbett. 2005. The syntax-morphology interface.
A study of syncretism. Cambridge: Cambridge University Press.
Baker, Mark C. 1988. Incorporation: A theory of grammatical function changing. Chicago: University of
Chicago Press.
Baker, Mark C., and Ruth Kramer. 2010. “Prepositions” as case morphemes inserted at PF in Amharic.
Handout of a talk at the BCGL5 conference, 2–3 December, at CRISSP, Brussels.
Baltin, Mark. 2006. The non-unity of VP-preposing. Language 82: 734–766.
Bittner, Maria, and Ken Hale. 1996. The structural determination of case and agreement. Linguistic Inquiry
27(1): 1–68.
Blake, Barry J. 1994. Case, 2nd edn. Cambridge: Cambridge University Press.
Bobaljik, Jonathan. 2002. Syncretism without paradigms: Remarks on Williams 1981, 1994. In Yearbook
of morphology 2001 eds. Geert Booij and Jaap van Marle, 53–86. Dordrecht: Kluwer.
Bobaljik, Jonathan. 2012. Universals in comparative morphology: Suppletion, superlatives, and the structure
of words. Cambridge: MIT Press.
Bobaljik, Jonathan. 2007. On comparative suppletion. Ms., University of Connecticut.
Borer, Hagit. 2005. Structuring sense: An exo-skeletal trilogy. New York: Oxford University Press.
Caha, Pavel. 2009. The Nanosyntax of case. PhD thesis, CASTL, University of Tromsø.
Caha, Pavel. 2011a. Case in adpositional phrases. Ms., CASTL, Tromsø.
Caha, Pavel. 2011b. The parameters of case marking and spell out driven movement. Linguistic Variation
Yearbook 2010 10: 33–78.
Calabrese, Andrea. 2008. On absolute and contextual syncretism. In Inflectional identity, eds. Asaf
Bachrach and Andrew Nevins, 156–205. Oxford: Oxford University Press.
Cardinaletti, Anna, and Michal Starke. 1999. The typology of structural deficiency: A case study of three
classes of pronouns. In Clitics in the languages of Europe, ed. Henk van Riemsdijk, 145–233. Berlin:
Mouton.
Carnie, Andrew. 2008. Constituent structure. Oxford: OUP.
Chomsky, Noam. 1995. The minimalist program. Cambridge: MIT Press.
Chomsky, Noam. 2008. On phases. In Foundational issues in linguistic theory: Essays in honor of JeanRoger
Vergnaud, eds. Robert Freidin, Carlos P. Otero, and Maria Luisa Zubizarreta, 133–166. Cambridge:
MIT Press.
Cinque, Guglielmo, and Luigi Rizzi. 2010. The cartography of syntactic structures. In The Oxford handbook
of linguistic analysis, eds. Bernd Heine and Heiko Narrog, 51–65. Oxford: Oxford University
Press.
Embick, David, and Rolf Noyer. 2007. Distributed Morphology and the syntax–morphology interface. In
The Oxford handbook of linguistic interfaces, eds. Gillian Ramchand and Charles Reiss, 289–324.
Oxford: Oxford University Press.
Guerssel, Mohand, and Jean Lowenstamm. 1996. Ablaut in Classical Arabic measure I active verbal forms.
In Studies in Afroasiatic grammar, eds. Jacqueline Lecarme, Jean Lowenstamm, and Ur Shlonsky,
123–134. The Hague: HAG.
Haegeman, Liliane. 1994. Introduction to government and binding theory. Oxford: Blackwell.
Halle, Morris. 1997. Impoverishment and fission. In Papers at the interface, eds. Benjamin Bruening, Y.
Kang, and Martha McGinnis. Vol. 30 of MITWPL, 425–449. Cambridge: MIT Press.
Explaining the structure of case paradigms by the mechanisms 1065
Halle, Morris, and Alec Marantz. 1993. Distributed Morphology and the pieces of inflection. In The view
from building 20: Essays in linguistics in honor of Sylvain Bromberger, eds. Kenneth Hale and
Samuel Jay Keyser, 111–176. Cambridge: MIT Press.
Halle, Morris, and Bert Vaux. 1998. Theoretical aspects of Indo-European nominal morphology: The nominal
declensions of Latin and Armenian. In Mir Curad: Studies in honor of Clavert Watkins. Jay
Jasanoff, H. Craig Melchert, and Lisi Olivier. Innsbrucker Beiträge zur Sprachwissenschaft, 223–
240. Innsbruck: Institut für Sprachwissenschaft der Universität Innsbruck.
Harley, Heidi. 2008. When is syncretism more than a syncretism? Impoverishment, metasyncretism and
underspecification. In Phi theory: Phi-features across modules and interfaces, eds. Daniel Harbour,
David Adger, and Susana Béjar, 251–294. Oxford: Oxford University Press.
Jackendoff, Ray. 1983. Semantics and cognition. Cambridge: MIT Press.
Jakobson, Roman. 1962. Beitrag zur allgemeinen Kasuslehre: Gesamtbedeutungen der russischen Kasus.
In Selected writings, Vol. 2, 23–71. The Hague: Mouton.
Johnston, Jason Clift. 1996. Systematic homonymy and the structure of morphological categories. Some
lessons from paradigm geometry. PhD thesis, University of Sydney.
Julien, Marit. 2007. On the relation between morphology and syntax. In The Oxford handbook of linguistic
interfaces, eds. Gillian Ramchand and Charles Reiss, 209–238. Oxford: Oxford University Press.
Kayne, Richard S. 1994. The antisymmetry of syntax. Cambridge: MIT Press.
Kayne, Richard S. 2005. Some notes on comparative syntax, with special reference to English and French.
In The Oxford handbook of comparative syntax, eds. Guglielmo Cinque and Richard S. Kayne, 3–69.
Oxford: Oxford University Press.
Kiparsky, Paul. 1973. ‘Elsewhere’ in phonology. In A Festschrift for Morris Halle, eds. Paul Kiparsky and
Steven Anderson. New York: Holt, Rinehart and Winston.
Koopman, Hilda, and Anna Szabolcsi. 2000. Verbal complexes. Vol. 34 of Current studies in linguistics.
Cambridge: MIT Press.
Kratzer, Angelika. 2000. Building statives. Berkeley: Berkeley Linguistic Society.
Krause, Todd B., and Jonathan Slocum. no date. Classical Armenian online. http://www.utexas.edu/cola/
centers/lrc/eieol/armol-0.html.
Marantz, Alec. 1995. ‘Cat’ as a phrasal idiom: Consequences of late insertion in Distributed Morphology.
Ms., MIT.
McCawley, James D. 1968. Lexical insertion in a transformational grammar without Deep Structure. In
Papers from the fourth regional meeting of the Chicago Linguistic Society, eds. B. J. Darden, C.-J. N.
Bailey, and A. Davidson. Chicago: University of Chicago.
McCreight, Katherine, and Catherine V. Chvany. 1991. Geometric representation of paradigms in a modular
theory of grammar. In Paradigms: The economy of inflection, ed. Frans Plank, 91–112. Berlin,
New York: Mouton de Gruyter.
McFadden, Thomas. 2007. Default case and the status of compound categories in Distributed Morphology.
Penn Working Papers in Linguistics (Proceedings of PLC 30) 13.1: 225–238.
McFadden, Thomas. 2004. The position of morphological case in the derivation: A study on the syntaxmorphology
interface. PhD thesis, University of Pennsylvania.
Müller, Gereon. 2003. A Distributed Morphology approach to syncretism in Russian noun inflection. In
Proceedings of FASL 12, eds. Olga Arnaudova, Wayles Browne, Maria Luisa Rivero, and Danijela
Stojanovic, 353–374. Ann Arbor: Michigan Slavic Publications.
Muriungi, Peter. 2008. Phrasal movement inside Bantu verbs: Deriving affix scope and order in Kiitharaka.
PhD thesis, CASTL, Tromsø.
Neeleman, Ad, and Kriszta Szendr˝oi. 2007. Radical pro-drop and the morphology of pronouns. Linguistic
Inquiry 38: 671–714.
Olsen, Birgit Anette. 1999. The noun in Biblical Armenian. Berlin, New York: Mouton de Gruyter.
Pantcheva, Marina. 2010. The syntactic structure of locations, goals, and sources. Linguistics 48: 1043–
1081.
Pantcheva, Marina. 2011. Decomposing Path. The Nanosyntax of directional expressions. PhD thesis,
CASTL, Tromsø.
Pesetsky, David. 1995. Zero syntax: Experiencers and cascades. Cambridge: The MIT Press.
Phillips, Colin. 2003. Linear order and constituency. Linguistic Inquiry 34: 37–90.
Plank, Frans. 1991. Rasmus Rask’s dilemma. In Paradigms: The economy of inflection, ed. Frans Plank,
161–196. Berlin: Mouton de Gruyter.
Plank, Frans. 1995. (Re-)introducing suffixaufnahme. In Double case: Agreement by suffixaufnahme, ed.
Frans Plank, 3–112. Oxford: OUP.
1066 P. Caha
Plank, Frans. 1999. Split morphology: How agglutination and flexion mix. Linguistic Typology 3: 279–
340.
Radkevich, Nina. 2009. Vocabulary insertion and the geometry of local cases. Ms., UConn, downloadable
at: http://ling.auf.net/lingBuzz/000958.
Ramchand, Gillian. 2008. Verb meaning and the lexicon. Cambridge: Cambridge University Press.
Richards, Norvin. 2007. Lardil “case stacking” and the structural/inherent case distinction. Ms., MIT,
downloadable at: http://ling.auf.net/lingBuzz/000405.
Schmitt, Rüdiger. 1981. Grammatik des Klassisch-Armenischen. Vol. 32 of Innsbrucker Beiträge zur
Sprachwissenschaft. Innsbruck: Institut für Sprachwissenschaft der Universität Innsbruck.
Schütze, Carson. 2001. On the nature of default case. Syntax 4(3): 205–238.
Starke, Michal. 2009. Nanosyntax. A short primer to a new approach to language. In Nordlyd 36: Special
issue on Nanosyntax, eds. Peter Svenonius, Gillian Ramchand, Michal Starke, and Tarald Taraldsen.
1–6. Tromsø: University of Tromsø. Available at www.ub.uit.no/munin/nordlyd/.
Starke, Michal. 2010. Universal grammar vs. lexically driven derivations, once more. Talk at the SmaSh
workshop. ˇCeské Budˇejovice: University of South Bohemia, June 23.
Starke, Michal. 2005. Nanosyntax class lectures. Spring 2005, University of Tromsø.
Starke, Michal. 2011. Towards elegant parameters: Language variation reduces to the size of lexically
stored trees. Transcript from a talk at Barcelona Workshop on Linguistic Variation in the Minimalist
Framework. Available at http://ling.auf.net/lingBuzz/001183.
Šurkalovi´c, Dragana. 2011. Modularity, linearization and phase-phase faithfulness in Kayardild. Iberia:
An International Journal of Theoretical Linguistics 3(1): 81–118.
Taraldsen, Tarald. 2010. The Nanosyntax of Nguni noun class prefixes and concords. Lingua 120(6): 1522–
1548.
Travis, Lisa. 1984. Parameters and effects of word order variation. PhD thesis, MIT.
Vangsnes, Øystein A. 2013. Syncretism and functional expansion in Germanic wh-expressions. Language
Sciences 36: 47–65.
Weerman, Fred, and Jacqueline Evers-Vermeul. 2002. Pronouns and case. Lingua 112: 301–338.
Wiese, Bernd. 2003. Zur lateinischen Nominalflexion: Die Form-Funktions-Beziehung. Ms., IDS
Mannheim.
Wiese, Bernd. 2004. Categories and paradigms. On underspecification in Russian declension. In Explorations
in nominal inflection, eds. Gereon Müller, Lutz Gunkel, and Gisela Zifonun, 321–372. Berlin:
Mouton de Gruyter.
Williams, Edwin. 1981. On the notions “lexically related” and “head of a word”. Linguistic Inquiry 12.2:
245–274.
Williams, Edwin. 1994. Remarks on lexical knowledge. Lingua 92: 7–34.