Judith L. Aissen-RelationalGrammar

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 40

2.

Relational Grammar

Judith L. Aissen

2.0 Introduction

Relational Grammar (RG) is a theory of clause structure which takes gram-


matical relations (e.g., subject, direct object, predicate, etc.) as primitives.
Its principal goals are to define the class of possible clause structures
through the articulation of universal laws, to identify those structures which
occur in particular languages and, where possible, to provide valid crosslin-
guistic definitions for traditional terms of linguistic theory, e.g., passive,
antipassive, ergative, raising, voice, impersonal construction.
Historically, RG represents one reaction to the excessive dependence
of standard transformational grammar (TG) on language-particular config-
urationality. The earliest RG work (presented in 1974 lectures by David
Perlmutter and Paul Postal) saw that the centrality of phrase markers in TG
and its conception of 'rule' as a mapping between phrase markers pre-
vented it from providing the basis for a universal theory of grammar. Cen-
tral to RG is the claim that a universal theory of syntax must be grounded
in grammatical relations, and that grammatical relations cannot be defined
universally in terms of linear order and/or constituent structure, as assumed
in TG, and reaffirmed in more recent theories of configurationally defined
government. Linear precedence is defined only for elements of surface
structure, and hence plays a minor role in the theory.
Despite the rejection of some basic TG ideas, other central ideas were
retained (though in rather different form), in particular, the distinction
between an underlying level of syntactic representation and a superficial
level. The multistratal character of RG distinguishes it from monostratal
theories like Lexical Functional Grammar and Generalized Phrase Struc-
ture Grammar, and unites it with Government Binding theory.
64 JUDITH L. AISSEN

In RG, the structure of a sentence is represented as a uninetwork of


grammatical relations holding between elements at a number of distinct
levels. The class of well-formed structures is not characterized generatively
as in TG, but through well-formedness conditions on networks. Conse-
quences of this approach include the non-existence of lexical insertion
rules, base rules, transformational rules, derivations, rule ordering, or the
cycle. Since there is no rule application, constructions (e.g., passive) cannot
be identified with particular rules. Particular constructions are charac-
terized instead in terms of subnetwork configurations.
RG recognizes two classes of well-formedness conditions: universal
conditions, termed 'laws', which constrain all structures in all possible
human languages, and language-specific conditions, termed 'rules', which
constrain structures in some languages but not others. The formal nature of
these rules has received little explicit discussion within RG, though all
heretofore proposed laws can be interpreted as material implications in
predicate logic. However, there is no general agreement on the form of
either language-particular rules or laws. In Arc Pair Grammar (APG) — a
formalized theory presented in Johnson & Postal (1980) which builds on
basic RG ideas — all laws and rules are material implications.
Notable collections of RG work are Perlmutter (1983a) Perlmutter &
Rosen (1984), and Joseph & Postal (1990). See also Dubinsky & Rosen
(1987).

2.1 Passive as illustration

To ground some of these ideas in a concrete example, consider how the


relation between the farmer killed the duckling and the duckling was killed
by the farmer is characterized in RG. The active sentence has three parts,
each bearing a grammatical relation to the clause: the farmer is subject, kil-
led is predicate and the duckling is direct object. The passive sentence has
analyses at two distinct levels or strata, a first stratum and a second stratum.
The analysis in the first stratum is (nearly) identical to that of the active: the
duckling is direct object, the farmer is subject and (some form of) kill is
predicate. In the second stratum, the duckling is subject, and the f armer is
chomeur (see Section 2.3 below). Some form of kill remains predicate. The
analysis of the passive is represented in (1).
RELATIONAL GRAMMAR 65

(1) the duckling (was) killed (by) the farmer


Stratum 1 Direct Object Predicate Subject
Stratum 2 Subject Predicate Chomeur
Thematic relations are associated with stratum 1 (initial stratum) relations.
Hence, the synonymy of active and passive pairs — so far as concerns their
thematic relations — follows from the identity of their initial strata. On the
other hand, various grammatical phenomena are determined by final
stratum relations. (Here, the second stratum = final stratum). These
include control of agreement and case marking. In English, for example,
the final subject controls agreement and (ignoring complications) the final
subject is nominative, while other final relations are non-nominative.
A passive clause is defined universally as any one whose representation
includes the following subpart:
(2) Nominal Nominal
χ y
Stratum. Direct Object Subject
Stratum.i+1 Subject Chomeur or Ø
What is crucial to passive is the advancement of a direct object to subject,
and concurrent demotion of the earlier subject to chomeur, or, if it is
unspecified, its absence in the later stratum. By this definition, the duckling
was killed by the farmer is a passive clause.

2.2 Representation of basic clause structure

The structure of a clause is construed in RG as a set of linguistic states,


where a linguistic state is characterized by the existence of a grammatical
relation between two elements at some level or levels. Hence there are
three kinds of entities involved in linguistic states: linguistic elements (e.g.,
phrases, words, sounds), grammatical relations (e.g., subject, predicate),
and levels.
The formal representation of clause structure is based on the arc,
which represents a linguistic state. Accordingly, arcs are made up of three
sorts of formal entities: nodes, which represent linguistic elements; R(ela-
tional)-signs, which represent grammatical relations, and coordinate indi-
ces, which represent levels. Since arcs represent linguistic states, it follows
that an arc will consist of a pair of nodes representing the elements in the
66 JUDITH L. AISSEN

relation, an R-sign representing the relation, and a coordinate sequence


representing the level or levels at which the relation holds.
There are two conventional ways to represent graphically an arc,
shown in (3a) and (3b). Both represent the fact that a bears the relation
GR to b at ci, at ck, and at all strata between  i and ck, but not at i-1or at
C
k+1'

(3) a. b. [GR x (a, b) <ci ck>]

The (a) ('arrow') notation is generally used in representations of clause


structure; the (b) notation is used in the statements of rules and laws.
The relation between a and b is asymmetric: b governs a; conversely, a
is a dependent of b. The dependent element is the head of the arc, while the
governing element is its tail.
Since the set of elements in natural language is infinite (recall that this
set includes phrases), the set of nodes is also infinite. The set of levels is
finite, and small; so is the set of coordinates used to represent them. The
set of primitive grammatical relations is finite, and relatively small. This set
includes the members in (4), widely used in RG work. Each is associated
with an R-sign name.
(4) Grammatical relation R-sign
subject 1
direct object 2
indirect object 3
benefactive Ben
instrumental Inst
goal Goal
locative Loc
chomeur Cho
dead Dead
head H
genitive Gen
predicate Ρ
union U
conjunct Con
RELATIONAL GRAMMAR 67

These relations are classified into various sets through a classification of R-


signs. The most important classes are represented in (5), from Perlmutter
and Postal (1983b).
(5) Classes of R-signs

Since a clause is a set of linguistic states, its representation is simply a


set of arcs with the same governing node, the node representing the clause.
Under the analyses sketched in Section 2.2, the f armer killed the duckling
and the duckling was killed by the f armer have the representations in (6).
68 JUDITH L. AISSEN

Such representations are called relational networks (RNs). (6a) contains 3


arcs, represented by A, B, and C. Arc A represents the fact that the farmer
is subject ('1') of the clause labelled 80 at   ;   the fact that killed is predi-
cate ('P') of 80 at the same level; and   the fact that the duckling is direct
object of 80 ('2'), also at c1 In (6b), these elements bear relations at two
distinct levels, c1 and c2. Arc   represents the fact that the duckling is 2 in
clause 90 at   , Ε the fact that it is a 1 at c2. A represents the fact that the
farmer is 1 in 90 at c1 and F the fact that it is a Cho at c2 D represents the
fact that killed is Ρ at both  1 and c2 (The passive auxiliary be is treated as
predicate of a higher clause, and thus involves ascension (see Section 2.7.1
below).) Note that while all three elements bear relations at both levels, kil-
led heads one arc, while the duckling and the farmer head two each. The
reason is that an arc is associated with a single grammatical relation, and
with the entire sequence of coordinates at which that relation is borne.
Hence, killed heads one arc because it bears only one relation (P) in the
clause. Each of the two nominais bears two distinct relations, and hence
each heads two arcs.
A clause is a set of arcs. It can also be defined as a set of strata, where
each stratum is the maximal set of arcs with the same tail node sharing a
coordinate. 90 in (6b) contains two strata, the c1 stratum and the c2 stratum.
Stratum C1 contains A, C, and D, while c2 contains D, E, and F. The  1
stratum of a node b is its initial stratum. The   stratum of b is its final
1
stratum if b has no c.i±1 stratum.

2.3 Two fundamental laws and their interpretation

Two RG laws apply to a wide range of clause types. These are the Stratal
Uniqueness Law and the Motivated Chomage Law.
RELATIONAL GRAMMAR 69

The Stratal Uniqueness Law requires that no stratum contain more


than a single 1, 2, or 3.
(7) RG Stratal Uniqueness Law
If A is a Termx arc in stratum  i and   is a Termx arc in stratum
 ithen A = B.
The Stratal Uniqueness Law comes into play when a term relation borne by
a nominal in one stratum is assumed by a distinct nominal in the subsequent
stratum. The Stratal Uniqueness Law asserts that the former nominal
ceases to bear the relation in question. Consider the passive RN (6b). The
farmer is 1 in the first stratum; the duckling is 2 in the first stratum, but 1 in
the second. While the Stratal Uniqueness Law does not specify what rela-
tion (if any) the f armer bears in the second stratum, it eliminates the possi-
bility of the 1 relation.
In fact, the farmer is hypothesized to bear the chomeur relation. This is
permitted by the Motivated Chomage Law, which specifies necessary con-
ditions for the chomeur relation. The essential insight is that the chomeur
relation exists to obviate violations of the Stratal Uniqueness Law. Accord-
ingly, a nominal may assume the chomeur relation only when persistence of
its earlier relation would result in a violation of stratal uniqueness.
The statement of the Motivated Chomage Law is facilitated by intro-
ducing the APG notion overrun. Overrun reconstructs, in terms of arcs, the
situation in which one nominal 'assumes' the grammatical relation previ-
ously borne by another, thereby potentially producing a violation of stratal
uniqueness. One arc overruns a second when the two have the same tail
node and term R-sign, and the first coordinate index of the former is +1 of
the last coordinate index of the latter. In (6b), E overruns A because both
arcs have tail 90, both have the R-sign 1, and E's first coordinate is c?, while
A's last coordinate is   .
This allows the following statement of the Motivated Chomage Law
(see also Section 2.7.2 below on unions):
(8) RG Motivated Chomage Law
An RN containing   = [Cho (a, b)<cicw>] also contains
A = [Term x(a,b)<c uci-1> ] , where A is overrun.
(9) represents necessary conditions for Cho arcs. A and   in (9) instantiate
A and   in (8). Arc A is overrun by  
70 JUDITH L. AISSEN

In (6b), the overrun arc which sanctions the Cho arc F is A.


It is now possible to redefine passive clause making use of overrun:
(10) Def: a is a passive clause iff a contains
A = [2(a, b)<cuci-1>],   = [l(a, b)<ci.ck>], and some C overrun
by B.
Early versions of RG maintained that the conditions which allow
chomeurs (i.e., those which figure in the Motivated Chomage Law) actually
require chomeurs. That is, that an overrun arc always induced a chomeur.
This condition was termed the Chomeur Law. Note that the Chomeur Law
is not an entailment from the Stratal Uniqueness Law and the Motivated
Chomage Law. Chomage could be one way to avoid violations of stratal
uniqueness without being the only way. The head of an overrun arc could
bear no relation in the next stratum, or it could bear some relation other
than chomeur. Perlmutter and Postal (1983b) propose an analysis of
Kinyarwanda (in Tanganyika) which violates the (earlier) Chomeur Law.
Under that analysis, a nominal which heads an overrun 2 arc is demoted to
3 rather than to chomeur. The APG treatment of agentless passives also
violates the (earlier) Chomeur Law. Under that analysis, the head of the
overrun 1 arc bears no relation at all in the next stratum.
One might consider weakening the Chomeur Law to require chomage
just in case the head of an overrun arc (i) bears a relation in the subsequent
stratum, and (ii) bears no term relation in that stratum. However, this
appears to be unnecessary since the theory makes no non-term relation
other than chomeur available (see Section 2.4.1 on the Oblique Law). In a
sense, the chomeur relation can be viewed as the default for guaranteeing
stratal uniqueness, a possibility which is permitted for individual grammars
unless overridden by other general or language-particular conditions. One
focus of the following sections is to show how the Stratal Uniqueness Law
RELATIONAL GRAMMAR 71

and the Motivated Chomage Law make correct predictions about a wide
range of clause types.

2.4 Voice

2.4.1 Constraints on clause structure

RG contains a quite precise and inclusive theory of the domain traditionally


termed 'voice'. It provides a reasonable definition of 'voice', characterizes
the set of possible voices in natural language, and makes possible defini-
tions of traditional voice categories like 'passive', 'benefactive voice', 'in-
strumental voice', 'antipassive', and 'inversion'.
We will use the term 'revalue' in what follows. A nominal a (locally)
revalues if the relations it bears in successive strata (of the same clause) are
distinct, i.e., if it heads arcs of the form: [GRx(a, b)<cick>] and [GR (a,
b)<ck+1cm>], where  ≠  .The terms 'advancement' and 'demotion' refer to
the hierarchy
(11) 1 > 2 > 3 > Non-term
A revaluation is an advancement if the earlier relation is lower on the hier-
archy than the later relation; it is a demotion if the earlier relation is higher
on the hierarchy than the later relation. Passive, for example, involves an
advancement of the 2 to 1.
Chart 1 represents the set of possible (local) revaluations, derived by
assuming that every relevant relation can be either the earlier or the later
relation. Shaded boxes represent revaluations which are ruled out by prop-
osed laws, discussed immediately below.
Impossible revaluations are ruled out by two laws, the Oblique Law
and the Chomeur Advancement Ban (Perlmutter and Postal 1983b). Both
restrict revaluations of a single nominal. The Oblique Law prohibits revalu-
ation to an oblique relation, while the Chomeur Advancement Ban pro-
hibits revaluation of chomeur to a term relation:
(12) RG Oblique Law
If A is an oblique arc, then A is   1arc.
(13) RG Chômeur Advancement Ban
If an RN contains an arc of the form [Cho(a,b)<cx i> ] ,
then it contains no arc of the form [Term (a,b)<ci+1cw>].
72 JUDITH L. AISSEN

Chart 1 Possible Revaluations

These two laws prohibit all the revaluations in the 4th column (revaluations
to oblique), and all the remaining revaluations in the bottom row (revalua-
tions from chomeur). One further consequence of the Motivated Chomage
Law is that there can be no revaluation from oblique to chomeur, for this
would entail a revaluation to oblique. The remaining revaluations consti-
tute the set of claimed possible revaluations, and all have been attested.
Sections 2.4.2-2.4.5 discuss several.
However the set of possible revaluations is not equivalent to the set of
possible clause structures. Several laws constrain the membership of single
strata, regardless of what revaluations may have resulted in that stratum.
The Stratal Uniqueness Law (7) is an example; so is the Final 1 Law which
RELATIONAL GRAMMAR 73

requires that the final stratum of a clause contain a 1 arc. Further, the Moti-
vated Chomage Law constrains the cooccurrence of multiple revaluations
by requiring that revaluation to chomeur always cooccur with some revalu-
ation to a term relation. Finally, the  haracterization of some clauses, e.g.,
those involving ascension or union, requires that the structure of its com-
plement be taken into account as well.

2.4.2 Revaluations to 1

There are three candidates for advancement to 1: 2s, 3s, and obliques. The
Stratal Uniqueness Law makes a uniform prediction for all three cases: an
overrun 1 arc will not persist into the next stratum (i.e., the earlier 1 will
cease being a 1). To my knowledge, this prediction is universally borne out.
There are no structures containing two 1s in the same stratum. The
Chomeur Law allows the earlier 1 to assume the chomeur relation, but does
not require it. There are cases, generally termed 'inversion', where the 1
revalues to 3 (see Section 2.4.5 below).
Advancement of 2 to 1 (passive) is apparently the most widely attested
revaluation. Most languages appear to have passives, and if a language has
revaluations, it generally has passive. Advancements to 1 from other rela-
tions are less common. They have been described chiefly for Malayo-
Polynesian (Austronesian) languages, 1 but are found in other languages as
well, e.g., Southern Tiwa (Tanoan), described below.
Allen and Frantz (1986) argue that Southern Tiwa allows advancement
to 1 of Goals (these should perhaps be identified as 3s, rather than Goals).
Compare (14), which involves no advancement, with (15) which does. 2
(14) Seuanide 0- wan- ban na- 'ay.
man 3iSG-come-PAST 1sG-to
T h e man came to me.'
(15) In- seuan-wan- ban (na).
1sG/3iSG-man- come-PAST 1SG
T h e man came to me.'
The various differences between the two sentence types follow from an
advancement analysis of (15), represented in (16b), and a non-advance-
ment analysis of (14), represented in (16a).
74 JUDITH L. AISSEN

The first person pronoun is final oblique in (14) and therefore cooccurs with
the postfix -'ay. It is final 1 in (15), and therefore occurs alone and may be
silent. In (14), seuanide is final 1, while in (15), it is final chomeur. There
are two arguments that seuan- is final chomeur in (15). Both are more than
arguments that it is not a final 1, for both depend on general properties of
absolutive chomeurs in Southern Tiwa. (An absolutive chomeur is a nomi-
nal whose relation in the pre-chomeur stratum is absolutive, either subject
in an intransitive stratum or direct object in a transitive one.) Such
chomeurs arise not only in goal-to-1 advancement structures, but also in 3-
to-2 advancement structures and in possessor ascension clauses. The first
argument concerns incorporation. Allen, Gardiner, and Frantz (1984)
argue that all absolutive chomeurs obligatorily incorporate. Under (16b),
seuan is a final absolutive chomeur, accounting for its obligatory incorpora-
tion in (15). The second concerns agreement. Agreement in Southern Tiwa
is determined by final 1s, final 2s, and final absolutive chomeurs (Allen and
Frantz 1983). In clauses containing more than one controller, all controllers
jointly determine a single affix. The only legal agreement controller in (14)
is seuanide, final 1. But under (16b), (15) contains two: na T , the final 1,
and seuan- 'man', final (absolutive) chomeur. Accordingly, the agreement
affix in- is jointly determined by both, as it is in any structure containing a
1st person singular pronoun as final 1 and a noun belonging to the same
class as seuan- (class i) as absolutive chomeur. The Stratal Uniqueness Law
requires that seuan- not be final 1, and the Motivated Chomage Law allows
it to be final chomeur. (See Rosen (1990) for a different RG analysis.)

2.4.3 Revaluations to 2

There are two candidates for advancement to 2: 3s and obliques; advance-


ments of both types are fairly widely attested. The Stratal Uniqueness Law
RELATIONAL GRAMMAR 75

makes the following prediction for such cases: the head of an overrun 2 arc
will not persist as 2 into the next stratum (i.e., the earlier 2 will assume
some other relation). The Motivated Chomage Lav allows assumption of
the chomeur relation, but assumption of the 3 relation is also possible.
These predictions can be tested in an interesting way in morphologi-
cally ergative languages. A number of Mayan languages have advance-
ments to 2: Tzotzil (and other members of the Tzeltalan branch) have
advancement from 3 (Aissen 1983, 1987); Quiché (and some other Eastern
Mayan languages) have advancement from Instrument. I discuss only Tzot-
zil 3-to-2 advancement here, though the same conclusions hold for Quiché
instrument advancement.
Tzotzil is a predicate-initial, subject final language. Non-emphatic pro-
nouns are not pronounced. The predicate agrees both with the final 1 and
2. Final 1s of transitive clauses (ergatives) determine one form of agree-
ment, while final 1s of intransitive clauses and final 2s (absolutives) deter-
mine another.
Advancement of 3 to 2 (32A) is obligatory in Tzotzil (i.e., there are no
superficial 3s in Tzotzil). Its application is marked on the verb by the suffix
-be, glossed ΊO' below, and its most salient consequence is control of
absolutive agreement by the initial 3. The (b) sentences below involve 32A.
(17) a. L- i- y- ak'.
ASP-ABSl-ERG3-giVe
'He gave me.'
b. L- i- y- ak'- be chitom.
ASP-ABSl-ERG3-giVe-IO p i g
'He gave me pigs.'
(18) a. Ch-a- j - toj.
ASP-ABS2-ERGl-pay
'I will pay you.'
b. Ch-a- j - toj-be chitom.
ASP-ABS2-ERGl-pay-IO p i g
'I will pay you for the pig.'
(19), proposed for (17b), correctly predicts that as final 2, Τ controls
absolutive agreement. Under (19), chitom fails to control agreement
because it is final chomeur.
76 JUDITH L. AISSEN

The passive versions of (17b), (18b) show that chitom is not a 2, in accord
with the Stratal Uniqueness Law. Such clauses involve two advancements,
32A, marked by -be, and passive, marked by -at. The initial 3 advances to
2 and then to 1, where it controls absolutive agreement:
(20) L- i- 7ak'-b- at chitom.
ASP-ABs1-give-IO-PAss pig
'I was given pigs.'
(21) Ch- a- toj-b- at chitom.
ASP-ABS2-pay-IO-PAss pig
'You were paid for the pig.'
Significant here is that the final 1 controls absolutive agreement, not erga-
tive agreement, i.e., these clauses are finally intransitive, not finally transi-
tive, as represented in (22). The chomeurhood of chitom, induced by 32A,
is what makes the clause finally intransitive. If chitom were not put in
chomage by 32A, but remained a 2, subsequent passive would result in a
transitive clause, as in (23). However, (23) is patently incorrect since it pre-
dicts ergative agreement with the final 1. The ill-formedness of (23) follows
from the Stratal Uniqueness Law.
RELATIONAL GRAMMAR 77

A number of serious challenges to the Stratal Uniqueness Law involve


advancement to 2 structures, particularly in several Bantu languages. These
all involve cases where the earlier 2 continues to display properties taken as
diagnostics for 2s, implying that the earlier 2 persists as 2, in violation of the
Stratal Uniqueness Law. Perlmutter and Postal (1983b) argue that such
cases have an account consistent with the Stratal Uniqueness Law. The
account has two essential properties: (a) the earlier 2 demotes not to
chomeur, but to 3; (b) those properties taken as diagnostics for 2s are really
diagnostics for 2s and 3s. Hence, both the advanced 2 and the earlier 2
exhibit the properties in question.

2.4.4 Unaccusative structures

There is one revaluation structure which never involves overrun. These are
cases of 'unaccusative advancement'. Unaccusative structures have played
an important role in RG (and more recently, in other theories as well).
Perlmutter (1978) proposed that intransitive strata (= strata containing
a single nuclear term (1 or 2) arc) are of two types: those containing a 1 arc
but no 2 arc ('unergative strata') and those containing a 2 arc but no 1 arc
('unaccusative strata').
78 JUDITH L. AISSEN

When   is the initial stratum, the distinction clearly has semantic correlates
(Perlmutter 1978. Rosen 1984), with non-agentive or stative predicates
tending to occur in initially unaccusative strata and agentive or active pred-
icates tending to occur in unergative strata. However, the distinction has
been justified for particular languages on syntactic grounds.
While   in (24a) may be the initial stratum, it may not apparently be
the final stratum. That is, all clauses appear to have final 1s, a condition
expressed in the Final 1 Law:
(25) RG Final 1 Law
If clause a has a stratum   but no stratum   ,,, then   contains
η η+Ρ η
a 1 arc.
The Final 1 Law can be satisfied in unaccusative structures by advancement
of the 2 to 1, where the advancement is termed "unaccusative advance-
ment". Note that structures involving unaccusative advancement are not
passive structures, for while they involve advancement of 2 to 1, no 1 arc is
overrun in unaccusative structures.
Italian perfect auxiliary selection provides evidence for the unaccusa-
tive hypothesis. See Rosen (1981) for discussion, or Rosen (1984) for a
summary of the arguments. Italian has two perfect auxiliaries avere 'have'
and essere 'be'. All transitive (non-reflexive) clauses select avere.
(26) Mario ha/*è difeso Luigi.
'Mario defended Luigi.'
Certain intransitive constructions, including passive, are restricted to
essere.
(27) I palazzi sonol*hanno stati ricostruiti.
'The buildings have been rebuilt.'
Other intransitive clauses are of two types: those which take avere and
those which take essere.
(28) a. Mario ha/*è telefonato/barato/starnutito.
'Mario telephoned/cheated/sneezed.'
b. I ragazzi sonol*hanno rimasti/venuti/nati.
'The boys stayed/came/were born.'
If those intransitive clauses which select essere have initial unaccusative
strata, and those which select avere initial unergative strata, the distribution
of the auxiliaries can be characterized as follows:
RELATIONAL GRAMMAR 79

(29) Italian Auxiliary Selection


A clause selects essere iff it contains a nominal which heads both
a 1 arc and an object (2 or 3) arc.
Unaccusative clauses, as in (28b), satisfy (29) by virtue of unaccusative
advancement, while unergative clauses, as in (28a), do not. Passive clauses,
as in (27), also satisfy (29), explaining why they invariably select essere.
((29) also accounts for the invariant use of essere in reflexive structures; see
Section 2.5).
Several other phenomena in Italian bifurcate intransitive clauses along
exactly the same lines as auxiliary selection. One involves the partitive clitic
ne which is linked to a quantified nominal and indicates that it is headless.
In transitive clauses, ne can only be linked to a 2, not to a 1:
(30) a. Ne ho visti tre.
'I saw three (of them).'
b. *Ne li hanno visti tre.
(Three of them saw them.')
Among intransitive clauses, some allow ne, while others do not. Those
which take the auxiliary essere allow ne, while those which take avere do
not.
(31) a. Ne sono rimasti/venuti/nati tre.
'Three of them stayed/came/were born.'
b. *Ne hanno telefonato/barato/starnutito tre.
(Three of them telephoned/cheated/sneezed.')
Passive clauses allow ne:
(32) Ne sono stati ricostruiti tre.
Th ree have been rebuilt.'
If predicates which select avere are initially unergative, and those
which select essere initially unaccusative, then the generalization regarding
ne can be stated as follows: if ne is linked to a nominal, then that nominal
heads a 2 arc. 2s in transitive clauses (as in (30a)) can be linked to ne, while
1s in such clauses (30b) cannot be. As an initial 2, the initial nuclear term in
an unaccusative clause can be linked to ne (31a), while that of an unergative
clause cannot be (31b).
Clearly, the evidence for unaccusativity is particularly convincing in
Italian because several distinct phenomena bifurcate intransitive predicates
80 JUDITH L. AISSEN

in exactly the same way. Of course, there are many ways to classify predi-
cates, including purely diacritic ones. The case for distinguishing these clas-
ses according to initial unergativity/unaccusativity depends on the nature of
the phenomena involved. Some (e.g., ne) distinguish 1s and 2s in transitive
clauses, thus providing apparent tests for 1s and 2s. The unaccusative/uner-
gative distinction is motivated by the fact that the nuclear term in purported
unaccusative clauses aligns with the 2 in transitive clauses, while the nuclear
term in purported unergative clauses aligns with the 1.

2.4.5 Demotions

The RG theory of voice predicts three possible demotions:


(1) 1 to 3
(2) 1 to 2
(3) 2 to 3
Structures involving demotions to 3 are well-attested, with demotion of 1 to
3 termed 'inversion', and demotion of 2 to 3 often termed '2-to-3 retreat'.
For treatments of inversion, see particularly Harris (1981, 1984) on Geor-
gian; see Berinstein (1985) on 2-to-3 retreat in K'ekchi (Mayan).
1 to 2 demotion is probably the most contested clause-internal revalua-
tion. Nothing in RG rules it out, but evidence for it is scarce. Postal (1977)
suggested that 'antipassive' involved 1 to 2 demotion, largely as a way of
obviating violations of the Motivated Chomage Law which would be
entailed by the spontaneous demotion of 2 to chomeur represented in
(33a).

If, on the other hand, (33b) is the structure of an antipassive clause, demo-
tion of 1 to 2 puts the 2 in chomage and satisfies the Motivated Chomage
Law, since the initial 2 arc is overrun by A. Readvancement of the demoted
2 to 1 is required, but this is independently guaranteed by the Final 1 Law.
RELATIONAL GRAMMAR 81

More recently, Davies (1984) has provided empirical evidence for (33b),
arguing that the final 1 in Choctaw antipassive structures must also head a
2 arc to correctly characterize agreement.

2.5 Reflexive passives and copy structures

There are two further construction types traditionally considered 'passive'.


One is 'reflexive' passive, the other 'impersonal' passive. This section and
the next present analyses of these constructions, focussing first on how
these structures differ from 'plain' passives, and second on how they
nonetheless satisfy the definition of passive clauses. Reflexive passives are
illustrated by the Spanish sentences in (34-5).
(34) Las pirámides se edificaron por esclavos.
'The pyramids were built by slaves.'
(35) Estas obras se venden por todos los libreros.
'These works are sold by all booksellers.'
These share with coreferential reflexive sentences the reflexive clitic se.
They share with passive sentences the fact that what corresponds to 2 in an
active sentence here functions as final 1, and what corresponds to 1 in an
active sentence here functions as object of the preposition por (cf. las
pirámides fueron construidas por esclavos 'The pyramids were built by
slaves'). (In Spanish reflexive passives, the object oí por cannot be defi-
nite.) Ideally, one wants an account of reflexive passives which accounts
both for what they share with coreferential reflexive structures and for what
they share with plain passives.
I start with the representation of coreference. An important concep-
tual property of RG is that a single element may bear more than one rela-
tion in a sentence. This has already been illustrated several times in the rep-
resentation of revaluations, where one element bears at least two relations,
though not in a single stratum. A nominal may bear more than one relation
in the same stratum as well. If that stratum is the initial stratum, the result-
ing configuration is associated with a coreferential interpretation, as one
would expect, given the general principle that thematic relations are
associated with initial arcs. The initial strata of (36)-(37) are represented in
(38a, b) respectively.
82 JUDITH L. AISSEN

(36) Marta se nominó (a sí misma).


'Marta nominated herself.'
(37) Marta nos permitió nominarla (a ella).
'Marta permitted us to nominate her.'

Marta heads two initial arcs in both structures, and is said to be 'multi-
attached'. As arcs with a shared head, A and   'Overlap' in both structures.
Nosotros ('we') is also multi-attached in (38b);   and D overlap. 3
While Marta heads distinct initial arcs in both structures, it does not
head distinct final arcs in either. In both, one of the two arcs is replaced by
an arc headed by a pronoun. This is required unless one of the two arcs is,
in APG terms, 'erased' through Equi, as it is in (37). In (38a), Marta is
replaced by the reflexive pronoun si because, roughly, the two overlapping
arcs are 'neighbors' (have the same tail, hence are in the same clause). In
(38b), it is replaced by the non-reflexive pronoun ella because, roughly, the
two overlapping arcs are not neighbors. (Direct objects referring to defi-
nite humans are preceded by a in Spanish.)
(39a, b) represent these replacements.
RELATIONAL GRAMMAR 83

Note now that these pronouns constitute necessary and sufficient conditions
for the obligatory clitic pronouns se and la in (36)-(37), for in Spanish, per-
sonal object pronouns require a cooccurring clitic pronoun which agrees in
person, number, gender, and reflexivity. Spanish personal nonclitic pro-
nouns may 'drop', and they usually do when nonemphatic, but the clitics do
not.
The crucial point is that anaphoric pronouns are introduced in struc-
tures containing overlapping arcs. When those overlapping arcs are initial
arcs, the pronoun in question is a coreferential pronoun. However,
anaphoric pronouns may also be introduced in structures containing over-
lapping arcs which are not both initial arcs, but rather a pair of arcs defining
a revaluation structure.
Consider from this point of view (40), the proposed structure of (34),
a reflexive passive clause. (40) contains the three arcs which define passive
structures: the 2 and 1 arcs which define the revaluation, A and B, and the
overrun 1 arc, D. A and   overlap, and as overlapping neighboring arcs,
one may be replaced by an arc headed by a reflexive pronoun. A is replaced
by C, and C's head, sí, requires the reflexive clitic se.

Here, the pronoun sí is not a coreferential pronoun, but a copy pronoun.


Copy pronouns are involved when the overlapping arcs which induce the
pronoun define a revaluation structure.
(40), as structure of (34), provides an account both of what reflexive
passives share with plain passives — the triplet of arcs which define passive
structures — and of what they share with coreferential reflexives — a pair
of overlapping 1 and 2 arcs which constitute sufficient conditions for an
anaphoric reflexive pronoun.
84 JUDITH L. AISSEN

Under this account of copy structures, one is led to expect copy ver-
sions of the full range of revaluation structures. These are not in general
well-attested, perhaps suggesting the need for principles to exclude some.
However, copy unaccusative structures are not uncommon. (41)-(42)
exemplify non-copy and copy unaccusative structures, respectively, for
Spanish.
(41) a. Floreció.
'It bloomed.'
b. Verdeció.
'It turned green.'
  Envejeció.
'He grew old.'
(42) a. Se enrojeció.
'It turned red.'
b. Se enriqueció.
'She became rich.'
  Se desvaneció.
'It vanished.'
As these examples suggest, whether a predicate 'triggers' reflexive or plain
unaccusative advancement is a lexically idiosyncratic matter.
It is well-known that coreferential clauses and revaluation clauses are
often formally similar, reflexive morphosyntax being but one example.
Multi-attachment provides a shared common property, one which can serve
as the motivation for shared morphosyntax. Consider again in this light Ita-
lian auxiliary selection. As noted in Section 2.4.4, reflexive clauses (both
coreferential and copy) select essere, while all non-reflexive transitive
clauses select avere. But this follows from (29) if reflexive clauses involve
overlapping 1 and 2 or 1 and 3 arcs, for such structures then satisfy the con-
ditions for essere.

2.6 Impersonal passives and other dummy structures

(43)-(44) from Dutch and (49)-(50) from Turkish are 'impersonal passives'.
(43) Er wordt voor de koning geknield.
'It is kneeled before the king.'
RELATIONAL GRAMMAR 85

(44) Door de kinderen wordt altijd gelachen.


'By the children is always laughed.'
(45) Bu kapi-dan sokag- a çik- il- ir.
this door-from street-to exit-PAss-AOR
Through this door one gets out to the street.'
(46) Oda- m- da sigara iç- il- me- ζ.
room-my-in cigarette smoke-PAss-NEG-AOR
' O n e does not smoke in my room.'
('No smoking in my room.')
These are termed 'passives' because in both languages such examples have
the same morphology that characterizes plain personal passives (worden +
past participle in Dutch, -Il- in Turkish). Further, in Dutch, the agent is
expressed as object of door, as in personal passives. (The Turkish construc-
tion does not permit expression of the agent, but one is implied, as in pas-
sives.) The term 'impersonal' apparently refers to the fact that none of
these examples have active transitive counterparts, i.e, none have as 1 a
nominal which corresponds to 2 of a corresponding active. (A more
adequate definition of 'impersonal' is offered below.)
Such examples apparently pose problems for several RG laws. It is not
clear that they satisfy the Final 1 Law, since neither Turkish example has a
surface 1, nor does (44). If (44) lacks a final 1, then door apparently marks
a spontaneous chomeur, entailing a violation of the Motivated Chomage
Law.
The RG analysis of impersonal passives makes crucial appeal to dum-
mies, and to the possibility that they may be silent. The RG analysis of (43)
is represented in (47). The crucial property of impersonal passives is the
advancement of a dummy 2 to 1 and the resultant overrunning of a 1 arc.
86 JUDITH L. AISSEN

Note first that the clause in (47) is passive: it contains a 2 which advances to
1, and the initial 1 arc is overrun. Hence, passive morphology — in both
personal and impersonal passives — can be linked to this configuration.
(47) also accounts for the fact that some Dutch examples have overt dummy
1s. The Turkish examples, as well as (44), require the recognition of "silent
dummies", i.e., dummies with no phonological realization. Since anaphoric
pronouns in many languages are silent, and since dummies are generally
identical to pronouns, often anaphoric pronouns, the existence of silent
dummies in some languages is not unexpected. 4
Support for the dummy advancement analysis of impersonal passives
has come principally from two sources. The first is the existence of both
plain and reflexive impersonal passives, as in the following German pair:
(48) Es wird hier getanzt.
There is dancing here.'
(Tt is danced here.')
(49) Es tanzt sich gut hier.
There is good dancing here.'
(Tt dances itself well here.')
If the German reflexive sich requires overlapping of a 1 arc with a neighbor-
ing non-1 arc, then the analysis of impersonal passives in (47) provides the
necessary conditions for the reflexive.
The second line of argument comes from the absence of certain imper-
sonal passives. Impersonal passives of personal passive clauses and of unac-
cusative clauses are systematically absent. (50)-(52) illustrate these gaps for
Dutch unaccusatives (from Perlmutter & Postal 1984):
(50) a. De planten zijn al gerot.
T h e plants have already rotted.'
b. *Door de planten werd al gerot.
(51) a. In dit weeshuis groeien de kinderen erg snel.
In this orphanage the children grow very fast.'
b. *In dit weeshuis wordt (er) door de kinderen erg snel gegroeid.
(52) a. Alleen zijn moeder overleefde.
'Only his mother survived.'
b. *Er werd alleen door zijn moeder overleefd.
RELATIONAL GRAMMAR 87

Cf.
(53) Door deze mensen wordt er altijd gevochten.
'By these people it is always fought.'
(54) Er wordt in deze kamer vaak geslapen.
'It is often slept in this room.'
(55) Er wordt hier veel geskied.
'It is skied here a lot.'
These gaps follow from (47) if the 1 Advancement Exclusiveness Law
(1AEX) is assumed. Under the 1AEX, no clause may involve more than
one advancement to 1 (see Perlmutter & Postal 1984 for further discus-
sion). (50b) would require the structure in (56).

(56) contains two advancements to 1: (unaccusative) advancement of de


planten in the second stratum, and (passive) advancement of the dummy in
the third. Hence, (56) violates the 1AEX. The advancement of the dummy
to 1 in the third stratum is what makes (56) an (impersonal) passive. The
advancement of de planten to 1 in the second stratum is necessary in order
to make the advancement of the dummy in the third an instance of passive,
for passive requires an overrun 1 arc. Without unaccusative advancement in
the second stratum, (56) would have no 1 arc to be overrun, and hence
would not be an impersonal passive at all. In short, impersonal passives to
unaccusative predicates (or to personal passives — which also involve
advancement to 1) necessarily involve two advancements, in violation of
the 1AEX. Hence, this account predicts their non-existence. The appeal of
the account lies in the fact that it says nothing special about passives, imper-
sonal passives, or unaccusative structures at all.
88 JUDITH L. AISSEN

2.6.1 Theory of dummies

Dummies also figure in the analysis of clauses which are not passive, nota-
bly existential clauses, clauses with weather predicates, and clauses involv-
ing extraposition. For example:
(57) There are leaks in the pipe.
(58) It is unlikely that he knew it.
(59) It was raining.
(60) Il est arrivé trois filles.
Three girls arrived.'
(61) Es kamen zwei junge Leute.
T wo young people came.'
Perlmutter (1983b) defines impersonal clauses as clauses with dummies as
final 1s. Under this definition, (57)-(61) are all impersonal clauses.
The RG theory of dummies involves the following components:
a. the assumption that dummies are not in the initial stratum;
b. two laws, the Nuclear Dummy Law and the Active Dummy Law;
  the 'brother-in-law' relation.
The first assumption is necessary because dummies are not semanti-
cally relevant, and all initial arcs are semantically relevant. The Nuclear
Dummy Law and the Active Dummy Law narrowly restrict the class of
structures containing dummies. The Nuclear Dummy Law restricts dum-
mies to the heads of 1 and 2 arcs, thus ruling out any indirect object dum-
mies, oblique dummies, chomeur dummies, dead dummies, and overlay
dummies (i.e., no dummy topics, dummy interrogatives, etc.) (see Perlmut-
ter & Postal (1983b) for discussion). It explains, for example, the ungram-
maticality of (62) which contains a dummy chomeur:
(62) *I was surprised by it that he left.
(cf. It surprised me that he left.)
(63) RG Nuclear Dummy Law
If a is a dummy, then a heads a nuclear term arc.
The Active Dummy Law requires roughly that a dummy head some arc
which is an overrunner. This further constrains possible structures contain-
ing dummies. Under the Active Dummy Law, (64a,b) are possible struc-
tures for (57), but (64c) is not, since no arc is overrun by an arc headed by
the dummy.
RELATIONAL GRAMMAR 89

An important consequence of the Active Dummy Law is that it forces


advancement of a dummy 2 to 1 in impersonal passives of unergatives.
Without the law, nothing would prevent a dummy from entering as 2 and
not advancing.
Overrunning, however, is only required in cases where there is some-
thing to be overrun. Clauses containing weather predicates apparently con-
tain no initial 1 or 2, hence, the entry of the dummy cannot satisfy the
Active Dummy Law. Nonetheless, dummies regularly occur in such sen-
tences.
(65) RG Active Dummy Law
If cn is the first stratum containing a dummy a and cn-l contains a
nuclear term arc, then a heads some overrunner arc.
90 JUDITH L. AISSEN

The 'brother-in-law' relation is central to an account of agreement in imper-


sonal clauses. As is well-known, some nominal other than the final 1 fre-
quently determines agreement in such structures. In (57) and (61), for exam-
ple, leaks and zwei junge Leute determine plural agreement on the verb.
The brother in-law relation holds between a dummy and a nominal which
heads an arc overrun by the first arc headed by the dummy. The brother-in-
law of a dummy may control agreement in place of the dummy. (61), for
example, has roughly the structure in (64a), where zwei junge Leute is ini-
tial 1. The dummy enters as 1 in the second stratum, with the arc it heads
overrunning the one headed by zwei junge Leute. Since A is the first arc
headed by the dummy, zwei junge Leute is its brother in-law and may con-
trol agreement in its place.
Exploiting the RG theory of dummies, Perlmutter (1983b) provides an
unexpected account of so-called 'free' word order in Italian, a freedom
illustrated by (66)-(67).
(66) Molti stranieri hanno lavorato in quella fabbrica.
'Many foreigners worked in that factory.'
(67) Hanno lavorato molti stranieri in quella fabbrica.
'Many foreigners worked in that factory.'
Pairs like (66) and (67) are generally taken to have the same relational
structure but to exhibit free order between subject and predicate. In con-
trast, Perlmutter argues that the linear relation of the subject and predicate
is fixed and that (66) and (67) have significantly different relational struc-
tures. Assuming that in Italian a (final) 1 must precede its predicate, then
molti stranieri is final 1 only in (66). The final 1 in (67) is a silent dummy,
whose entry makes molti stranieri a chomeur.
Evidence for this relational difference comes from the possibility of
control into a variety of non-finite complements. Control is possible only
from 'subjects' which precede the predicate, not from those which follow.
(68) illustrates this contrast for participial clauses, (69) illustrates it for
adverbial infinitival clauses. The only surface difference between the (a)
and (*b) sentences is the position of the controller relative to its predicate.
(68) a. Entrati nel paese clandestinamente, dei profughi ungheresi
sono rimasti a Roma.
'Having entered the country illegally, some Hungarian
refugees remained in Rome.'
b. *Entrati nel paese clandestinamente, sono rimasti dei profughi
ungheresi a Roma.
RELATIONAL GRAMMAR 91

(69) a. Dei profughi ungheresi sono rimasti a Roma tanto tempo dopo
la guerra da sentirsi romani.
'Some Hungarian refugees remained in Rome so long after
the war as to feel themselves to be Romans.'
b. *Sono rimasti dei profughi ungheresi a Roma tanto tempo
dopo la guerra da sentirsi romani.
Perlmutter (1984) motivates a condition on control in Italian which restricts
it to nominais which head 1 arcs and final term arcs. (This allows control by
inversion' nominals which are initial 1s but final 3s.) Under the proposed
analysis, (68a, 69a) contain possible controllers, while (68b, 69b) do not,
since the post-verbal nominals in those examples are final chomeurs.
The agreement in (67) then is an instance of brother-in-law agreement:
mold stranieri is brother-in-law of the final 1 since it is put in chomage by
the dummy's entry, and thus may control agreement.
One prediction of this account is that in contrast to various other
impersonal structures, impersonal passives to intransitive predicates will
never manifest brother-in-law agreement. That is, the absence of brother-
in-law agreement in the Dutch examples (43)-(44) and (53)-(55) is neces-
sary and not accidental. As (47) makes clear, impersonal passives (to
intransitive predicates) contain no brother-in-law, for the first arc headed
by the dummy will always be a 2 arc, and the previous stratum will never
contain a 2 arc. Dutch does manifest brother-in-law agreement in other
impersonal struc tures, ones analogous to the Italian example (67) where
the dummy (er) enters as 1, putting the earlier 1 in chomage:
(70) a. Twee kinderen spelen in de tuin.
Two children are (PL) playing in the garden.'
b. Er spelen twee kinderen in de tuin.
T wo children are (PL) playing in the garden.'
(71) a. Veel huizen werden door de terroristen verwoest.
'Many houses were (PL) destroyed by the terrorists.'
b. Er werden veel huizen door de terroristen verwoest.
'Many houses were (PL) destroyed by the terrorists.'
In (70b), twee kinderen is the dummy's brother-in-law, and controls agree-
ment; in (71b), the brother-in-law is veel huizen. Note that (71b) is the
impersonal version of a personal passive: veel huizen advances from 2 to 1,
and is then put in chomage by the dummy. Hence, while (71b) is imper-
sonal and passive, it is not an impersonal passive.
92 JUDITH L. AISSEN

2.7 Multiclausal structures

2.7.1 Ascensions

The remaining major domain of RG laws is multiclausal structures, in par-


ticular, the phenomena termed 'ascensions' and 'unions'.
Ascensions have the property that a single dependent of a subordinate
node 'assumes' a grammatical relation to a higher node. The best-known
ascensions are so-called 'subject raising', and 'object raising', exemplified
in English by (72)-(75).
(72) There seem to be leaks in the pipe.
(73) I believe it to be inside.
(74) Mistakes are difficult to locate in such texts.
(75) Games are easy for children to appreciate.
In (72) and (73), both there and it bear final grammatical relations in
the main clause: there is final 1 and it is final 2. But neither bears any initial
grammatical relation in the main clause at all. This follows from the
assumption that all and only members of initial strata are relevant to the
construction of a logical representation of sentences. But both can be moti-
vated as final 1s in the subordinate clause. Hence, they must be raised into
the main clause.
Early work (Perlmutter & Postal 1983a) posited the following RNs for
(72) and (73).
RELATIONAL GRAMMAR 93

These structures raise the following question: what principle governs the
grammatical relation of the raised element? Why does there raise as 1 in
(76a), and not as 2 or 3? Why does it raise as 2 in (76b), and not as 1 or 3?
The answer is embodied in the Relational Succession Law (Perlmutter &
Postal 1983a), which requires that a raised nominal assume the grammatical
relation of the constituent out of which it is raised (its host). There raises as
1 because it is raised out of a 1; it raises as 2 because it is raised out of a 2.
Notice that the Relational Succession Law would entail violations of
the Stratal Uniqueness Law unless the host too assumed some new relation.
In these cases, it apparently becomes a chomeur, an assumption which
accounts for the post-verbal position of raising hosts in English.
The Relational Succession Law correctly predicts that in (74) and (75),
the objects of the subordinate clauses {mistakes, games) raise as 1s, for their
hosts are 1s (Cf. To find mistakes in such texts is difficult. To appreciate
games is easy for children.)
More recent work analyzes all sentential complements as 2s. In this
view, structure (76a) is replaced with one in which there raises out of a 2. By
the Relational Succession Law, there raises as 2, and then advances to 1 by
unaccusative advancement to satisfy the Final 1 Law.
Ascensions from clausal hosts are also governed by a second law, the
Host Limitation Law, which limits ascension hosts to terms. The above
cases are consistent with this, as all involve hosts which are 1s or 2s.
Because clauses rarely, if ever, function as 3s, cases of 3 clausal hosts are
correspondingly rare.
All the ascensions illustrated above involve clausal hosts, but ascen-
sions with nominal hosts also exist. These appear to be limited to nominals
containing a possessor, and it is the possessor which raises ('possessor
ascension'). Consider the following examples from Tzotzil (Aissen 1987).
94 JUDITH L. AISSEN

(77) Ch- i- s- toyilan- be j - jol.


ASP- ABs1-ERG3-keep lifting-ίο ERG1-head
'He kept lifting my head.'
(78) L- a- j -    - be l- a- tot- e.
ASP-ABS2-ERG1-meet-IO the-ERG2-father-clitic
'I met your father.'
Note that in (77), the initial 2 has a genitive (cross-referenced on the
head by means of the same prefixes used to cross-reference ergatives in
clauses), and further that the verb is suffixed with -be, suggesting 3-to-2
advancement (see Section 2.4.3). Absolutive agreement shows that the
final 2 in (77) is not jjol 'my head', but the first person pronoun, i.e, the
possessor of the initial 2. (78) has analogous properties. Aissen (1987)
proposes that these cases involve ascension of the genitive of the initial 2 to
clausal 3. Like all other 3s, these must advance to 2, thus accounting for the
presence of -be on the verb, as well as the form of absolutive agreement.
Once having advanced to 2, such 2s can passivize. Possessor ascension is
productive in Tzotzil; in some cases it is obligatory, in others optional, this
depending largely on the person of the genitive.
The ascension in (77) is consistent with the Host Limitation Law, since
the host is a 2. But it is not consistent with the Relational Succession Law
since the genitive ascends from a 2 host to clausal 3. While possessor ascen-
sion sometimes obeys the Relational Succession Law (see Allen, Gardiner
and Frantz (1984) on Southern Tiwa, for example) this is clearly not inevit-
able. Hence, the Relational Succession Law must except possessor ascen-
sion from its scope.
The Host Limitation Law, on the other hand, may be too lax. While
isolated examples have been claimed to show that non-terms may host pos-
sessor ascension, I am unaware of any sustained argument for this point. In
fact, possessor ascension appears to be overwhelmingly attested only for 2
and 'inactive' 1 hosts. If the latter are initial unaccusatives, it may be possi-
ble to restrict possessor ascension hosts to 2s. In some languages, (e.g.,
Tzotzil, Sierra Popoluca) only transitive 2s host possessor ascension. In
others (e.g., Acehnese), only 'inactive' 1s host possessor ascension. Finally
there are languages in which both 1s and 2s are claimed possible hosts, but
examples cited for 1s include only inactive 1s (e.g., Southern Tiwa). This
remains an area for further work.
RELATIONAL GRAMMAR 95

In sum, the Relational Succession Law and the Host Limitation Law
constrain ascensions with clausal hosts, but not possessor ascension. Posses-
sors may ascend in accord with the Relational Succession Law, but ascen-
sion to 3 must also be permitted. Further, it may be possible to restrict pos-
sessor ascension hosts to 2s.

2.7.2 Unions

'Union' refers to structures in which all (eligible) dependents of a subordi-


nate node assume grammatical relations to a dominating node. (On 'eligi-
bility', see below.) Hence, unions are like ascensions in that both involve
raising of some element. The two are distinguished by the fact that ascen-
sions have exactly one such element, while unions generally have more than
one.
The best known union is 'causative clause union'. Like other clause
unions, causative clause union involves two clauses, a main clause and a
subordinate clause. All (eligible) final dependents of the subordinate node
become dependents of the main clause node. As a result, the subordinate
clause ceases to bear any relation to the main clause.
The following examples from French and Turkish exemplify causative
clause union.
(79) On a fait sortir Jean de sa chambre.
'They made Jean come out of his room.'
(80) Elle a fait visiter la ferme à ses parents.
'She had her parents visit the farm.'
(81) Ayşe çocug-u cik- ar- di.
Ayşechild-ACCleave-CAus-PAST
'Ay§e made the child leave.'
(82) Hasan kasab- a et- i kes-tir- di.
Hasan butcher-DAT meat-ACC cut-CAUS-PAST
'Hasan had the butcher cut the meat.'
(79) and (80) are hypothesized to have the following structures.
96 JUDITH L. AISSEN

Note that in both cases, the complement predicate bears the Union ('U')
relation in the main clause, and that the complement clause itself bears no
final relation in the main clause. (83a, b) claim that the complement 1 of
(79) (Jean) and the complement 2 of (80) (la ferme) raise as 2, while the
complement 1 of (80) (sesparents) raises as 3 (and is hence marked with à).
Evidence of various sorts shows that the complement nominals do in
fact bear final relations in the main clause. In French, clitics associated with
such nominals cliticize to the main verb:
(84) OnI'αfait partir de sa chambre.
They made him come out of his room.'
RELATIONAL GRAMMAR 97

(85) Elle la leur a fait visiter.


'She made them visit it.'
In Turkish, a raised nominal which is accusative can passivize in the main
clause. Hence, both (81) and (82) have passive versions.
(86) Çocuk Ayşe tarafindan cik- ir- il- di.
child Ay§e by leave-caus-PAss-PAST
T h e child was made to leave.'
(87) Et kasab- a Hasan tarafindan kes-tir- il- di.
meat butcher-DAT Hasan by CUT-CAUS-PASS-PAST
T h e meat was caused by Hasan to be cut by the butcher.'
Assuming that only a dependent of the main clause can cliticize in the main
clause, and that passive is a clause internal advancement, facts like these
indicate that dependents of the complement have indeed been raised into
the main clause.
A major focus of research within RG concerns the discovery of princi-
ples which determine how grammatical relations are assigned in union con-
structions to the raised nominals. The French and Turkish examples illus-
trate a recurrent pattern, one in which a complement subject raises as 2
when it raises from an intransitive clause, but as 3 when it raises from a
transitive clause. This explains why çocuğu is accusative in (81), while
kasaba is dative in (82) ; and why Jean is unmarked in (79), while ses parents
is object of à. This also accounts for the clitic forms in (84-5), and for the
passive possibilities in (86-7).
These patterns, summarized in (88), were proposed in lectures by
Perlmutter & Postal (1974).
(88) a. Complement 1 of intransitive clause raises as 2;
b. Complement 1 of transitive clause raises as 3;
c. Complement 2 of transitive clause raises as 2;
Under early proposals, all other complement nominals (i.e., 3s, obliques,
chomeurs) were assigned the Dead relation in the main clause, and the
complement predicate was assigned the Union relation.
d. All other complement nominals raise as Dead;
e. Complement predicate raises as U.
The universality of (88b, c) was challenged by Cole & Sridhar (1977), who
argued from Hebrew and Kannada that complement ergatives can raise as
98 JUDITH L. AISSEN

2s. (88a) was challenged by Gibson & Raposo (1986), who argued that an
intransitive 1 can raise as 3. They also challenged (88d), arguing that, in
general, raised obliques and 3s bear the same relation in the main clause
that they bear in the complement clause. They termed this principle the 'In-
heritance Principle', and observed that it predicts (88c) as well.
Under the now generally accepted Gibson/Raposo proposal (1986),5
the complement 1 raises as an object, either 2 or 3, this choice being deter-
mined by particular grammars. All other nominals bear the same relation in
the union clause as that borne in the complement clause, unless the result
would violate the Stratal Uniqueness Law. So, for example, if the comple-
ment 1 in a transitive clause raises as 2, the complement 2 cannot also raise
as 2. If the complement 1 raises as 3, a complement 3 cannot also raise as
3. Gibson & Raposo propose that in such cases, the complement 2 or 3
raises as chomeur, and that in general, potential violations of the Stratal
Uniqueness Law are obviated through the chomeur relation.
This proposal allows an RN in which the first relation borne by a nom-
inal to a clause is the chomeur relation:

The complement 1 raises as 2, hence the complement 2 must raise as


chomeur. However, (89) violates the Motivated Chomage Law (see (8)),
which requires that the first stratum containing a Cho arc be preceded by a
stratum containing a term arc with the same elements as head and tail, i.e.,
that any chomeur bear some earlier term relation in the clause in which it is
chomeur. (89) contains no such term arc. Hence, a revision of the Moti-
vated Chomage Law is required to permit these cases. The Gibson/Raposo
proposal is summarized in (90).
RELATIONAL GRAMMAR 99

(90) Gibson/Raposo Union Proposal


Let x be the R-sign of a final (eligible) dependent a in the union
complement, and let y be the R-sign of the first arc it heads in the
union clause. Then:
a. If x = P,y = U ;
b. If x .= 1, y — 2 or 3;
c. If x ≠ 1,  =   or   = Ch o .
While clause union is most common when the main predicate is
'cause', there are other varieties. Johnson & Postal (1980) identify 'Equi
Unions' and 'Raising Unions'. Such cases involve both raising or equi, and
union. The following Spanish sentences illustrate these constructions (see
Aissen and Perlmutter 1983 for discussion):
(91) No se lo puedo dar a Marta.
'I can't give it to Marta.'
(92) Estos mapas no se quieren discutir.
' O n e doesn't wish to discuss these maps.'
(91) involves raising, while (92) involves equi. Both also involve union. The
complement objects in (91) must raise into the main clause in order to
cliticize in that clause. Estos mapas must raise into the main clause in order
to (reflexively) passivize there. In APG terms, these cases are characterized
by the fact that 'erasure' of the final complement 1 arc through raising or
equi is a necessary condition for raising all the other final complement
dependents into the main clause, i.e., for union.
The absence of a notion like 'erasure' makes it difficult to characterize
such structures in RG terms. But these cases show that there are cases of
union where all final complement dependents do not ascend into the main
clause. In (91), where the final complement 1 arc is 'erased', its head clearly
cannot raise into the main clause. Nonetheless, all other final complement
dependents ascend. Further, while the final complement 1 in (91) does
ascend into the main clause, it does so because the main predicate is a rais-
ing trigger, and raising of all other dependents is contingent on raising of
the subject. Hence, one wants to define union in such a way that it does not
require raising of all final complement dependents into the main clause, but
only a subset (the 'eligible' ones).
100 JUDITH L. AISSEN

2.8 Conclusion

Having sketched the main areas of RG proposals, it may be useful to indi-


cate several areas in which RG proposals are either lacking or are limited in
scope.
RG has no characterization of a structure which corresponds to the
surface structure of other theories. This should not be confused with the
false claim that RNs do not or cannot represent constituent structure. RNs
do represent constituent structure: each node corresponds to a constituent.
In general, however, the internal structure of constituents other than
clauses is suppressed since RG has few proposals regarding sub-clausal
structure. Nothing prevents the representation of internally-complex con-
stituents other than the present lack of relevant R-signs. Categorial infor-
mation is also generally suppressed because it plays little role in RG. But
little enrichment of RG is required to allow for labelling of nodes (see
Johnson & Postal 1980). What is lacking in RG is an explicit theory of how
particular RNs are associated with word order, surface bracketing, and a
phonetic representation. Consider, for example, an RN like (76a) in which
there heads two arcs. Clearly, only the main clause 1 arc is relevant to pro-
nunciation: there surfaces as main clause 1, not as complement 1. Nothing
in RG presently picks out that subset of arcs in an RN which determines
pronunciation. See Johnson & Postal (1980) for a proposal within APG,
however. Also lacking is any explicit proposal for the structure of preposi-
tional and postpositional phrases. In RG, there are no adpositions in initial
strata; hence these must somehow be incorporated into representations.
Again, see Johnson & Postal (1980) for a proposal.
As the references to Johnson & Postal (1980) should suggest, there is
no necessary obstacle to an account of surface structure within a relational
theory. It is simply that such proposals have been limited to APG.
RG has generally been silent on the question of the lexicon. The posi-
tion taken in Johnson & Postal (1980) is that there is no lexicon. This is con-
sistent with the more general fact that an APG grammar does not contain
components: there is no base component, no transformational component,
no semantic component, no phonological component. The representation
of a sentence — including syntactic, semantic, and phonological properties
— is a single object, and all rules (language-particular and universal) have
the same formal character: material implications.
RELATIONAL GRAMMAR 101

The claim that there is no lexicon should not be confused with the false
claim that the information usually found in the lexicon cannot be expressed
in a theory without a separate lexical component. In most theories with a
lexicon, the lexicon contains all conditions peculiar to particular lexical
items (e.g., subcategorization, morphological relations, meaning, pronunci-
ation). In APG, such conditions are formally expressed like all other rules
(material implications), but are restricted to particular lexemes. Sub-
categorization restrictions, for example, are conditions on the initial
stratum neighbors for particular predicates. Such a position appears consis-
tent with RG as well.
No phonological, morphological, or semantic analyses have been
articulated within APG or RG. Johnson & Postal (1980) suggest that the
theory of APG is rich enough to support the development of accounts of
phonology, morphology and semantics and they offer some preliminary
remarks about the nature of such representations. But these proposals
remain to be fleshed out.

Notes

1. Bell (1983) contains an extensive discussion of what she analyzes as advancement to 1


structures in Cebuano. However Gerdts (1988) convincingly reanalyzes analogous cases
in Ilokano as involving advancement to 2.
2. The following abbreviations appear in glosses: 3i 3rd person, noun class i; ABS Absolu-
tive; ACC accusative case; AOR aorist; ASP aspect;  AUS cause; DAT dative case; ERG
Ergative; ΙΟ indirect object voice suffix; NEG negative; PASS passive suffix; PAST past; PL
plural.
3. The question arises whether all forms of nominal anaphora are represented syntactically
with overlapping arcs. APG (Johnson and Postal 1980) has taken the position that they
are, while Rosen (1981) argues against such a unitary account of nominal anaphora.
4. However, the fact that they must be silent in Turkish remains unexplained, since
anaphoric pronouns are otherwise only optionally silent in Turkish.
5. Since this survey was written, a proposal by Davies & Rosen (1988) for Union construc-
tions has been widely adopted in RG work, superceding that of Gibson & Raposo. Davies
& Rosen propose that unions are monoclausal, and thus involve nothing like ascension
of complement dependents into a 'union' clause. The basic idea is that predicates
added to clauses, along with (roughly) the set of dependents subcatgorized by that predi-
cate. Thus a stratum may differ from the preceding one by virtue of containing a new
predicate and some additional nominal dependents. Union structures then are simply
those which contain more than one predicate, with the initial stratum containing the pred-
icate corresponding to that of the complement clause in the biclausal analysis. A predi-
cate added in a later stratum corresponds to the higher predicate in the biclausal analysis.
102 J UD I T H L. AISSEN

This proposal has several attractive features. First, it predicts the facts subsumed under
the 'Inheritance Principle' of Gibson & Raposo (formalized in (90c) below). 'Inheritance 1
is the unmarked case because 'complement' nominals now originate in the initial stratum
and simply 'fall through' into subsequent strata; no additional arcs are required and there-
fore there is no need to specify what grammatical relations are associated with those new
arcs. Second, it requires no change in the Motivated Chomage Law (see below) because
the conditions for chomage are all satisfied within a single clause, and Union construc-
tions thus fall under the scope of the Law, as originally formulated.
6. Another possibility is to recognize a distinct relation, say Dead, which obviates potential
violations of stratal uniqueness which arise in ascensions/unions, restricting the chomeur
relation to cases in which those violations arise in (clause-internal) revaluations.

You might also like