Ergativity R. M. W. Dixon

Linguistic Society of America
Ergativity
Author(s): R. M. W. Dixon
Source: Language, Vol. 55, No. 1 (Mar., 1979), pp. 59-138
Published by: Linguistic Society of America
Stable URL: http://www.jstor.org/stable/412519 .
Accessed: 27/06/2011 05:17
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=lsa. .
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Linguistic Society of America is collaborating with JSTOR to digitize, preserve and extend access to Language.
http://www.jstor.org
ERGATIVITY
R. M. W. DIXON
AustralianNational University
Morphological marking, whether realized by nominal case inflections or cross-
referencing on the verb, can be either (i) ergative, marking transitive subject [A]
function, vs. absolutive, marking intransitive subject [S] and transitive object [O];
or (ii) nominative, marking A and S, vs. accusative, markingO function. Absolutive
is always the unmarked term in an absolutivelergative opposition. Nominative is
most frequentlythe unmarkedterm in a nominativelaccusativesystem, but there are
some languages in which accusative is unmarked. A language whose morphology
mixes accusative and ergative marking has the split determinedby (a) the semantic
content of verbs, (b) the semantic content of NP's, (c) aspect/tense choice, or (d) a
combination of these.
A, S, and O are universal semantic-syntacticprimitives.A universal category of
'subject' can be defined as the set {A,S}, and is valid only for the level of deep
structure. Language-particular syntactic operations, such as coordination and
subordination,work in terms of a (shallow-structure)'pivot'; this is most often S/A,
but can be wholly or partly S/O (languages of the latter type are said to be ergative
at the syntactic level). A major function of antipassive or passive derivations is to
place A or O NP's (respectively)in derived pivot function, S. Many languages which
have some morphologicalergativityare entirelyaccusative(S/A pivot) at the syntactic
level. All languages that show syntactic ergativity have some morphological
ergativity.*
INTRODUCTION
1. 'Ergativity' is currently an 'in' term in linguistics. It is used by a wide variety
of linguists, with a whole range of different meanings. As a result, much confusion
exists at present about what an 'ergative' language is, and about the morphological,
syntactic, and semantic consequences of such a characterization.
* In preparing this paper, I have relied on information and ideas from many friends and
colleagues. A numberof the sections reflectideas which Michael Silversteinhad years ago, and
which he patiently tried, time and again, to convince me of. Bernard Comrie shared crucial
data and analytic ideas; Geoff Pullum provided a much-needed theoretical inspectorate;
Michael Halliday first explained to me what ergativitywas, and demonstratedits wide-ranging
relevance; Ken Hale fomented many ideas, unstintingly shared his data, and responded to
queriesconcerningit. The influenceof Anna Wierzbickahas been pervasive;she confirmedmy
view of the importance of semantics, and set an example in the use of semantic explanation.
None of these scholars is, of course, likely to agree whole-heartedlywith my use of their ideas.
I am gratefulto the following for providingcriticism and correction of a draft of this paper,
or for supplyingdata and ideas at an earlierstage: W. S. Allen, SteveAnderson, Avery Andrews,
Peter Austin, C. E. Bazell, Barry Blake, Dwight Bolinger, J. G. Breen, Victoria R. Bricker,A.
Capell, Ian Catford, Ray Cattell, Wallace Chafe, Sandra Chung, R. E. Cromack, Terry
Crowley, Diana Eades, Charles Fillmore, Bill Foley, Talmy Givon, Mary Haas, John Haiman,
M. A. K. Halliday, John Harris, Jeffrey Heath, Luise Hercus, Robert Hetzron, Rodney
Huddleston, Dick Hudson, Joyce Hudson, Paul Kay, Harold Koch, David Kilby, Hans Kuhn,
Ronald Langacker,MargaretLangdon, CharlesLi, Patrick McConvell, MaryalyceMcDonald,
Graham McKay, Sally McLendon, Francesca Merlan, George Milner, Pamela Munro, David
Nash, Geoff Pullum, Karl Rensch, Bruce Rigsby, Paul Schachter, Tim Shopen, Neil Smith,
Andrew Taylor, Sandra Thompson, Larry Trask, Robert Van Valin, Hans Vogt, Michael
59
60 LANGUAGE, VOLUME 55, NUMBER 1 (1979)
This paper surveys the various properties that have been taken as examples of
'ergativity', and attempts to arrange and explain them. In ??2-4 I deal with types of
ergativity at the morphological level. In ?2 I discuss the rationale for nominal cases,
and survey 'markedness' within both absolutive/ergativeand nominative/accusative
systems. In ?3 I put forward semantic explanations to account for split case-
marking; the occurrence of ergative and accusative systems can be motivated by the
semantic nature of any of the three types of obligatory sentence component-verb,
NP's or aspect/tense specification. In ?4 I briefly review the diverse ways in which
ergative morphologies can evolve.
In ??5-7 I consider syntactic ergativity, and the difficulty of defining a universal
category of 'subject' that holds over ergative as well as accusative languages. It is
suggested that 'subject' is not the most fundamental category. The basic thesis of
this paper, developed in ?5, is that A (roughly: underlying transitive subject), O
(underlying transitive object), and S (underlying intransitive subject) are universal
syntactic-semantic primitives. 'Subject' as a universal category, which can be valid
only at the level of deep structure, involves a grouping of A and S.
In ?5.4,I show that certain widespread syntactic phenomena, including impera-
tives and jussives, necessarily link together S and A; they are a consequence of the
universal category of 'subject', and should not be taken as evidence for the
typological classification of any individual language. In ?6, I discuss derived
structures. In ?7,I look at language-particularsyntactic phenomena and investigate
whether the 'pivot '-for coordination, subordination, and the like-groups
together NP's in derived S and A functions ('accusative syntax') or those in derived
S and 0 functions ('ergative syntax'). Antipassive and passive derivations feed
these syntactic operations, placing an appropriateNP in a derived function that can
be 'pivot' for relativization etc.
The label 'surface subject' has been customarily used for what I term 'pivot'.
This does not lead to any difficulties in accusative languages. In languages with
ergative syntax, however, considerable confusion can arise if 'pivot' (a language-
particular, surface-syntactic category) and 'subject' (defined as a universal deep-
syntactic/semantic category) are not clearly distinguished. I discuss the difficulties
involved in Keenan's discussion of 'subject' (1976) and in the application of
Relational Grammar and the Accessibility Hierarchy to ergative languages; all
these result essentially from a failure to distinguish 'subject' from 'pivot'.
In ?8 I summarize the varying factors that motivate morphological marking, and
attempt to explain why syntactic ergativity is rather rare, although some degree of
morphological ergativity is quite commonly encountered.
1.1. INTRODUCTORY
EXEMPLIFICATION.
A language is said to show ergative
characteristics if intransitive subject is treated in the same manner as transitive
Walsh, Deirdre Wilson, Frank Wordick, and Stephen Wurm.It is difficult adequately to thank
Paul Black, Bernard Comrie, David Rood, Alan Rumsey, Anna Wierzbicka, and Anthony
Woodbury-each of whom provided most detailed and perceptive comments on a draft of the
paper, and freely shared with me their theoretical insights.
The basic outline of the paper evolved during a productive sabbatical spent in the stimulating
atmosphere of the Linguistics Section, University College London.
ERGATIVITY 61
object, and differently from transitive subject. There are many ways in which this
'treatment' can be realized:' perhaps the clearest is in terms of case inflections.2
We can illustrate with the noun paradigm for the Australian language Dyirbal.
Here transitive subject is marked by ergative case inflection; this has the form -ygu
on disyllabic stems ending in a vowel. In contrast, intransitive subject and transitive
object functions are marked by absolutive case, with zero realization:3
(1) yuma banaga+nYu 'Father returned.'
(2) yabu banaga+ rnyu 'Mother returned.'
+
(3) yuma yabu ygu buma + n 'Mother saw father.'
+
(4) yabu yzuma ygu bura + n 'Father saw mother.'
The object in 3, yuma 'father', has the same form as the subject in intransitive
sentence 1. Note also that, in the normal word order of Dyirbal (which I follow
here), object precedes subject; a more revealing way of putting this is to say that an
NP in absolutive case occurs sentence-initially.4(The intransitive verb banaga-y 'to
return' in 1-2 belongs to the -y conjugation, and selects the allomorph -rnyufor the
non-future tense. Transitive bura-l 'to see, look at' in 3-4 belongs to the -I con-
jugation, and takes non-future -n.)
We shall need continually to refer to the three core semantico-syntactic relations,
and it will be useful to employ abbreviatory letters:
Intransitive subject: S
Transitive subject: A
Transitive object: 0
The groupings of these relations for a nominative/accusative case system (e.g.
Latin) and for an absolutive/ergative system (e.g. Dyirbal) are:5
(A ERGATIVE
(5) NOMINATIVE
> ABSOLUTIVE
ACCUSATIVE OJ
1 See Fillmore (1968:52) and the discussion in the remainder of this paper.
2 The term 'ergative' originated as the name for a nominal case, and has been progressively
extended to the other uses shown below.
3 Each NP in Dyirbal also contains a 'noun marker' that agrees with the head noun in case,
shows its noun (gender) class, and indicates whether its referent is 'here', 'there', or 'not
visible'. To simplify the discussion here, noun markers-which have a slightly irregular
paradigm-have been omitted; they do not in any way affect the grammatical points being
made. The 'there' forms of the masculine noun marker are ABS bayi, ERG baygul, DAT bagul; of
the feminine marker, ABS balan, ERG baggun, DAT bagun. Full forms of the Dyirbal sentences are
thus: (1) bayi yuma banaganYu;(2) balan yabu banaganYu; (3) bayi guma baagun yabuagu buran;
(4) balan yabu baygul gumaygu buran; (9) bayi yuma buralyan"u bagun yabugu; etc.
4 In fact, word order is very free in Dyirbal, syntactic relations being shown by case inflections.
There is, however, a normal order of constituents; it is adhered to in the examples quoted here,
simply for pedagogic effect. (See the comments below on position of pronouns.)
5 'Nominative' was until recently employed as the complement of ergative (covering S and 0
functions) as well as the complement of accusative (S and A functions). Because of the confusion
that this engendered, 'absolutive' has recently been adopted from Eskimoist terminology.
That case which includes S function is most often the unmarked term in the
system. Thus an absolutive NP occurs sentence-initially in Dyirbal, and is the pivot
for various syntactic operations. (Transformations may be applied to bring an NP
into a surface function where it will take absolutive case, to satisfy coreferentiality
conditions etc.) Absolutive has in Dyirbal a very similar syntactic/morphological
status to that which nominative has in Latin.
If any case in an 'ergative' language has zero realization, it will be absolutive
(as in Dyirbal). Similarly, it is nominative that most frequently has zero realization
in an 'accusative' system. (Note, though, that the parallel between absolutive and
nominative is not complete here. There are a few well-attested instances where
accusative is unmarked, while nominative involves a positive affix; see ?2.33).
Languages with ergative case-marking may be in a minority, but they are known
from every corner of the globe; to mention only a few of the better-known
examples: Eskimo (Thalbitzer 1911, Woodbury 1975,1977), Georgian (Vogt 1971),
Tibetan (Regamey 1954), Basque (N'Diaye 1970), Hindi (Allen 1951), and Tongan
(Churchward 1953; see Hohepa 1969).
Moving on to another grammatical level, languages can also be said to have
'ergative syntax'; i.e., some rules of coordination and/or subordination will treat
O and S in the same way, and A rather differently. We can again exemplify from
Dyirbal.
Two clauses can be coordinated in Dyirbal if they involve a 'common NP' that
is in surface S or 0 function in each clause (i.e. in absolutive case, in terms of the
noun paradigm given above). The occurrence of the common NP in the second
clause is usually deleted, and the whole biclausal construction can comprise one
intonation group. (There is no overt coordinating particle in Dyirbal, similar to
Eng. and.) Thus, from 1 and 3 we can derive:
(6)yuma banaga+ nYu yabu+ ygu bura+ n 'Father returnedand was seen by
mother.'
Similarly, from 3 and 1:
+
(7)yuma yabuaigu bura + n banaga+nYu 'Father was seen by mother and
returned.'
If we wish to conjoin 1 and 4, we find that the syntactic condition on coordination
is not met. The NP yuma 'father' is common to the two clauses; but although it is
in S function in 1, it is in A function in 4. In such cases, an 'antipassive' trans-
formation must be applied, deriving a construction in which an underlying A NP
is in derived S function, to satisfy the coreferentiality condition on coordination.
Antipassive can be stated as follows:
(8) NPX NPo V + tense
=> NPS NPDAT V + ya-y + tense
This transformation derives an intransitive sentence from an underlying transitive
construction. The original A NP becomes S (derived intransitive subject); the
original 0 NP now takes dative case (this has the realization -gu with nouns and
-ngu with pronouns) ;6 and the verb is marked by the antipassive derivational suffix
6 Instrumental(-ergative)inflection is possible here as an infrequentalternativeto dative.
ERGATIVITY 63
-ya-y. (Reasons for regarding an antipassive construction in Dyirbal as a derived

intransitive are given in Dixon 1972:150-51.)
The antipassive version of 4 is:
(9) yuma bural+ lia + nYu yabu+gu 'Father saw mother.'
Note that 4 and 9 have the same cognitive meaning, in the same way as an active
sentence and its passive transform in English.
Now 1 and 9 can be coordinated, in either order:
(10) yuma banaga + nyu bural + qa + nyu yabu+gu 'Father returned and saw
mother.'
(11) yuma bural+ ya + nyu yabu +gu banaga + nYu 'Father saw mother and
returned.'
Many languages which have an ergative morphology do not have ergative
syntax; instead, syntactic rules seem to operate on an 'accusative' principle,
treating S and A in the same way (see ??5-7, below, and Anderson 1976). Dyirbal
is unusual in that all major syntactic operations-those of relativization and
complementation, as well as of coordination-treat S and O in the same way.
It appears that there are no languages that are FULLYergative, at either the
syntactic or the morphological level. I discuss in ?5.4 certain constructions that
must have accusative syntax (in Dyirbal, as in all other languages) as a consequence
of the universal category of 'subject'. At the morphological level, no language
consistently uses one case-marking for S and O functions, but a different marking
for A, across every possible constituent of an NP and in every construction type-
and also has S and O cross-referencedin identical manner on verbs.7This provides
another point of difference between the absolutive/ergative and nominative/
accusative alternatives: there ARElanguages that consistently treat S and A in the
same way, in both nominal case-marking and verbal agreement (e.g. Lithuanian,8
Quechua, and Telugu).
At the morphological level, Dyirbal is by no means a fully ergative language; it
is rather of the type that has been called 'split ergative'. Although nouns and
adjectives (and 3rd person pronouns) inflect on an absolutive/ergative pattern, 1st
and 2nd person pronouns show a straightforwardnominative/accusativeparadigm,
as in Table 1.
ROOT yabu 'mother' guma 'father' yana 'we all' nYura'you all'
A function yabu + gu auma + ygu

pana nYura
S function
yabu auma
O function dana + na nrura+ na
TABLE 1
7 Both Basque and (Paleo-Siberian) Chukchee have absolutive/ergative case marking on all
nominal constituents. But both languages also show verb agreement with core NP's; and this
appears to operate on a partly accusative, partly ergative basis.
8 Standard Lithuanian has lost the neuter gender and dual number, in which S, A, and O fell
together in Proto-Indo-European. There are only a few isolated survivals of the neuter, e.g. tai
'that', used for S, A, and O functions.
Whereas nouns use the simple root for absolutive (S and O functions), and
show ergative case (A function) by -ygu, pronouns have only the root for nomi-
native case (S and A functions) and an affix -na for accusative (O). Sentences
involving pronouns are:
(12) dana banaga+ nu 'We returned.'
(13) n'ura banaga+n}u 'You returned.'
(14) n1urayana+na bura+n 'You saw us.'
(15) yana n9ura + na bura + n 'We saw you.'
Note that a nominative pronoun will, in normal word order, occur sentence-
initially. Sentences, and even individual NP's, can involve any mixture of pronouns,
nouns, and adjectives without possibility of confusion. The skeptical reader is
invited to construct a few examples for himself.
An exhaustive examination of the literature suggests that all languages which
have been described as ergative, at the morphological level, are in fact 'split
ergative'. Splits can be of a number of different kinds; these are described and
explained in ??3.1-3.6.
I have distinguished 'morphological ergativity', where one case marks both S
and O functions, from 'syntactic ergativity', where certain types of syntactic rule
identify S and 0. These are, potentially, independent parameters (possibilities of
connection are summarized in ?8). In some languages that have a partially ergative
morphology, all syntactic rules operate on an accusative basis, treating S and A in
the same way. And although pronouns in Dyirbal show nominative/accusative
inflection, the rule for pronoun coordination is identical to that for noun coordi-
nation: there must be an NP common to the two clauses, and it must be in S and 0
function in each clause. Thus we can coordinate 12 and 14 in either order (just like
1 and 3), giving:
(16) yana banaga +n9u nYura bura ? n 'We returned and were seen by
you.'
(17) n9ura yana + na bura+n banaga+nYu 'We were seen by you and re-
turned.'
The occurrence of the common NP in the second clause can be omitted, as in 6-7.
If we try to conjoin 12 and 15, we find that, although there is a common NP, yana
'we', with identical FORM in the two sentences, it is in S function in 12, but in A
function in 15.9 Transitive sentence 15 must be antipassivized, giving:
(18) yana bural+ 7ga + nYu nYura+ygu 'We saw you.'
Now 12 and 18 can be conjoined, in either order:
(19) yana banaga + nyu bural + Ca + nYu nYura + ygu 'We returned and saw
you.'
(20) yana bural + a + nYu n9ura ? agu banaga + nYu 'We saw you and re-
turned.'
9 yana banaga + nYulnvura+ na bura + n can only mean ' We returned and you were seen (by
someone other than us).' That is, it would have to be said with sentence-final intonation on
banaga + nYu; the second sentence would then be taken to have an unspecified A NP. These four
words could NOT be understood as 'We returnedand we saw you.'
ERGATIVITY 65
The rule for coordination in Dyirbal refers to the syntactic function of NP's, not
their forms: it demands that the common NP be in (surface) S or O function in each
clause. This coincides with absolutive case for nouns, but cuts across the morpho-
logical paradigm of 1st and 2nd person pronouns. Thus ianain 12 and yana in 15
have identical forms but different functions; they do not satisfy this syntactic
condition.
Rules for the formation of relative clauses and complements in Dyirbal also
show an entirely 'ergative syntax', identifying S and O functions (see ?7.2).
Dyirbal is undoubtedly MORE ergative at the syntactic than at the morphological
level. (But it does have SOME accusative syntactic characteristics. I return to this in
?5.4.)
1.2. TYPESOFERGATIVITY. It will be seen that the term 'ergativity' can be used
to describe phenomena at a number of distinct grammatical levels. We must
distinguish at least the following:10
(I) DEEPSTRUCTURE
refers to underlying semantic structures, and their universal
properties.
(I -+ II) SINGULARY TRANSFORMATIONS operate on deep structures (passive, anti-
passive, reflexive etc.) These yield
(II) SHALLOW STRUCTURE. NP's which have a certain deep function may have a
different 'derived' function at this level, on the basis of which
(II -1III) GENERALIZED transformationsmay operate. These derive various types
of coordinate and subordinate constructions, and yield
(III) SURFACE STRUCTURE. It is at this level that morphological marking takes
place-adding case affixes or pre-/post-positions to NP's, cross-referencingperson
(and/or number etc.) of certain core syntactic constituents onto the verb, and so on.
In the remainder of this paper, I consider the applicability and nature of 'erga-
tivity' at each of these levels.
Ergative phenomena are encountered most frequently (and are best documented)
at the surface-structurelevel. In ?2-4, I survey types of morphological ergativity,
and put forward semantic explanations for the kinds of split system that occur. In
??5-7, I discuss deep and then shallow structures, separating universal syntactic-
semantic phenomena from language-specific constraints that are true criteria for
syntactic ergativity. Finally, in ?8, 1 return to morphological marking, relating it to
syntactic and semantic needs.
MORPHOLOGICAL ERGATIVITY
2.1. TYPESOFMORPHOLOGICAL
MARKING.There are three main ways in which the
function of an NP in a sentence can be shown.
2.11. CASEINFLECTIONS
can be used, as in such well-known languages as Latin and
Greek, and as illustrated above for Dyirbal. The ways in which case is marked on
an NP can vary: the inflection can occur just on the head word, or just on the last
10 I owe a considerable debt to Michael Silverstein for pointing out to me that any discussion
of ergativity must distinguish at least these levels. He may not, of course, approve of the way I
have set them out here.
word, or on every word (or sometimes, on every word ONLYIF they are non-
contiguous, being distributed through the sentence). These details are essentially
irrelevantto the present discussion. What is significantis whether case is obligatorily
marked on an NP (as in Latin and Dyirbal); or whether it is optional, being
included only when ambiguity would otherwise result (see ?2.3).
2.12. SEPARATE i.e. prepositions or postpositions, can mark syntactic
PARTICLES,
function. Particles show essentially the same possibilities as cases. Since a particle
usually has the phonological status accorded to a 'separate word', there will
normally be only one occurrence of the particle in an NP, whereas a case inflection
may be added to each word; but this is not significant for the present discussion.
What can complicate the syntactic picture is the tendency of particles to combine
information about syntactic function and 'topic', as in Japanese.
2.13. THE VERBOR A VERBALAUXILIARYmay include some indication of the
person, number, gender etc. of NP's in certain syntactic functions. There is tremen-
dous variation as to how much information is 'cross-referenced' in the verb, and
how it is realized. English shows minimal cross-referencing: most verbs have two
'present tense' forms, indicating whether or not the surface subject is 3rd person
singular (e.g. walks vs. walk). A more complex pattern occurs in Swahili, where the
verb contains separate affixes referring to subject and object. These specify the
person, number, and noun (i.e. gender) class (if 3rd person non-human) of each NP
cross-referenced; some affixes also have different forms depending on whether the
sentence is affirmativeor negative.
The form of bound pronominal affixes in the verbal word can be taken as
evidence of morphological 'accusativity' or 'ergativity', just like the form of case
inflections. If a certain affix cross-references an NP that is in S or 0 function (with
a different affix referring to an NP that is in A function), then the language could
be characterized as 'ergative' at this level; Abaza (Caucasus) is such a language.11
Consider these sample sentences in Swahili:
(21) Sisi tulianguka 'We fell down.'
(22) Ninyi mlianguka 'You all fell down.'
(23) Ninyi mlituona sisi 'You all saw us.'
(24) Sisi tuliwaona ninyi 'We saw you all.'
These reveal a pronominal affix paradigm:
SUBJECT(S/A) OBJECT(0)
(25) we tu- -tu-
you all m- -wa-
Note that the same form, -tu-, is employed to cross-referenceS, A, and 0 in the 1st
person plural. The existence of one form for S/A, but another for 0, in the 2nd
person plural (as well as 2nd person singular and 3rd person singular) establishes a
'nominative/accusative' pattern. Perhaps the most important detail here is the
11 In Abaza, S and 0 share an identical system of markers, occurring in initial position; the
A system occurs in non-initial position, but differs from the S/O system only in the 3rd person
and the relative pronoun. Note also that as many as four Abaza NP's may be marked in the
verb-e.g. causative agent, subject, object, and indirect object, as in 'The old man couldn't
make the boys give the girl her dog back' (Allen 1956:139 and p.c.)
ERGATIVITY 67
POSITION of the pronominal affix within the verb word: bound forms which cross-
reference S or A occur word-initially, whereas the affix which refers to the O NP
comes between tense (here, 'past' -li-) and the root (in these examples, intransitive
-anguka 'fall' or transitive -ona 'see').
The free-form pronouns sisi 'we' and ninyi 'you (pl.)' would normally be
omitted from 21-24; they are used mainly for emphasis. But where A and O NP's
are identical as to information that is cross-referenced, as in 'The man saw the
boy', full NP's would of course be required.
Bound forms which cross-referencethe core NP's are usually, but not invariably,
attached to the verb; when they are attached to the verb, they can either be affixes
fully integrated into word structure(as in Swahili, where the object affix intrudes
between tense and root), or they can be more loosely connected, as clitics. The
alternative is for the bound forms to make up a separate constituent (cf. the
'auxiliary' in Walbiri of Australia; Hale 1973)-or, say, to be attached to the first
word of the sentence; in the latter case, the affixes usually have the status of clitics.
2.14. Although case inflection (or particles) and bound pronominal cross-
referencing affixes are both morphological indicators, and can both be said to
pattern either 'ergatively' or 'accusatively', it is by no means obvious (on a-priori
grounds) that they are equivalent. When surveying kinds of 'split ergative' systems
in ??3.1-3.6, we must be careful to note the type of morphological marking
involved, and to see whether a particular sort of conditioning for the split system
can apply equally to both types (see also ?2.3).
2.15. There are, of course, languages which do not use morphological means to
mark syntactic function, but instead utilize contrastive word order. Although this
falls within the traditional division of syntax, contrastive word order plays the
same role as case inflection or bound pronominal affixes, and should show the same
typological possibilities.
We could suggest that a language which employed the order AVO for transitive,
and VS for intransitive sentences, is working on an 'ergative' principle: both S and
O follow the verb, and A precedes it. Similarly, AVO/SV or OVA/VS would be
'accusative' (and OVA/SV another 'ergative' variety). Note that this sort of
categorization may be possible only for verb-medial languages: thus, with orders
VAO and VS, one could argue either that S and A are treated in the same way
(since they both immediately follow the verb), or that S and 0 are equivalent (since
they both occur finally).
With case or cross-referencing systems, we basically have three possibilities:
(a) S and 0 are treated the same; (b) S and A are treated the same; or (c) S, 0, and
A are all treated differently. With 'contrastive word-order' languages, if we
LOOKEDfor some patterning (in verb-medial languages, where this question can be
asked), we would be confined to (a) or (b). The criteria are much slighter than for
cases or for cross-referencingaffixes, and there is a risk of forcing a characterization
of a language as 'accusative' or 'ergative' where none is really justified. (Cf.
Swahili, where the morphological positioning information-S/A prefix before tense
affix, 0 prefix between tense and root-is reinforced by the FORMALidentity of S
and A cross-referencing forms, and the partial dissimilarity of 0 forms.)
For these reasons, 'accusativelergative' characterization based on word order

should be viewed with caution, unless corroborated by evidence of some quite
different sort (or unless there are something like conditioned rules for ordering that
yield an unequivocal pattern).l2 Note that word order, in ALLlanguages, is exploited
to convey information about 'topic' or 'theme'; it can be hard to separate these
various roles.
2.2. SEMANTICBASES.It is a simple matter to provide a-priori semantic justifi-
cation either for absolutive/ergative or for nominative/accusative morphological
marking. For the first alternative, we could suggest that an intransitive sentence
essentially describes 'something happening to something/someone' (e.g. The bough
broke), whereas a transitive sentence describes 'someone making something happen
to someone/something' (e.g. The man broke the bough.) By this argument, S is
semantically identified with 0, and it is natural for them to be marked by the same
case inflection. Alternatively, we could suggest that an intransitivesentence involves
'someone doing something' (e.g. I'll fight tomorrow), and a transitive sentence
'someone doing something to someone/something' (e.g. I'll fight John tomorrow.)
With this semantic characterization, it is surely natural to use a single case to mark
S and A functions.
The explanations given in the last two paragraphsare mutually inconsistent; and
each is oversimplistic. They imply that it is sufficient to talk in terms of 'transitive
verbs', 'intransitive verbs', and 'noun phrases' (filling S, 0, and A functional
slots). But there is a wide range of semantic variation within the verb class and
among NP's; and note that the semantic type of a verb shows a strong correlation
with the nouns/pronouns that would be expected to co-occur with it. Words from
certain lexical domains behave like those used in the first paragraph above: break,
open, wake can function transitively or intransitively in English, with correspon-
dence between the NP's that fill 0 and S slots. Words in other lexical domains
behave as shown in the second paragraph: fight, eat, follow, answercan also occur
with either transitivity value, but here it is A that corresponds to S. (It is certainly
the case that EVERY language mingles 'ergativity' and 'accusativity' in the
structure of its lexicon. In this paper, I am concerned with the ways in which S/0
and S/A equivalences are established at the morphological and syntactic levels;
but, as I shall show in ?3.1, it can sometimes be difficult to draw the line between
lexical and morphological characteristics.)
The question posed here is WHYthere should sometimes (in some languages, or
in some places in some languages) be nominative/accusative case-marking, and at
other times absolutive/ergative marking. And a prior question is: WHY SHOULD
THERE BE CASE-MARKINGAT ALL?
We can first note an empirical universal. All languages appear to distinguish
activities that necessarily involve two participants from those that only necessarily
involve one (I use 'participant' to describe a person, an animal, or a thing-in fact,
12
In fact, among languages that are, on other grounds, 'ergative', very few show a preferred
verb-medial word order. Tsimshian (see ?3.5) appears to show basic orders AVO and VS (Boas
1911:298). However, Rigsby (1975:353) suggests that in Nass-Gitksan, a Tsimshian language,
the order AVO, which applies only in embedded clauses, can be derived from the basic main-
clause ordering VAO.
ERGATIVITY 69
any referent of an NP). Then all languages have classes of transitive and intransitive
verbs, to describe these two classes of activity. In some languages, there is a very
clear division of verbs into transitive and intransitive types, and some derivational
affix is required to 'mark' a basically intransitive verb used in a transitive con-
struction, or vice versa. In other languages, the extremes will be clear enough, but
there may be some middle ground where it is difficult to decide whether a given
verb is basically transitive or basically intransitive (or both, or neither); English
falls into this category with respect to verbs like break andfight (see ?5.2 and ?6.3,
fn. 94).
Now the core constituents of a transitive sentence are a verb and two NP's. The
critical point-and the ultimate basis for any system of case assignment etc.-is
the need to distinguish which NP is which: i.e., which has A and which O function.
If case inflection is adopted (rather than, say, contrastive word ordering), then one
of the core transitive NP's must receive a non-zero desinence. The other NP can
also receive a positive inflection, or it can be left unmarked. (The contrast between
marking and lack of marking means that there is a system of TWOcases, one having
zero realization.)
Since an intransitive sentence contains only one core NP, in S function, there is
no need for any positive marking on it (as there is need for positive marking on
ONEof the core NP's in a transitive sentence). Certainly, there is no need for an S
NP to receive a case-marking that differs from both A and 0 possibilities: an S
item occurs in a different sentence-type from A and 0 phrases, and can most
economically receive the same marking as EITHERA OR0. If either A or 0 is shown
by a case that has zero realization, then we might expect S also to be phonologically
unmarked; this is the most frequent situation (although there is a set of languages
in which both A and S receive positive marking, and 0 is unmarked; see ?2.33).
If both A and 0 involve a non-zero inflection to the nominal root, then S could
conceivably fall together with either of them.
In fact scarcely any language is known that consistently, across all nominal
constituents, marks S, A, and 0 differently (although a fair number of languages
have three distinct inflections for a SMALLSUBCLASS of NP constituents; see ?3.23).13
13 Motu has been taken to have differentmarkingfor A, S, and 0, entirely on the basis of the
slender data given by Capell (1969:36) As Lister-Turner& Clark (1930:34 ff.) point out, the
syntactic function of an NP in a Motu sentence is 'indicated by the demonstrativeadjectivesor
articles ese, se, be, and na; by suffixes; by prepositions; and sometimes by the position of the
word in the sentence'. The ergativeparticleese is used when it is not clear on semantic or other
grounds which NP is A, and which is 0 (see ?2.31). The particle na, said by Capell to mark S
function, has complex behaviorwhich is far from being fully understood-it can be used as the
copula in verbless sentences; it may occur after an S NP; and it is occasionally found after an
0 NP, most commonly when the word order is OAV rather than the more usual AOV. In
addition, na is sometimes encounteredafter an A NP, with the sequence ese na being attested.
The use of na characterizesthe western dialect; it is used much less in eastern Motu. (I am
grateful to Andrew Taylor for all this information.)
The only languages that do appear to have different marking for S, A, and 0, across NP
constituents, are the 'Ngura group' of the Australianfamily. Breen 1976 reports that, in Wang-
kumara, suffixesmarkingA, S, and 0 are bound forms of 3sg. pronouns; in an earlierstage of
the language there may have been a split case system, of the type described for Cashinawa
(?3.23), with distinct A, S, and 0 forms only for the free-form 3rd person pronouns. Data on
Similar remarks apply for cross-referencingpronominal affixes: either the A or the

0 form is also used to refer to S NP's, in almost every instance.14
We have established that, for a case-marking language, there should be distinct
inflections for NP's in A and O functions; these can be referred to as Ca and CO
respectively. And we have noted that one of these cases will normally mark an NP
in S function in an intransitive sentence. IfS is marked by Ca, we have a nominative/
accusative system; but if COis used, the paradigm is of the absolutive/ergative
variety.
There must be a reason for the use of ONEof the transitive cases for the intransi-
tive core NP. Why is Ca used in some circumstances, and C,Oin others ? The answer
to this question lies in the range of semantic types among intransitive verbs.
That NP whose referent can initiate and control the activity (if anything can) is
recognized as being in A function in a transitive sentence. The core NP whose
referent cannot be controller is taken to be in O function. (Fuller discussion of these
points is given when I deal with the notion of'subject', in ?5.2.) For some intransi-
tive verbs, the S NP CANbe controller of the action (e.g. jump, speak); this could be
described as Sa. It is then semantically plausible for Ca to be used for Sa, as for
A (cf. John spoke and John told Mary). For other intransitive verbs, the S NP can
not control the action (e.g. break, die, yawn), and has the same semantic status as
an O NP ('that which something happens to'); we can refer to this as subtype SO,
of S. It is now natural to use COfor both 0 and SOfunctions.
A nominative/accusative case system always uses one case (Ca) for A and S
functions, making no distinction at all between Sa and SO.The semantic relation
that holds for verbs like jump and speak is generalized to hold, as a GRAMMATICAL
relation, for all intransitive verbs. Similarly, an absolutive/ergative system
generalizes from the semantic relation that holds for verbs like break and die, and
provides GRAMMATICALIDENTIFICATION of 0 and S (through a common case
inflection, CO)over all intransitive verbs; it too ignores the putative distinction
between Sa and SO.
It is convenient to have a grammatical generalization here, of either sort. Our
examples have been of clearly 'controlled' or definitely 'non-controlled' verbs.
But these lie at the extremes of a continuum, with most intransitive verbs some-
where in between. With laugh, the activity is sometimes involuntary, but at other
times contrived and controlled; cry is similar, although most often it may be
involuntary, and more difficult to contrive (there will be differences from culture
the related language Galali suggest a similar state of affairs, except that there the accusative (0)
and nominative (S) suffixes seem to be used predominantly to mark definite NP's (see McDonald
& Wurm, MS). Note, though, that the data on both these languages are slim, comprising
isolated sentences elicited from the last speakers (whose fluency was not perfect), and that it was
not possible to carry out exhaustive checking.
14 Exceptions are found where a single 'portmanteau' affix refers to person/number/gender
of both A and 0; these cannot be analysed into separate A and 0 forms, and cannot be related
to the S affixes. But in most such cases it is possible to reconstruct distinct A and 0 forms, at
least one of which is related to original S forms. Thus Hinton & Langdon 1976 explain the
development of portmanteau prefixes in modern Yuman languages from series of free pronouns
in the proto-language.
ERGATIVITY 71
to culture and from person to person). Whether or not a verb is controlled often
depends on the semantic nature of the referent of the S NP. As Lyons points out
(1968:350-65), in his exemplary discussion of this topic (on which the present
section is closely based), It moved would be taken as non-agentive, while He moved
could be either agentive or non-agentive (Lyons suggests 'ideal' representations
He moved for the agentive and Him moved for the non-agentive sense). Thus, for
mariy intransitive verbs, it is difficult to decide whether they basically belong to the
'controlled' or the 'non-controlled' class; i.e., it is difficult to determine whether
the S NP is of subtype Sa or SO,.Most languages avoid decisions in this area of
semantic fuzziness by simply generalizing either Ca or COto mark S, for all intran-
sitive verbs.
I have already noted that, although there are some 'pure' nominative/accusative
languages (which is certainly related to the universality of 'subject', and the crucial
role this category plays in the grammar of every language; see ?5), there is no
language with 100%7ergative morphology. Every so-called ergative language has a
split case system (or whatever), mixing ergative with accusative, and/or with a
three-way case assignment (having separate indications for each of S, A, and 0).
These 'splits' can be conditioned by a variety of factors.
First, and most obvious, there can be a split according to the semantic content
of verbs; some languages avoid marking every S NP by Ca or CO,and use one or the
other transitive desinence according to the semantic type of the intransitive verb-
roughly, employing Ca for Sa, and COfor SO.
A second type of split is conditioned by the semantic content of the NP's
(remember that it is NP's which bear case inflections). Something that can function
as controller of an action (a human, or perhaps a higher animal) is unmarked in its
normal S or A function, but the NP receives a non-zero case inflection in 0
function; similarly, something inanimate, which would not normally initiate or
control any activity, may be unmarked in S or 0 function, but receive a positive
desinence if it does happen to occur in A function.
A third type of split can be conditioned by a further semantic component of an
sentence-the tense/aspect choice. Something that is complete can be viewed either
from the point of view of the patient ('Something happened to X') or of the agent
(' Y did something'); but a prospective activity is best viewed in terms of a proclivity
of an agent. In the latter case, there is pressure for S and A to be dealt with in the
same way.
In ??3.1-3.5, I discuss each of these types of conditioning factor for 'split
ergative' morphologies.
2.3. MARKEDNESS. We have seen that case inflectionsexist primarilyto distinguish

between the two core NP's in a transitive sentence. The simplest way to do this is
just to mark ONEof the semantic-syntactic functions A and 0, leaving the other
unmarked. One alternative is to use ergative case to mark A function-that NP
which is or can be the agent; then the 0 NP is left unmarked, as is the S NP (and
these are referred to as being in 'absolutive' case). The other alternativeis to use a
positive accusative inflection to mark 0 function-that NP which can NOTinitiate
or control the action; under this scheme, the A NP is left unmarked, like S (and
these are said to be in 'nominative' case). Note that here the 'identification' of O
with S, or of A with S, is quite negative in character; S is always left unmarked, and
so will fall together with whichever transitive NP is not marked. Such a residue-type
identification surely has little grammatical significance; this gives some explanation
of why morphological ergativity/accusativity does not always correlate with
syntactic ergativity/accusativity (the latter can only be recognized in terms of
positive behavior; see ?7). Discussion of this topic will be continued in ?8.
I am suggesting that 'absolutive' is always the unmarked term in an absolutive/
ergative system, and that 'nominative' is always unmarked within a 'nominative/
accusative' system (it will be necessary later to modify the last 'always' to' usually').
Generally, the case that is functionally (i.e. morphologically) unmarked is also
formally (i.e. phonologically) unmarked-i.e., it has zero realization; this is so for
absolutive in Dyirbal and nominative in Telugu, among many examples. BOTH
cases-ergative and absolutive, or accusative and nominative-may involve some
non-zero desinence added to the basic nominal stem; i.e., both may be phono-
logically marked. In such languages, it is still usually true that the absolutive or
nominative (that case whose scope includes the S function) will be MORPHO-
LOGICALLY unmarked. Thus, in Latin, a nominative form (e.g. servus) is gram-
matically unmarked with respect to accusative (e.g. servum); it is the nominative
that is used in syntactically unmarked circumstances such as citation.'5
My thesis is that ergative is always the marked term in an absolutive/ergative
system, and that accusative is usually the marked member in a nominative/
accusative opposition. There is a further possibility, less common but quite
adequately attested, whereby nominative can be morphologically marked with
respect to accusative. We can now exemplify these three possibilities, one at a time.
2.31. ERGATIVE ASTHEMARKED CASE. It is not uncommon to find an ergative case
inflection described as 'optional'. For Motu, an Austronesian language of coastal
New Guinea, 'ese is the transitive subject particle ... it need not appear when there
is no possibility of the object NP being taken as the subject' (Taylor 1970:30). In
'The boy saw the girl', ese will be included after 'the boy'; but this particle is not
required, and is unlikely to be included, in 'The snake bit the boy.' Another
15 That case form which includes S within its list of functions will be the unmarkedterm, in
most instances.It is always this case form that is used for the 'topic' of equational sentences-
which show only a copula, or (in many languages)no verbat all: e.g. Dyirbal yindabulgan'You
are big', (bayi) pumabulgan' Father is big.'
It is the unmarkedcase form that is employed in citation (indeed, this is one criterion for
markedness).But one or two languagescustomarilymark A, S, and 0 by non-zero inflections,
and employ the bare stem in citation. In Creek,e.g., nominative(S/A) is -t and accusative(0) is
-n; the bare stem is used in citation and for a title, spoken at the beginningof a story. But a bare
stem is sometimes found at an A, S, or 0 slot within a sentence, if the function of the NP is
clear from the context etc. (data on Creek from Mary Haas). It seems that the Creekinflections
-? and -n are USUALLY included on core NP's, but they are NEVER used in citation. All known
languages which have OBLIGATORYinflections for A, S, and 0 must use one inflectional form
(never the bare stem) in citation.
The use of case in Kemant, a Cushitic language, shows some similarities to the Creek
situation (see Hetzron 1976:16 ff.); but insufficientdetail is available to check whetherthe bare
citation form can be used in core function, in place of a normal inflected form.
ERGATIVITY 73
example is Murinypata, from the Port Keats area of North Australia. Here there
are pronominal prefixes to the verb, cross-referencingS/A and O NP's by (roughly)
person, number and gender. There is also an ergative inflection on nouns, -te --e;
but this is normally used only when information about which NP is in A, and which
in O function, is not inferrablefrom either (i) the semantic nature of the NP's and of
the verb, or (ii) the cross-referencingprefixes. That is, -te - -ie is likely to be used
when A and O NP's are both 3rd person and both make the same choice from
masculine/feminine; it is unlikely to be included in most other circumstances
(although it CAN always be present).16 Note that, although Murinypata has an
ergative nominal inflection, verbal prefixes work on a 'nominativelaccusative'
paradigm, with one series for S or A, and another for O reference. (Data on
Murinypata are from Walsh 1976a,b, and p.c.)
My assertion that ergative is used to 'mark' an A NP (with O and S NP's being
simply left without any marking) is supported by examples such as these. The
ergative particle (in Motu) or inflection (in Murinypata) is normally used only
when the identity of the A NP can NOTbe inferred from any other grammatical or
semantic information in the sentence.
In most languages in which the ergative occurs, it is obligatory; an NP of a certain
semantic type (see ?3.2) MUSTtake ergative inflection when it is in A function. But
there can still be evidence that it is grammatically the 'marked' case, as in YidinY,
from North Queensland. YidinYappears to work on these principles: (a) that NP
which is marked by ergative case is the 'controlling agent' of a transitive action;
and (b) that NP which is the controlling agent of a transitive action is (if non-
pronominal) marked by ergative case. Any deviation from this is shown by a
derivational affix-.:dYi-n on the verb (coming between the verb root and the final
tense-type inflection).
Consider a regulartransitive sentence in Yidin5:
(26) wagudYa+ ,gu dYugi+0 gunda+1 (galba:n+ da)
man + ERG tree+ABS cut+PRES axe+INST
'The man is cutting a tree (with an axe).'
Absolutive case (covering S and 0 functions) has zero realization; ergative is here
-ygu, and instrumental -da. (In fact, instrumental has the same realization as
locative, but there are important syntactic criteria for distinguishingthe two cases.)
Present tense is -l with verbs from the predominantly transitive -l conjugation, here
gunda-l 'cut'. (The occurrence of vowel length in these examples is inserted or
deleted by regular phonological rules; see Dixon 1977a,b.)
For 26, both (a) and (b) are satisfied: wagudYa 'man' is the controlling agent of
the activity, and is marked by ergative case. But consider 27, the antipassivized
counterpart. Here the deep A NP is brought into surface S function (normally, to
meet syntactic conditions on subordination and coordination), and the deep 0 NP
receives locative inflection; a number of grammaticaltests show that an antipassive
construction must be considered intransitive.
16 Other examples of 'optional ergative' include the Australian

language Dalabon (Capell
1962:111), and Hua of the New Guinea highlands(Haiman,MS).
(27) wagu-da + 0 gunda+ dYi+r dyugi+ :1 (galba: n + da)

man + ABS cut + :dYi+ PRES tree + LOC axe + INST
'The man is cutting a tree (with an axe).'
Here condition (b) is broken: wagudcais still the controlling agent of the verb; but
in this derived intransitive construction, it does not receive ergative marking. The
infringement of this condition is shown by -:dYi-n on the verb. (Note that -y is
present-tense inflection on the predominantly intransitive -n conjugation; -:1 is the
locative inflection on dyUgi'tree'.)
Similarly, a reflexive sentence-with the agent intentionally doing something to
himself-is also a derived intransitive:
(28) wagu:dYa+0 gunda+ : dyi+ y (galba:n + da)
man + ABS cut + :dYi+ PRES axe + INST
'The man is cutting himself (with an axe) [on purpose].'
Here wagudYa is in derived S function and absolutive case; but it is the 'deep A'
NP of an underlying transitive verb; -:dYi-n is again included, to indicate that the
controlling agent is not in ergative inflection.
It is important to note that a construction like 28 indicates a purposeful reflexive
activity. Eng. The man cut himself could also be used of an accidental injury; this
must be rendered in Yidin5 by the following, which differs from 28 primarilyin that
galban 'axe' takes ergative -du, rather than the instrumental inflection -da (the
allomorphs that occur after a stem ending in n):
(29) galba: n + du wagu: dva+ 0 gunda+: dYi+ y
axe + ERG man + ABS cut +: d i + PRES
'An axe cut the man (= The man cut himself on an axe, accidentally.)'
Here the man could have injured himself by accidentally standing on the axe, or
letting it drop on his foot, or nicking himself in the neck while swinging it back.
Now 29 is, by an array of syntactic tests, a transitive sentence. Indeed, it contains
an ergative (A) and an absolutive (0) NP. But the ergative inflection is not here
marking a 'controlling agent' (there is NO controller, for an accident of this sort),
and condition (b) is broken; thus the verb is marked by -:dvi-.l7
As a final example we can contrast the following:
(30) wagudva+ygu bana+0 wawa+/
man + ERG water + ABS see + PRES
(31) wagudva+ygu bana+0 wawa+ :dvi+y
Sentence 30 is, like 26, a normal transitive construction,18 'The man sees the
17 In fact, -:dYi-n is used to mark an inanimate agent only with transitive verbs from the
'affect' semantic class ('hit', 'cut', 'split', 'spear', 'burn' etc.; full details are in Dixon 1977a:
287).
18
Like Dyirbal, YidinY has absolutive/ergative case-marking for nominals, but a nominative/
accusative paradigm for 1st and 2nd person pronouns. Thus the S/A of the lsg. pronoun is
yayu; the 0 form is yanYan9. The transitive/intransitive status of the sentences given here can be
seen from examination of the pronominal equivalents:
(26') yayu dvugigundal (galba: nda) ' I am cutting a tree (with an axe).'
(27') yayu gunda: dviy dougi:l (galba:nda) 'I am cutting a tree (with an axe).'
(28') yayu gunda: dvi (galba:nda) ' I am cutting myself on purpose (with an axe).'
ERGATIVITY 75
water', with the presumption that he was looking for some water, and found it.
Sentence 31 is also transitive; unlike 29, it has an ergative NP with human reference
that COULDbe the controlling agent. But the insertion of :-dyi- into the verb of 31
indicates that in this instance 'the man' did not simply do what he had set out to do.
In contrast to 30, it means 'The man sees the water accidentally': he may have
been, say, chasing a dog or looking for a place to defecate, when he came across a
stream of fresh water. In 29, the NP COULDNOT have controlled the activity; in
30-31, he COULD-but, just in case he achieves some result by pure chance, -:dYi-n
is inserted in 31, to mark the non-satisfaction of condition (a).
It will be seen that -: dyi-n has a wide range of uses.19In 27-28, it marks a derived
intransitive construction (antipassive in 27, reflexivein 28); in 29 and 31, it indicates
that, in a transitive construction, the referent of the ergative NP does not control
the activity. These apparently diverse syntactic and semantic effects can be related
as involving non-satisfaction of conditions (a) and (b). It appears that the YidinY
ergative definitely MARKSone NP in a transitive construction as 'controlling agent'
for the activity described by the verb. Here the ergative contrasts with the unmarked
absolutive case, which appears on S and O NP's.
In many ergative languages, the absolutive NP must obligatorily be included in
each sentence, but an ergative NP may be deleted (this holds for Dyirbal and for
Eskimo; Woodbury 1975:113); this provides further support for absolutive as the
'unmarked' and ergative as the 'marked' case. In every ergative language known
to me, the absolutive is the sole citation form.
2.32. ACCUSATIVE AS THE MARKED CASE. In most languages that have a nomi-
native/accusative case system, it is the nominative that is morphologically un-
marked; if any NP is obligatory in a sentence, it will be the one in the nominative
case. Nominative forms are used in citation.
Accusative is then the 'marked' case. Paralleling the ergative examples above,
there are instances where an 0 NP need not (and does not) receive accusative
marking when other factors show which NP is in A and which in 0 function. Thus in
Finnish the 0 NP usually receives the accusative inflection -n; but in a 1st or 2nd
person imperative construction, there is no overt expression of the subject, and
here the 0 NP does not take accusativeending. However, in a 3rd person imperative
(e.g. 'Let him eat the fish! '), there can be overt expression of the subject, and here
the direct object is in the accusative case (Comrie 1975:115-16, Moreau 1972).20
(29') galba:ndu taanYanY gunda:d9ig 'An axe cut me (= I cut myself on an axe,
accidentally).'
(30') rayu banawawal 'I see the water(that I was looking for).'
(31') yayu bana wawa: d9if 'I see the water (by chance).'
19An additional, minor use of-:d9i- is to mark some activity as 'continuous'; in this sense,
it can be added to a transitive or an intransitive stem, and preservesthe transitivity. There
appears to be no connection between this sense and those given above (a full discussion is in
Dixon 1977a:273-93).
20
The Finnish accusative has (to my mind, rathermisleadingly)been termed 'anti-ergative'
by Comrie 1975. Comrie uses 'anti-ergative' for an object inflection that applies only when a
subject is present; this is seen as the mirror-imageof ergative, which is a 'subject' inflection
applying only when an object is present (i.e. only in transitive sentences). This ignores the
A similar situation prevails in other Balto-Fennic languages (Collinder 1965:

54-5), in Australian 'nominative/accusative' languages like Ngarluma (O'Grady,
Voegelin & Voegelin 1966:102) and Lardil (Klokeid 1976:197),21and in Southern
Paiute (Sapir 1930:179-81, 235) and other Uto-Aztecan languages.22In each of
these languages, the nominative is morphologically and phonologically unmarked,
whereas the accusative involves a positive affix. Thus the accusative can be thought
of in terms of a special marking of the object that can be omitted whenever its
identity can be inferred in certain other ways; if the subject of an imperative must
be 2nd person, then (whether or not this subject is expressed) any non-2nd person
core NP must be in O function.
2.33. 'NOMINATIVE'
ASTHEMARKED
CASE.We began with the thesis that morpho-
logical marking can be used for either A or O function. S is normally unmarked,
since there is no other core NP in an intransitive sentence from which it must be
distinguished; it then falls together with the unmarked transitive function. But
there is a special semantic connection between A and S: these are the NP's whose
referents can control and/or initiate an event, if anything can. A and S are joined
together, at the deep syntactic level, in the universalcategory of'subject' (discussed
in some detail in ?5).
It is thus quite natural that the positive marking on A in transitive sentences
should be extended to cover S function. This type of system differs from those of
??2.31-2.32 in that it is the marked transitive case, not the unmarked case, which
is used for S in an intransitive sentence. One case does cover A and S functions, and
another case 0 function; so this kind of system could be termed 'nominative/
accusative'. It is, however, radically different from the kind of 'nominative/
accusative' system discussed under ?2.32; there the 0 function was marked by a
positive accusative case, but here it is the A function in a transitive sentence that is
marked. As far as transitive sentences go, the present type is-on semantic grounds
-most similar to the ergative case system, dealt with under ?2.31. But since the
'special A-marking' is extended-so that it is in fact a 'special subject marking'-
there are distributional (although scarcely semantic) similarities to the 'unmarked
nominative/marked accusative' type of morphology.
Midway between the type of ?2.31, marked case for A function, and the present
type, marked case for A and S functions, there are some languages where a marked
case is used for all A NP's in transitive sentences and for SOMES NP's-just those
where the S NP does have agentive force, e.g. Bats (Comrie 1973:241, Catford
1975:20-21); see the discussion in ?3.1, below. In the latter circumstance it is
usually said that the 'ergative case' CAN also be used to mark intransitive subject.
crucial difference between transitive and intransitive sentence-types, and conflates A and S
functions. I attempt to demonstrate throughout this paper, however, that A and S must be
clearly distinguished if any progress is to be made in an investigation of ergativity and in a
generaltheory of case-marking.
21
The situation in Lardilis slightly more complex, in a ratherinterestingway; details are in
?3.23.
22
For discussion of the syntactic circumstancesin which an object NP can be marked with
'nominative' in North Russian dialects, see Timberlake 1974.
ERGATIVITY 77
This terminology could appropriately be taken further; the name 'extended

ergative' (rather than 'marked nominative') could be used when we encounter a
marked case employed for A and for ALLinstances of S function. Using labels of
this type would ensure that 'ergative' and 'accusative' are always used to name
marked case choices, and 'absolutive' and 'nominative' unmarked choices.
Languages showing the 'marked nominative' (= 'extended ergative') include
some members of the Cushitic family. Here it is the unmarked 'accusative' case
that is used in citation forms-and, interestingly, as the complement of the verb
'to be'.23 A similar situation is found in the Yuman family of California. Proto-
Yuman used the stem form for 'object', but added a suffix -c for (transitive and
intransitive) subject; this system is followed in most modern Yuman languages.
Thus in Mojave 'nouns are usually cited in their unmarked form, often with a -d
vowel added, but sometimes they are cited with -c(N + c may constitute an elliptical
"It's a ..." sentence, which could explain this)' (Pamela Munro, p.c.) Wappo
appears to be another language of this type; Li, Thompson & Sawyer 1977 suggest
that an original A marker has been extended to cover S function in main clauses,
but not yet in relative clauses or equational sentences. If this diachronic hypothesis
is correct, 'extended ergative' would be a most appropriate label for the marked
'subject' inflection.
Having the unmarked nominal case only for O function (and as the normal
citation form) appears to be a well-established grammatical characteristic of the
Yuman and Cushitic families. In other languages, phonological changes may lead
to an accusative form being phonologically unmarked with respect to nominative;
but this may never correlate with morphological marking, and is then likely to be a
transient stage of development. An example of this is found in some of the older
Germanic languages-where, e.g., the IE masc. nom. sg. *-os still appears as -s or
-r, but acc. *-an has completely disappeared (the -n was lost first, and then the -a;
-a is still found in Runic Norse). There is no evidence that accusative ever
functioned as the unmarked case (that it was ever, say, employed in citation).
Further changes have led to nominative and accusative falling together (Meillet
[1917] 1970:91 if.)
It is interesting to compare the 'marked nominative' systems of Yuman and
Cushitic, and the regular 'unmarked nominative' system of Germanic (with
unusual phonological realization), with case-marking in Maidu, a Californian
Penutian language. Here 'subject case' involves the addition of-m to the 'object'
form. Shipley (1964:29-30) reports that older speakers employ the 'subject' form
for naming (i.e. citation), but that younger speakers use the 'object' form; each
speaker is quite consistent in his citation forms. An immediate suggestion here is
that Maidu was originally of the Germanic type, with nominative as the functionally
unmarked case (we would then wonder whether, in an earlier stage of the language,
accusative would have had some non-zero realization), but that the youngest
23
E.g. Oromo, Dasenech, and Kambata (Bender 1976:182, 205, 253). Other Cushitic lan-
guages have a marked nominative only in SOME noun classes, or use the zero accusative only
for INDEFINITE objects. A furthergroup appear to have non-zero inflectionfor both nominative
and accusative, with accusative being used in citation (informationfrom Robert Hetzron and
Dick Hudson).
generation of speakers has extended formal markedness to apply also at the

functional level, making Maidu more like Cushitic and Yuman in this respect. (We
would of course need to know a good deal more about the use of cases in Maidu to
be sure of this shift in markedness. Note also that the 'younger speakers' referred
to by Shipley were the remnant last generation, speaking a language on the point
of extinction. We cannot say that this provides a natural example of language
change, under normal conditions of use.)
In summary, we have distinguished three kinds of 'markedness' among case
inflections covering the three core syntactic functions A, O and S. Basically, either
of the transitive functions can be marked. If O is marked (by 'accusative case'),
then the unmarked 'nominative case' is used for S and A functions, and is used in
citation etc. If A is marked (by 'ergative case'), then both O and S may be shown
by the unmarked 'absolutive case', which will again be used for citation. But the
marking on A can also be extended to cover S, with the unmarked case being
confined to O function and most instances of citation. Strictly speaking, none of
the terms 'nominative', 'accusative', 'absolutive', or 'ergative' are really appro-
priate for this third possibility; the alternatives seem to be to speak of 'extended
ergative' and 'restricted absolutive' (maintaining the convention that ergative is
always marked with respect to absolutive), or of 'marked nominative' and 'un-
marked accusative' (this respects the convention that nominative always covers
S and A functions, while accusative is confined to 0 function).
The extension of 'marked A case' to S can be explained in terms of the universal
syntactic-semantic identification of A and S as 'subject' (see ?5). There is a more
slender semantic link between 0 and S, so that the fourth logical possibility-
'marked 0 case' being extended also to cover S-appears not to occur.24
2.34. In ?2.1 I noted that languages can be characterized as 'ergative' or
'accusative' at the morphological level, in terms of case inflection on NP's, or in
terms of bound pronominal-type affixes (usually attached to the verb) which
cross-reference certain features of core NP's. The discussion in ?2.2, on the
semantic basis of morphological marking, applied equally to the two possibilities.
However, the present discussion of 'markedness' applies only to case inflections.
Some form of a noun-either just the root, or else a particular non-zero inflec-
tional form-must be used in citation. In many languages, one NP (in a particular
case) must be present in each sentence. There is usually a restriction (in terms of
surface syntactic function and/or case inflection) on the NP's that can act as
'pivots' in subordination or coordination. All these considerations, and others
besides, will provide criteria for recognizing markedness in case systems. In
addition, one case often has zero realization: in most instances, this correlates
24
Australia offers examples where the accusative case is extended to cover S as well as 0
function for SOME TYPES of nominal constituent only. Thus in both Warluwaraand the Western
Desert language, the pan-Australianaccusativesuffix -nva marks 0 and S functions on proper
names (Dixon 1970:95; furtherexamplesin Dixon, MS b, ch. 11.) Note that in Proto-Australian
an accusative inflectionoccurred only with pronouns, demonstratives,and propernouns (as in
most modern Australianlanguages);common nouns used the bare stem-absolutive case, with
zero inflection-for S and 0 functions.
ERGATIVITY 79
with morphological markedness (the Germanic example quoted in ?2.33 is one of

the few exceptions).
The only way in which we might talk about 'markedness' with respect to
pronominal affixes is in terms of WHICHamong A, S, and O is cross-referenced in
this way, in a language with only partial cross-referencingof core NP's. If only A
and S are cross-referenced, this could be taken as evidence for an (unmarked)
nominative / (marked) accusative system.25 Note here that the unmarked term is
the one that has some positive realization; this is the inverse of the situation with
cases, where an unmarked case is the most likely candidate for zero realization.
There may be one or more forms in a cross-referencingparadigm that have zero
realization (usually including 3sg.), though this is evidence about a quite different
type of markedness, viz. markedness within person/number systems. But if there
are more zero forms in the A than in the O prefix paradigm (as Alan Rumsey, p.c.,
reports for the Australian language Ungarinjin), this could conceivably be taken as
evidence for A being relatively 'unmarked' with respect to 0.
Many languages have non-zero affixes cross-referencing S, A, and 0; even for
those that do not, the criteria for 'markedness' outlined in the last two paragraphs
are slim, and need corroboration from other types of grammatical criteria. This
contrasts with the several strong criteria available for deciding on markedness
within a system of case inflections. This discrepancy in the recognition of'marked-
ness' constitutes an important difference between cases and cross-referencing
affixes.
A further difference between the two sorts of morphological systems is the fact
that case inflections can ALWAYSmark which NP is A, and which 0, without
reference to any other grammatical information. But verbal affixes provide only
limited semantic information about core NP's-usually person and number, and
sometimes gender or noun class. If both A and 0 make the same choices from
these cross-referencing systems-say, if they are both 3sg. masculine-then some
other way must be found to distinguish A NP from 0 NP. (Word order is the usual
fall-back, although nominal case inflections can be used, as with the 'only-if-
needed' ergative in Murinypata, mentioned in ?2.31.)
These are fundamental differences. If we describe a language as 'ergative' in
terms of case inflection, or in terms of the paradigm of cross-referencingaffixes, we
are describing distinct types of grammaticalphenomena. We cannot expect that the
types of conditioning for 'split ergative systems' will necessarily apply equally to
the two phenomena.
TYPES OF SPLIT CASE SYSTEMS
3. We can now return to the discussion begun in ?2.2, concerning the kinds of
factor that condition split case systems-namely, if some part of the morphology
of a language shows absolutive/ergative patterning, and some other part a nomi-
native/accusative paradigm, what is it that determines this split? There appear to
be three basic types of factor: the semantic nature of the main verb, the semantic
25 Latin and English are essentiallyof this type. The Caucasianlanguage Avar
representsthe
other type of system-where, roughly, S and 0 are cross-referencedin the verb, but not A
(Anderson 1976:4,Cerny 1971).
nature of the core NP's, and the tense or aspect of the clause. I consider these in
turn.
BYSEMANTIC
3.1. SPLITCONDITIONED NATUREOFVERB.For nearly every transitive
verb, the state of affairs to which it refers CANbe controlled by an agent; that core
NP which refers to the controlling participant (if there is a controller) is said to
be in A function, and the other core NP is in O function. (Note that the referent
of an A NP does not ALWAYS control a potentially controllable activity. People can
look for something, and as a result see it; but they can see something accidentally,
without meaning to. A non-controlled token of a transitive verb can be indicated in
some special way; in ?2.31 we saw how YidinYemploys an affix -:dyi- to mark this
phenomenon.)
Intransitive verbs, referringto events that involve a single core participant, span
a wide semantic range; we will say that the single core NP of an intransitive
construction is in S function, regardless of the semantic type of the verb. With some
verbs, the referent of the S NP will almost always be controlling 'agent' (e.g. 'run',
'jump'); with others, it will normally not be (e.g. 'yawn', 'be hungry'); with a
further set of intransitive verbs, the referent of the S NP may sometimes have a
measure of control over the activity and sometimes not (e.g. 'sleep', 'cough'). The
semantic nature of the S NP is relevant here: something inanimate could never be
'controller'. Thus It fell must be non-agentive; but with He fell, it is conceivable
that the participantfell on purpose!
Languages show a tendency to ignore these semantic niceties and to generalize
ONEgrammatical form to mark S NP's with all types of intransitive verbs. This can
be the same form used to mark 0 function in a transitive sentence ('absolutive'
case), or the same form as for A ('nominative' case).
To avoid ambiguity in a transitive sentence, there must usually be a strict gram-
matical convention that A NP's be marked in one way and 0 NP's in some other
way. Once these two types of morphological marking are established, the semantic
nature of an intransitive sentence can perspicuously be shown by marking S
according to the A convention whenever, for that referential token of the verb, it
ACTUALLY ISthe controller, and marking S in the same way as 0 when it exercises
NOcontrol over the activity-i.e. by using one case inflection for A and Sa, and
another for 0 and SO(cf. ?2.2).
This scheme, which we can call 'fluid S-marking', is found in just a few languages.
It is said that in Bats, a Northeast Caucasian language, some intransitiveverbs (e.g.
'go', 'play', 'look', 'speak') MUSThave a 1st or 2nd person pronominal S NP in
ergative case, while others can mark S function by EITHER absolutive case (implying
that the action was involuntary) ORby ergative case (implying that the referent of
the S NP controlled the activity, or that it was his fault). Verbs in the latter set
include 'fall', 'get drunk', 'fear', and 'lie down'.26 The use of ergative or absolutive
26
Nineteenth-centuryBats behaved in the way described here (Comrie 1973:241, Catford
1975:20-21). In the modern language, some intransitiveverbs demand ergative inflection on
their S NP's, but others take absolutive: in effect, a 'fluid S-marking'system has evolved into a
'split S-marking' language. The meaning distinctions that could previously be shown by
alternation of ergative and absolutive inflectionsare now dealt with by lexical resources.(I am
grateful to C. E. Bazell for pointing this out; see Desheriev 1953.)
ERGATIVITY 81
on an S NP appears to be semantically determined: instead of having to recognize

classes among intransitive verbs, we can simply say that ergative case is used for a
controlling S, while absolutive case is used for S NP's in other instances. The
semantic nature of intransitive verbs dictates that for some the S NP is ALWAYS
agentive, for some it CAN BE, and for others it NEVERis. (The ergative case in Bats is
seen to be clearly 'marked'; cf. ?2.31.)
Another example of this type is Eastern Pomo, a Hokan language of Northern
California. There are some intransitive verbs (e.g. 'fall', 'sneeze') whose S NP
always receives case inflection typical of an O NP in a transitive sentence: these
describe an event that CANNOTbe controlled. At the opposite end is a group of
intransitive verbs (e.g. 'sit', 'go') where the patient ALWAYS exercises control; the
S NP receives marking that applies to the A NP in a transitive sentence (at least for
pronouns, kin terms, and proper nouns). Between these two extremes is a further
group of intransitive verbs whose S NP's can be inflected like an A or an O NP,
depending on whether the participant referredto does or does not exercise control,
e.g. 'slide/slip' (McLendon 1978).
In Bats and Eastern Pomo, fluid S-marking is expressed by case inflections. It
appears that the Siouan language Crow (Kaschube 1967; Avery Andrews, p.c.)
has the same sort of phenomenon realized in terms of bound pronominal affixes.
An intransitiveverb in Crow bears a prefix,cross-referencingS NP, that corresponds
either to A or to 0 transitive prefixes, according to whether or not the 'subject' has
volitional control over the activity. By their semantic nature, some intransitive
verbs occur only with A prefixes ('run' is always voluntary) and others only with 0
prefixes ('fall over' is never voluntary)-whereas a middle group (e.g. 'go') can
take either type of prefix, depending on whether volition is involved in a particular
token of use.27 A similar situation appears to prevail in Choctaw (although there
may be less fluidity here): most verbs are restricted either to A or 0 prefixes, but
there are a few that may take either type of affix (Byington 1870 gives 'sleep' as an
example).28
27 There are some complications in Crow. There is a class of verbs (e.g. 'to not know') that
are semantically non-agentive but take agentive prefixes, except for 1st person plural (this
pattern of prefixationis also shown by a class of prepositions, with respect to their objects).
The verbs 'to tattle' and 'to tell lies' take non-agentive (0) prefixes although they behave
syntacticallylike agentive verbs (intransitiveverbs taking A prefixes).
28 Jeffrey Heath reports that, in his recent field research on Choctaw, he 'encountered no
systematic use of variation in case-marking with individual roots for marking nuances, as
Byington suggested'. This COULD be taken to indicate that Choctaw has changed toward a
'more grammatical' system (as described in the next paragraph)during the last century; but
there are too many other variablesinvolved to make this more than a speculation. The gram-
matical system of Choctaw is in fact slightly more complex than suggested here, with a few
intransitiveverbs taking neither A nor 0 but 'dative' prefixes,to cross-referencethe S NP (cf.
Byington 1870and Heath 1977).
It is interesting to compare Bats and Crow with Cupeno (as describedin Hill 1969). In all
threelanguages,a single lexical item can be used to referto an involuntaryor voluntaryinstance
of some activity, through appropriategrammaticalmarking. But whereas Bats uses different
case inflections and Crow differentpronominalprefixes,in Cupeno and related languageseach
verbal word must be specifiedas to whetherit is describinga 'natural event', (zero markingon
verb), something volitional (suffix -ine), or something involuntary (suffix -yaxe). The Cupeno
I have followed other writersin using 'ergative' and 'absolutive' case labels in the
last few paragraphs. But on distributional grounds, these are by no means the
uniquely appropriatedesignations. For some verbs, S is marked in the same way as
A (a 'nominative/accusative' system); for others, S is marked in the same way as
O (an 'absolutive/ergative' distribution); and for a few verbs, both alternativesare
possible. The use of 'ergative' would have to be justified in terms of markedness
(cf. ?2.3); it may well be that this can be done for Bats, or for any other case-
marking language of this type. But for morphological systems which work in terms
of cross-referencingverbal affixes, this solution is not available. It will be simplest
to specify-for Crow, Choctaw, Bats, or Eastern Pomo-whether an S NP receives
'A-marking' (i.e. that morphological markingwhich applies for A NP's in transitive
sentences) or '0O-marking'.
Grammatical untidiness of the Bats/Crow sort-semantically revealing though
it is-is, in fact, not tolerated in more than a handful of languages. What is more
common is for intransitive verbs to be strictly classified into two sets: one, whose
actions are always (or almost always) 'controlled', marks an S NP like an A NP;
the other, whose actions are seldom or never controlled, marks it like an O NP. In
languages of this 'split S-marking' type, there are no intransitive verbs that can
take EITHER A OR 0 marking.
The Siouan language Mandan is a clear example of this. Kennard 1936 distin-
guishes verbs which indicate an activity from those which indicate a state or con-
dition. The first class (of 'active verbs') can be transitive, occurring with subjective
and objective pronominal suffixes (e.g. 'ignore', 'tell', 'give', 'see', 'name'), or
intransitive, occurring just with subjective suffixes (e.g. 'break camp', 'enter',
'arrive', 'think it over', 'go'). The second class (of 'neutral verbs') takes only the
objective pronominal suffixes; they include 'fall', 'be lost', 'lose balance', and
verbs covering concepts that would be included in an adjectival class for other
languages, such as 'be alive', 'be brave', 'be strong'.
Guarani, from Paraguay, provides a further example of the split S-marking
type.29 Gregores & Suarez 1967 distinguish three classes of verb. 'Transitive verbs'
distinctionscross transitivityboundaries(cf. volitional -wecdxnen 'throw down' with voluntary

-wacdxneyex 'drop accidentally') in a way that the Bats and Crow distinctions cannot, by the
nature of their grammaticalmarking.
29
A more complex (and also more interesting)example of split S-markingoccurs in Tunica.
Haas 1940 first distinguishes'active' from 'static' verbs. Active verbsall take a prefixindicating
person/numberof the subject(A or S), and also the mood of the clause; they can be subdivided
into transitive verbs, which also take an object prefix, and intransitiveverbs, with no object
affix. Static verbs (a small class-only some thirty membersare known) take a differentprefix,
cross-referencingthe S NP.
Now the static prefix is identical to the prefix on nouns that marks inalienable possession
(e.g. 'my father'). The object prefix on a transitiveverb is identical to the alienable possessive
prefix on nouns (e.g. 'my hog'). Furthermore,alienable prefixes appear to be derived from
inalienableprefixesby the addition of -(h)k. Thus the static S prefixdoes not coincide with the
O prefix; but it has the same formal relation to it as inalienable nominal prefixes have to
alienable ones. This suggests tempting lines of philosophical speculation, e.g. that an S NP is
more closely attached to ('inalienably possessed by') an intransitiveverb than an 0 NP is to a
transitiveverb. It would be inappropriateto pursue this further in the present context.
ERGATIVITY 83
(e.g. 'give', 'steal', 'know', 'order', 'suspect', 'like') take prefixes from both
subject and object paradigms. 'Intransitive verbs' ('go', 'remain', 'continue',
'follow', 'fall') take subject prefixes. Both classes can occur in imperative inflection
-unlike the third class, which Gregores & Suarez call 'quality verbs'; these take
prefixes that are almost identical to object prefixes on transitive verbs. Most
quality verbs would correspond to adjectives in other languages, although the class
does contain 'remember', 'forget', 'tell a lie', and 'weep'.
There are languages of the split S-marking variety where the two intransitive
subclasses do not have as good a semantic basis as Mandan and Guarani. Thus in
Hidatsa, another Siouan language (Robinett 1955), the intransitive verbs with A-
marking include volitional items like 'talk', 'follow', 'run', 'bathe', and 'sing'-
but also 'die', 'forget', 'have hiccups', that are surely not subject to control. And
the set with O-marking includes 'stand up', 'roll over', and 'dress up', in addition
to such non-volitional verbs as 'yawn', 'err', 'cry', 'fall down', and 'menstruate'.
It may be that what began as a semantic distinction has, since it was grammatical-
ized into two non-overlapping classes, come under other pressures,and has gradually
moved away from its original semantic basis.
The most frequently quoted example of a split S-markinglanguage is undoubtedly
Dakota, another member of the Siouan family (Boas & Deloria 1939, Van Valin
1977; see also Sapir 1917, Fillmore 1968:54). There are many other languages of
this type among the Caddoan, Siouan, and Iroquoian families, e.g. Joway-Oto
(Whitman 1947) and Onondaga (Chafe 1970).
The size of the class of intransitive verbs that must take '0 affixes' (so-called
'stative verbs') varies a good deal. In Guarani it appears to be an open class (some
hundreds of members are listed by Gregores & Suarez), including all concepts that
would be rendered through a class 'adjective' in other languages. But the class of
stative verb roots in Seneca (an Iroquoian language) has only three dozen or so
members; for Arikara, a Caddoan language, barely two dozen stative verb roots
are reported-here most adjectival concepts are dealt with by a quite different
grammatical class.30
Just because a language has split S-marking in some part of its morphology does
not, of course, imply that it cannot identify a macro-class of intransitive verbs with
respect to which the unique intransitive core function S is recognized. My basic
thesis is that all languages work in terms of universal syntactic-semanticfunctions
A, 0 and S. Although surface realizations may group A with S, or 0 with S, or
split S into two sub-types (Sa identified with A, and SOwith 0), the underlying
pattern of A, 0, and S still exists; and this is always needed to explain and relate
the morphological, syntactic and semantic possibilities for a language.
Wichita, a Caddoan language, is a classic example that shows some 'ergative'
and some 'accusative' characteristics in surface structure, in addition to split S-
marking. Rood 1971 describes how transitive verbs take 'subjective' and 'objective'
30
Details of sizes of stative verb classes in Iroquoian and Caddoan languages were supplied
by Francesca Merlan. It seems that languages with a closed class of stative verbs generally have
an open class of active intransitive verbs. The reverse, however, may not hold; David Rood
reports that Lakhota (closely related to Dakota) has hundreds of members in both its active
intransitive and stative intransitive verb classes.
prefixes, cross-referencingA and O NP's respectively (but note that there are non-
zero prefixes only for 1st and 2nd persons). One set of intransitive verbs (e.g. 'go')
must take subjective prefixes; another set must take objective prefixes (examples
are 'be cold' and 'be hungry'). We could say that S is subcategorized as Sa or SO,,
depending on the subclass of intransitive verb with which it occurs. But Rood
quotes two grammaticalprocesses that group together O and S (and take no account
at all of the distinction between Sa and SO):any O or S NP (but not A NP) can
optionally be incorporated into a verb word, and a single set of verbal affixes
indicates plural O or S (another set is used for plural A). Finally, S and A behave
the same way in word ordering: an O NP (if there is one) must precede the verb,
then the subject (A or S NP) can either precede or follow this complex.31
Choctaw (Heath 1977) provides another example of a language that has different
kinds of identification in different parts of the grammar. I have already mentioned
that an S NP can be cross-referencedonto the verb by A, or by O, or in one or two
cases by dative prefixes (and a few verbs may take either an A or O prefix depending
on whether they are understood as volitional in that instance). In contrast, the core
NP's themselves show straight nominative/accusative case-marking, -n being used
for O and -t - -s for A and for S (regardless of how S is cross-referenced in the
verb). And rules of coordination and subordination refer directly to 'subject' NP's
(i.e. A and S) regardless of the prefix system used to cross-referenceS.
This paper began by pointing out that characterizations of 'ergativity' could
apply at the morphological or at various syntactic levels, and that a language which
took one value at a certain level could show a quite different value at some other
level. It was then indicated that no 'pure ergative' language is known (at any level).
In morphology, ergative characteristicsoccur within a 'split case system', and there
can be a number of diverse semantic factors conditioning a split. Wichita and Choc-
taw show that things are not even THAT tidy: differentparts of the morphology may
require quite different characterizations. Basically, there are a number of different
ways in which S can be related to A, or to 0-or partly to each, according to
syntactic and semantic factors; each language can exploit ONEor MORETHAN ONE
of these ways.
This section has discussed (i) fluid S-marking, where the S NP of any intransitive
verb can potentially be marked either on the A or on the 0 pattern, and (ii) 'split
S-marking', where intransitive verbs fall into two mutually exclusive subclasses,
one using A-marking and the other 0-marking for its S NP.32 For (i), the marking
31
Rood points out (p.c.) that thereare distinct Saand SOprefixesonly for 1st and 2nd persons,
but the processes of incorporationand pluralizationwhich group together S and 0 apply only
to 3rd person forms. He suggeststhat this could be evidencefor a Silverstein-typesplit (cf. ?3.2)
-with, roughly, Sa being grouped with A for 1st and 2nd persons, but included with SOand 0
in an 'absolutive' groupingfor other nominalconstituents.
Rood also points out (modifyinghis statementin Rood 1971:101) that word orderin Wichita
shows considerablefluidity: although AOV and OVA are the most frequent,and represent 'the
firstinterpretationof sentenceswith two NP's, neitherincorporated',OAV and AVO have been
encountered.
32
Americaniststypicallycharacterizelanguageswith split S-markingas showing a distinction
between 'active' and 'stative' verbs. Klimov 1973 (see Comrie 1976b)simply calls these 'active'
languages.
ERGATIVITY 85
is entirely conditioned by the referential token of that instance of the intransitive

verb; for (ii), the grammatical subclassification is based, to a greater or lesser
extent, on the semantic type of the verbs.
In verb-conditioned splits of these types, there is consistent treatment of A and O
NP's within a transitive sentence. The split focuses on how the S NP is treated, in
terms of the transitive marking possibilities.
3.2. SPLITCONDITIONED
BY SEMANTIC
NATUREOFNP's. We can enquire why any
language should invest in the complications of a split case-marking system, rather
than just employing a straightforwardnominative/accusative or absolutive/ergative
marking. The answer lies in the fact that a split case system enables the language
to reflect the semantic nuances of different types of event-and, in some instances,
to reflect both the universal category of subject {S,A} and a language-particular
pivot that equates S with O (see ??5-8). In the last sub-section, I showed how a split
conditioned by the semantic nature of (intransitive) verbs can perspicuously
indicate which members of this class can be 'controlled' (like transitive verbs).
We now consider the semantic nature of NP's, continuing to use the label
'participant' for the referent of any NP. Some kinds of participant will typically
function as controller or agent of an action referred to by a transitive verb; others
are seldom likely to function as agent; and a further set are never likely to function
in this way.
3.21. Most discourse, in any language, is oriented to the people involved in the
speech act, and pre-eminently to the speaker. Every human language has a class of
pronouns, with 'shifting reference'. ('Shifters' are referentially so different from
other parts of speech that it is rather striking that ALLlanguages should show this
class.) In the speaker's view of the world, as it impinges on him and as he describes
it in his language, he will be the quintessential agent. Radiating out from this
egocentric focus, the next most likely agent will surely be the addressee; then
specific '3rd person' humans referred to by demonstratives; then known humans
referred to by proper names; then other humans (described just through common
nouns); then higher animals such as dogs, and on down the scale of animacy until
inanimate participants are reached. The last type could never be expected to
function as controlling 'agents' (even though they could, on occasion, realize A
function in a transitive sentence).
We can represent the 'potentiality of agency' scale diagrammatically as in
Figure I.33
Demonstratives
Human Animate Inanimate ...
1st person 2nd person 3rd person Proper t
pronoun pronoun pronouns nouns Common nouns
-*
likelihood of functioning as transitiveagent
FIGURE 1
33 Whether 1st person shouldprecede 2nd person on the hierarchy,or vice versa, is a con-
troversial question; there is evidence for either ordering. Perhaps we should just recognize
three types of NP constituent in terms of which case splits are to be explained: pronominal
shifters (1st and 2nd person forms), other pronominal-type forms (3rd person pronouns,
deictics, propernouns), and common nouns.
There are, of course, different semantic types within the class of transitive verbs;
and the 'agent propensity' of differentkinds of NP does vary a little with the verbal
type. Verbs like 'eat' or 'bite' can have virtually any human or animate agent;
'notice' and 'hear' are limited to higher animals and humans; 'speak', 'calculate',
or 'lend' are surely restricted to human agents. But the hierarchy given here
roughly indicates the over-all 'agency potential' of any given NP.
Which type of participant is most likely to function as patient (realizing the O
slot in a transitive sentence) varies much more with the semantic nature of the
verb. One can 'see' the speaker or any other human just as well as anything
inanimate; and human participants are MORELIKELYto be patients of 'hear' than
most items to their right in the hierarchy. TheO NP of 'pick up' is LIKELYto be
something from the far right of the hierarchy, but the patient for 'spear' or 'shoot'
will be centered on the animate set of participants. When the 'potentiality of being
patient' is averaged out, as it were, over the whole class of transitiveverbs, the same
hierarchy will apply-although not so homogeneously as for the 'agent' slot. The
further to the left a participant is, the more likely it is to function as agent; the
further to the right, the more likely it is to be the patient.
3.22. We have been viewing the hierarchy as an explanation of what is likely to

O
function as an A or NP (essentially treating these as two distinct phenomena).
It is perhaps more revealing to look at things the other way around-to take each
class of participants on the hierarchy, and to ask whether it is MORE LIKELY to
occur in A than in 0 function. When we do this, the variations according to
semantic type of verb are reduced: it is certainly the case that, as one moves along
the hierarchy from the left, the expectation of occurring in an A slot more often
than in 0 function decreases. What I am saying is that the speaker will think in
terms of his doing things to other people to a much greater extent than of having
things done to him ;34 he will think of people doing something to animals or things
much more often than the reverse; and so on.
It is plainly most natural and economical to 'mark' a participant when it is in an
unaccustomed role. That is, we could expect that a case-marking language might
provide morphological marking of an NP from the right-hand side of the hierarchy
when it is in A function, and of an NP from the leftmost end when in 0 function
(as an alternative to providing ergative marking for ALL A NP's, of whatever
semantic type, or accusative marking for ALL 0 NP's).
A number of languages have split case-markingsystems exactly on this principle:
an 'ergative' case is used with NP's from the right-hand end, up to some point in
the middle of the hierarchy, and an 'accusative' case from that point on, over to
34
Discussion of this point should reallydeal in turn with each semantictype among transitive
verbs; differentconsiderations apply for each type. Thus most people would like to think of
themselvesas hitting more often than being hit. Witha verb like 'see', quite differentarguments
are applicable. The speaker will be seen as often as he sees another person, but the fact of his
seeing someone is more likely to be significantto him than the fact of his being seen. (Of course
instances of trying to avoid being seen do occur; but they are normallya minorityamong every-
day occurrencesof 'seeing'.) It is what is importantand significantto the speaker(ratherthan
'what happens') that largely determinesthe shape of the hierarchy.
ERGATIVITY 87
the extreme left of the hierarchy. The case-marking of Dyirbal, mentioned in ?1.1,
provides a straightforward example, as shown in Table 2.
A -0 -yjgu -ygu -ggu
S -0 -0 -0 -0
0 -na -0 -0 -0
1st & 2nd 3rd proper common
person person names nouns
pronouns pronouns
TABLE2
Here we have accusative -na vs. the unmarked nominative -0 for 1st and 2nd
person pronouns, but ergative -yiguopposed to the unmarked absolutive -0 for the
rightmost three columns.35
3.23. We can think of O-marking, extending in from the left, and A-marking,
coming in from the right, as essentially independent parameters.They can overlap,
so that something in the middle portion of the hierarchy will have different forms
for all three of the core functions S, A, and 0.36 Consider the Cashinawa language
of Peru, shown in Table 3.37
A -0 habui nasalization
S -0 habu -0
0 -a haa -0
1st and 2nd 3rdperson propernamesand
person pronoun common nouns
pronouns
TABLE3
An NP with a noun as head receives the ergative case-marking, realized as
nasalization of the last vowel in the final word of the NP. The 1st and 2nd person
pronouns show an accusative suffix -a. But the 3rd person pronoun has both types
of marking, and shows three different case forms (note that the root is habu for S
function, with nasalization added in A function as it is for nouns; the pronominal
accusative -a is added to a shorter root ha-).
35 The situation is in fact slightly more complex than this. The interrogative/indefiniteform
wan5a 'who, someone' has distinct forms for S, A, and 0 functions; and propernames, as well
as some nouns with human reference, can optionally take -n5a (cognate with pronominal
accusative -na) in 0 function only. (This suggests that they should be placed to the left of' 3rd
person pronouns', as least as far as Dyirbal is concerned.)
A language that does have mutuallyexclusive 'ergative' and 'accusative' marking, with no
overlap of any sort in the middle of the hierarchy,is Gugu-Yalanji; see R. Hershberger1964,
H. Hershberger1964.
36
There are examples of ergative case covering the whole length of the hierarchy, with
accusative being more limited in application. In Waga-Waga,from south Queensland,all NP
constituentstake the ergativeinflection;accusativemarkingapplies to pronouns, propernouns,
all common nouns with human reference, and just a few common nouns with non-human
reference(Wurm1976). Here thereare separateforms for S, A, and 0 at the left of the hierarchy,
but an absolutive/ergative system at the right (and no simple nominative/accusativecase-
markingin any part of the hierarchy).
37 Analysis of Cashinawais inferredfrom data provided in Merrifieldet al. (1965:140-43);
I am grateful to Robert E. Cromackfor supplyingadditional data.
There are many other languages where A and O markings overlap for some part
of the middle of the hierarchy, rather than ergative marking stopping at the place
where accusative begins. (Note, though, that the A and O markings, extending in
from opposite ends of the hierarchy, should at least meet; if they do not, other
means would have to be employed to distinguish A and O for the class of NP's that
show neither accusative nor ergative affixes.) In YidinY(cf. ?2.31), e.g., pronouns
have a nominative (-0)/accusative (-no -:n.) paradigm, while at the other end of
the scale, all nouns show absolutive ergative marking (-0 vs. -ygu -du etc.) But in
the middle region, there are separate forms for A, S, and O functions, for deictics
that have human reference, and for the human interrogative/indefiniteform 'who,
someone'. Deictics with inanimate reference can use the unmarked S form for O
function, OR they can use a special O form (in accusative -:ny); and the inanimate
interrogative/indefinite 'what, something' has one form for S and O functions,
exactly like nouns. Note that the ergative case in YidinYmarks any common or
proper noun, or deictic or interrogative/indefinite, when it is in A function (that
transitive function whose referent COULDbe controller/initiator of the action). But
in addition, as described in ?2.31, the verbal affix -:dvi- is brought in when the A
NP is inanimate (and thus Is INCAPABLEOF BEING the controller) or when the A NP is
human but IS NOT IN THIS INSTANCEcontrolling the activity.
This hierarchical explanation for split determined by semantic content of NP's
originated entirely with Michael Silverstein; this whole section is firmly based on
Silverstein's research and observations. Many further examples of split systems of
this type are given in Silverstein 1976, together with detailed discussion of the
principles of 'markedness' underlying the hierarchy. Note also that Silverstein's
hierarchy explains case splits outside the field of 'ergative languages'. Thus, in
most IE languages, pronouns and nouns from masculine and feminine declensions
have distinct nominative and accusative forms; but neuter nouns have a single
form for S, A, and 0 functions. We can say that accusative marking extends only
so far in from the left of the hierarchy, but there is no ergative marking on the
right. (One presumes that a transitive sentence whose A and 0 NP's both involved
neuter nouns would have to resort to something like word order to decide which
was A and which 0; such sentences are, of course, quite uncommon.) See also
Lyons' insightful discussion of this point (350-71).
A further instance of the nominal hierarchy explaining an oddity in the case
system of a basically accusative language concerns Lardil, spoken on Mornington
Island in the Gulf of Carpentaria, Australia. We saw in ?2.32 that for Lardil the
accusative case is marked on 0 NP's except in imperative constructions. There is
in fact an exception to the exception: accusative MUST be marked on a 1st person
pronoun, even in imperative sentences. Since 1st person is on the extreme left of
the hierarchy, it is the strongest candidate for accusative marking: it receives this
marking even when accusative case is suspended for all other pronouns and nouns,
in imperative constructions. (Data on Lardil are from Klokeid 1976:197 and
Nash, MS.)
3.24. There is one further complication in systems split according to the semantic
content of NP's: non-singular pronouns sometimes appear to be further to the left
ERGATIVITY 89
in the hierarchy than singular pronouns. This can be illustrated from Arabana,
spoken in South Australia,38as shown in Table 4.
A -0 adu -ru-ri -ru -ri
S -0 anda -0 -0
0 -na
n
ana -na -0
n n
non-sg. sg. pronouns proper common
pronouns (exemplified nouns nouns
here by Isg.)
TABLE4
There is absolutive/ergative inflection of common nouns, and a nominative/
accusative paradigm for non-singular pronouns (of all three persons), but three
distinct forms for proper nouns and for singular pronouns. (Singular pronouns are
not readily analysable; the actual Isg. forms are cited in the table.)
Silverstein mentions other examples of number-conditioned split, and attempts
an explanation in terms of markedness of features. From a semantic angle, it surely
cannot be argued that non-singular pronouns are more likely to be in A function
(and less likely to be in O function) than singular pronouns. But we can note that
'singular' is the universal unmarked term in number systems-and that, typically,
several cases that are distinguished in the singular will fall together in the plural
(with still more neutralization in the dual, for languages with a three-term number
system). If all pronouns are not to show the same case distinctions, then it is the
non-singulars that are likely to be deficient; for ergative case to extend in from the
right-hand side of the hierarchy as far as singular pronouns, but not to include
non-singulars, is thus perfectly natural. But strictly speaking, number distinctions
should not be included in the semantic hierarchy;this is a differentkind of phenom-
enon, involving a different type of 'marking', and providing a different sort of
explanation.39
3.25. In ?2.1, I described the two major mechanisms of morphological marking:
case inflections (or particles) on NP's, and cross-referencing pronouns (typically
attached to the verb). Attention was drawn in ?2.3 to some important differences
between these two sorts of mechanism; on a-priori grounds, we should not
necessarily expect every type of split to be distributed evenly over them.
In fact, in most examples of split conditioned by the semantic nature of verbs,
bound affixes are involved; whereas, in most examples of split conditioned by the
semantic nature of NP's, case-marking is involved. This is surely what could have
been predicted: that kind of morphological marking split which is CONDITIONEDBY
38There are slight complications
beyond the diagrammaticarray here. The accusativesuffix
-na MUST be used on proper nouns in 0 function, and on common nouns which bear a non-
singular numbersuffix; it CANbe used on some common nouns but is not frequenthere (data
from Luise Hercus).
39 Silverstein'shierarchy,along with the number system, provides explanations for the great
majorityof split case systems; but there are a few oddities that cannot be accounted for. Thus,
in the Australianlanguage Gumbaynggir, ldu. and 2sg. pronouns have one form for S and A
functions and another form for 0 function; but Isg., Ipl., 2du., and 2pl. have three distinct
forms for S, A, and 0 (see Eades 1979). There may well be diachronic explanationsfor some
exceptions of this type: e.g., phonological constraints may have blocked a certain change in a
particularenvironment.
the semantic nature of the verb is most often REALIZED BY an affix to the verb; that
variety of split which is CONDITIONED BY the semantic nature of NP's is usually
REALIZED BY affixes or particles attached to NP's. (In Choctaw, it will be recalled,
case inflections show a straight nominative/oblique pattern; but verbal prefixes are
split according to the semantic content of the verb.)
There is, then, an ASSOCIATION between types of conditioning and types of
morphological marking; but a few examples go against the majority pattern. Bats
and Eastern Pomo have been mentioned as examples of fluid S-markinglanguages,
which work in terms of case inflections (but no example is known of a case-marking
language with strict split S-marking). A language with split conditioned by the
semantic nature of NP's, but realized by cross-referencingaffixes, is more unlikely.
Indeed, a combination of this sort can be seen, on a-priori grounds, to be rather
implausible. An A NP would really need to be cross-referenced in the verb ONLY
if it were of a certain semantic type; and similarlyfor an O NP. Thus, whether the
verb showed concord with NP's in the sentence with it would depend on the
particular nature of the NP: for the application of a grammatical process to one
part of speech to depend on semantic information elsewhere in the sentence would
be an unusual state of affairs, something seldom encountered in natural language.
The normal situation is for a verb always to contain an affix cross-referencing
certain features on A and/or 0 NP's. (But note that '3rd person' or '3rd person
neuter' may often have zero realization; this could perhaps be taken as 'null
marking' in terms of the hierarchy for 0 cross-reference, though not for A.
Decisions between 'zero term from a system' or 'the system not applying' would
have to depend on detailed analysis of the language concerned.)
There is at least one example of a cross-referencinglanguage with a split of this
kind. Silverstein describes how in Chinook the S, A, and 0 NP's are always cross-
referenced by verbal prefixes. The various prefixes go into different positional slots.
There are identities and similarities between the forms filling different slots (for a
given person/number combination), and these are suggestive of a Dyirbal-type
split; in particular, reference to the A NP involves the addition of-k- to the regular
S/0 prefix form for certain persons/numbers. Here we do NOT have zero cross-
reference corresponding to the left of the hierarchy-but we DO have the addition
of -k- to the regular prefix, just for certain possibilities that Silverstein shows to be
hierarchicallydetermined.
A more common phenomenon is for bound prefixes to indicate the RELATIVE
positions of A and 0 on the hierarchy. We should expect A to be further to the left
than 0; choice of verbal affixes may depend on whether or not this does hold. In
Algonquian, e.g., each transitive verb selects one of four suffixes: (i) *-a.- denotes
'action by 1st or 2nd person on 3rd, and by 3rd person proximate on 3rd person
obviative'; (ii) *-ekw-, the inverse of (i), denotes 'action by 3rd person on 1st or 2nd
person, and by obviative on proximate'; (iii) *-e6e(ne)-, denotes 'action by 1st
person on 2nd person'; (iv) *-i- denotes 'action by 2nd person on 1st person'
(Goddard 1967:67).
The Algonquian type of marking is not normally regarded as an example of
'ergativity'. Yet it is plainly a phenomenon of the same sort as the NP-conditioned
split case systems we have been dealing with. This suggests that 'ergativity' is a
ERGATIVITY 91
misleading label, which can only indicate a part of the wider, integrated field of
'surface marking of syntactic-semanticfunctions'; I attempt, in this paper, to cast
my discussion in terms of this inclusive field.
3.26. I know no examples of languages that combine a split conditioned by
semantic content of verb (described in ?3.1) with a split conditioned by semantic
content of NP's, where both splits are realized in terms of morphological marking
of the same kind. Splits of the first type involve consistent marking of syntactic-
semantic functions for transitive clauses-but more fluid marking of S function,
depending on the semantic nature of the intransitive verb involved. Splits of the
second type provide consistent treatment of S NP's-but more fluid marking of A
and O functions in transitive sentences, depending on the semantic nature of the
NP's involved. In the one instance, semantic nuances in intransitive sentences are,
as it were, calibrated against a constant transitive schema; in the other, the semantic
orientation within transitive sentences is brought out against an invariable intran-
sitive matrix. If both were allowed to vary simultaneously-useful as this would be, to
bring out all the relevant semantic niceties-there would be no constant element,
and surely a likelihood of confusion and ambiguity. Grammatical structures and
rules, as abstractions from and idealizations of semantic relations, must organize
the material of a language in order to facilitate effective communication. Dual
conditioning of case-marking 'splits', of the type just suggested, might lead to
irresoluble anarchy, i.e. to semantically-sponsored variation that could go beyond
the limits allowable by a grammar.
However, a language could well have verb-conditioned and NP-conditioned
splits if these were realized by different morphological devices-one by case
inflections, and the other by cross-referencingaffixes. I return to this topic at the
end of the next section.
3.3. 'BOUND' VS. 'FREE'SPLIT.A further kind of split mentioned in the literature
on ergativity consists in different kinds of marking on FREE-formnominals (i.e.
case- or particle-markingon NP's) and in cross-referencingBOUNDaffixes. This is
best regarded not as a distinct kind of conditioning, but as a secondary pheno-
menon, explainable in the same terms as NP-conditioned split.
In ?2.1, I discussed the two major kinds of morphological marking: some
languages use cases exclusively, while others only employ cross-referencingverbal
affixes (and some use both mechanisms). I mentioned that a language can be
characterized as 'ergative' in terms of either type of marking.
These two morphological mechanisms may yield the same ergativity value. Thus,
in Latin, one case marks S and A NP's (and a different case marks 0 NP's); the
verb cross-referencesjust S and A. In Avar, a northeast Caucasian language, nouns
take ergative/absolutive (-as: - -c:a vs. 0) case inflections, and verbs cross-
reference gender and number only for S and 0 NP's (Anderson 1976:4). But they
may also be in conflict. We saw in ?2.31 that in Murinypata, one verbal prefix is
used to cross-reference S and A NP's, and another for 0 NP's (a 'nominative/
accusative' pattern), while NP's in A function can take ergative inflection (but there
is no inflection for S or 0 functions).
Plainly, this is a 'split' of a different kind from those discussed above. It might
seem at first glance that the split is not conditioned by the semantic nature of any
sentential constituent, but is instead a 'meta-split '-depending entirely on the
different GRAMMATICAL ways of realizing S/A/O identification. However, a close
examination of the phenomenon yields a semantic explanation.
Cross-referencingaffixes index a limited amount of information. They can make
choices from a number of grammatical systems: basically, person, number, and
gender. These systems provide a full characterizationof pronouns, but supply only
quite limited data on nouns. Verbal cross-reference makes free-form pronouns
virtually redundant (they tend to be used rather infrequently, mainly for special
emphasis); but at best it can only indicate the gender of a noun in a core syntactic
function.
Cross-referencingsystems are thus basically pronominal (with the affixes having
developed from free-form pronouns, in some earlier stage of the language). We
expect them to be on a nominative/accusative pattern, since this characterizes
pronouns, at the extreme left of the hierarchy. Case-marking on NP's is under no
such constraint, and can be either nominative/accusative or absolutive/ergative.
What we can predict is that, if there is a 'split' of this kind, then bound prefixes
will be accusative, and case-marking on free forms will be ergative. This is exactly
what is found. Both case-marking and cross-referencing affixes can be accusative,
or both can be ergative; but if there is a split, then bound forms will be accusative
and free forms ergative (as in Murinypata)-never the other way around. We can
thus regard this type of 'meta-split' as a corollary of the type discussed in ?3.2; it
can be given a semantic explanation in terms of Silverstein's hierarchy.40
Sometimes free-form pronouns can be straight ergative in their inflection, while
bound pronominal markers are pure accusative; this is so for Walbiri, among other
languages. The following fairly certain chain of historical development (cf. Hale
1973) provides an explanation:
(a) Originally there was a simple split case-marking system, conditioned by the
semantic content of NP's. Pronouns (for all persons) followed a nominative/
accusative paradigm, and nouns an absolutive/ergative pattern. At this time there
were no bound pronominal affixes.
(b) A system of cross-referencing suffixes developed, on the verbal auxiliary, as
reductions of free-form pronouns; these followed the free pronouns in having a
nominative/accusative paradigm. Since full pronominal information was now
obligatorily included in the auxiliary, the use of free-form pronouns diminished
(and was only necessary for emphasis etc.)
40
Rumsey, MS, arguesthat if a languagehas bound pronominalprefixeson verbs, in addition
to case-markingon nominals, then the latter sufficeto markthe syntacticfunction of a core NP
in a sentence, while the primaryfunctions of the former will be coreferenceand anaphora.The
prefixes will be basically oriented to the syntactic pivot (cf. ?6.2, below); it is because S/A is
the most common pivot that bound pronouns tend to have a nominative/accusativeparadigm,
Rumsey suggests. He predicts that if Dyirbal were to develop bound pronominal affixeson the
verb, these would follow an ergative/absolutivepattern-following the S/0 pivot in Dyirbal
(?7.2, below)-although free-form pronouns are nominative/accusative.There is undoubtedly
some value in Rumsey's suggestion, although it identifies at best only one of many factors
involved.
ERGATIVITY 93
(c) Walbiri morphology was simplified, in that the absolutivelergative case

system on nouns was generalized to apply to free pronouns. The original 'nomi-
native' pronominal forms now receive ergative inflection in A function.41
I suggested in ?3.25 that a split conditioned by the semantic content of verbs
could not be combined with a split conditioned by the semantic content of NP's,
because of the confusion this would engender. My arguments were valid for a
language which employs just one method of morphological marking. But for a
language that uses both cases and cross-referencing affixes (something that might
seem to imply a measure of grammatical redundancy), the possibility is open. No
example is currently known, but it is rather likely that a thorough search might
uncover one. (Choctaw points in the right direction. Here bound affixes on the
verb show a split of the verb-conditioned type; Choctaw also has case-marking on
NP's, but this seems not to involve any split.) We have noted that most examples
of verb-conditioned split use cross-referencing, while most instances of NP-
conditioned split employ case inflections; a combination of the two could prove
both semantically revealing and grammatically workable.
3.4. SPLITCONDITIONED
BY ASPECT/TENSE
OFSENTENCE.
Each natural language has
two varieties of 'shifters'. The pronominal system was mentioned in ?3.21; this
involves orientation to the speaker 'I'. The other area in which shifters are always
encountered is time reference. The focus here is primarily 'now', the moment of
speaking; there is also always a secondary focus 'today'. Relative to these shifting
origins, languages show either a grammatical system of tense inflection, or a
lexical class of time qualifiers,or both.42
Just as one type of ergativity split can be explained in terms of a semantic
hierarchyextending from 'I', through 'you' and other shifters, to nominal referents
that are increasinglydifferentfrom the speaker (?3.2), so a further type of ergativity
split can be explained with reference to the different ways of regarding events that
are established facts (roughly, completed before 'now') vs. those that are merely
prospective possibilities.
There are basically two ways of viewing time: first, as a gradually unfolding
scale, with 'now' as a point which moves along it at a steady rate, as in Figure 2.
past present future
> * >
snow'
FIGURE2
Alternatively, time can be viewed by looking in both directions from the constant
origin 'now', as in Figure 3.
past present future
'now'
FIGURE 3
41 It is possible to reconstruct stages earlier than (a); see Hale 1973, Dixon 1977c. Discussion
of these would add nothing to the example here.
42 Tense systems are always relative to
'now', but can sometimes also involve 'today'; see
Hymes 1975. Time qualifiersappearalways to referto 'today*-including items like 'yesterday',
In this perspective, there is continual movement of time from the right-hand

side, past the origin 'now', to the left, as the guessable but essentially unknowable
future becomes the documented past. Tense systems in a few languages appear to
operate according to Fig. 2, according equal grammaticalstatus to 'past', 'present',
and 'future' (Ancient Greek appears essentially to be of this type). But most
languages treat 'future' in a quite different way from 'past'; cf. English, which has
an inflection for past tense but employs an array of modals for future reference-
one must indicate whether something should or might or could happen, or is
predicted to happen, etc. (see Lyons, 304-17).
Every language has some syntactic means for linking descriptions of a series of
connected actions. There are basically two ways in which such a series can be
viewed:
(a) It may simply be seen as a series of actions that all happened to involve a
certain participant. Here no causal connection between the events need be stated (or
implied): it is simply that the events ARE documented, and linked through a
common participant. This viewpoint is most compatible with a retrospective time
perspective-looking backward from 'now', in Fig. 3, to a series of known and
documented events.
(b) It may be seen in terms of an agent initiating and controlling a series of
interconnected actions: he undertakesX so that he can then engage in Y, as a result
of which Z will follow, etc. The common participantto the events must, in this view,
be the referent of S and A NP's in the sequential clauses. Here the events follow a
causal sequence, moving forward through time; this viewpoint is compatible either
with Fig. 2 or with the future perspective in Fig. 3.
I have suggested that (b), which demands syntactic identification of S and A as
controlling 'agent', is the expected alternative for future-time discussion, and can
also be employed for past time, as in Fig 2. But with (a) we could equally well get
either S/A or S/0 as syntactic pivot; this viewpoint is most plausible in past time.
An analogy may help here. The classic crime thriller begins with a series of events
that have all befallen some participant-a victim, a sum of money, or whatever.
The detective notes the events and the connection between them. He then tries to
establish the agent and the chain of causality. Effectively, he begins with viewpoint
(a), and then re-interprets the events from viewpoint (b). Once this is successfully
done, he can project the modus operandi of the criminal into the future, predict his
likely actions, and perhaps trap him.
These different syntactic orientations for known events, on the one hand, and for
potential happenings, on the other, can aid in predicting the form of a morpho-
logical split conditioned by tense or aspect. If absolutive/ergative marking is found
in one part of the system, we would expect it to be in past tense or in perfect aspect,
where a series of completed events could be related to 0 and S as pivots. In non-
past tense or in imperfectaspect, nominative/accusativemarkingwould be expected.
Something that has not yet happened is best thought of as a propensity of the
'tomorrow'-but sometimes also involve 'now' (in the latter case, the class contains items
'earlier today' and 'later today'). (Further discussion and exemplification from the Australian
language family are given in Dixon 1977a: 498-500.)
ERGATIVITY 95
potential agent ('That man might hit someone', rather than 'That person might
get hit by someone'); this must involve A and S NP's as pivot. (There is useful
discussion of this point in Regamey.)
This is exactly what is encountered. Many languages can, of course, have
nominative/accusative marking in all aspects and tenses, and others have absolutive/
ergative marking unimpeded by aspect or tense (it may be conditioned by one of
the factors mentioned in ??3.1-3.3). But if a split is conditioned by tense or aspect,
the ergative marking is ALWAYSfound either in past tense or in perfect aspect.
Burushaski, a language isolate spoken in inaccessible mountain valleys of the
Karakoram Range, on the border between Kashmir and Tibet, shows this type of
split. Here an A NP is obligatorily marked for ergative case (with the suffix-E),
only if the transitive verb is in a past-based tense (i.e. preterit, perfect, pluperfect
past participle, or static participle active); if the verb is in any other tense, ergative
will not be used (Lorimer 1935:64).43 In non-past tenses, there are no inflections
for A, S, orO (as there are none for S orO in past tenses): Burushaski shows a split
of ergative in past tenses vs. no case-marking in non-past. (It appears that, in non-
past tenses, A and O are distinguished partly through word order and partly
through pronominal prefixes; the latter cross-reference S NP's for intransitive
verbs; indirect objects for ditransitive verbs like 'give', 'tell', or 'say to'; and
direct objects for transitive verbs like 'strike', 'see', and 'kill'; Lorimer, 192 ff.)44
Many other examples in the literature show this type of split; generally (by
contrast with Burushaski), explicit nominative/accusative marking occurs in non-
past/perfect sentences. Ergative/absolutive marking occurs only in past tense for
Tibetan (Regamey 1954) and only in perfective aspect for Hindi (Allen 1951,
Kachru 1965), Rajasthani (Allen 1960), Georgian (Vogt 1971), and Samoan
(Milner 1973) and only in perfective aspect in Yucatec Mayan (Bricker, MS). In
some instances, the marking involves case inflections (Burushaski, Georgian) or
particles (Samoan). There is generally positive marking for A function in past/
perfect, and for 0 function in non-past/imperfect-creating a genuine 'accusative'
vs. 'ergative' split (e.g., Samoan has ergative e and accusative i'i); absolutive/
nominative then has zero realization. In other languages, the marking is shown by
verbal affixes: in Yucatec, A is cross-referencedby a prefix and 0 by a suffix, while
S is marked by the prefix system in incomplete aspect and by the suffixin completive
aspect.
Some verbs in Abkhaz-Adyghe (West Circassian)can occur in two constructions,
43 Lorimer mentions two furtherpeculiarities:ergative MUST be used on the A NP of henWs

'to know' in non-past tenses, and it CAN be used on the A NP of SEfAS 'to say' in non-past
tenses. He mentions that ergative is permissiblewith other transitiveverbs in non-past tenses,
but is seldom used there; it is obligatory with all transitiveverbs in past-basedtenses.
44 Alan Rumsey has pointed out a universal tendency: those NP's which are likely to be
highest on the 'human/animate/inanimate'hierarchywill be cross-referencedin the verb. Many
languages have just two NP's cross-referenced:one is the NP in S or A function, while the
other is the indirect object (if there is one) or direct object (in the absence of an indirectobject).
In most sentences, the referentof the indirectobject will be higherthan that of the direct object.
Languages behaving in this way include the Australian Walmatjari (see ?7.1), Rembarnga
(McKay 1975), Ungarinjin and Nunggubuyu; the Siouan Lakhota (Van Valin 1977:7); and
Chukchee (Comrie, MS b), among many others.
with a semantic difference reminiscent of the splits just described. Briefly, with an
ergative A and absolutive 0, the implication is that something WASDONE;but with
absolutive A and oblique 0, the sense is that the agent was TRYINGto accomplish
something. Thus a single verb can mean 'strike, stab, beat' in absolutive/oblique
constructions, but 'kill' with ergativelabsolutive marking; and another verb is
translated as 'look at' and 'see' respectively (Catford 1975:32-3; cf. Anderson
1976:20-22).
I made the point in ?3.26 that languages do not (and would not be expected to)
mingle verb-conditioning and NP-conditioning of splits within the same system of
morphological marking. We do, however, encounter a mixture of one of these
types of split with a split conditioned by aspect/tense. Thus Burushaski has
distinct ergative forms for all nouns and for singular pronouns, but non-singular
pronouns have a single form for A, S, and O functions in all tenses.
Perhaps the most interesting example is Yukulta, from the Gulf Country of
Queensland. This shows a split case system, conditioned partly by tense/aspect-
type considerations and partly by the semantic nature of NP's; but the conditioning
is more complex than in Burushaski. Keen 1972 recognizes two kinds of transitive
construction, with the following inflections:
A 0 VERB MARKEDBY
(32) (a) ergative absolutive transitive suffixes
(b) absolutive dative intransitive suffixes
Ergative is basically -ya; dative is -nYdYa -dpi; and absolutive has a number of
allomorphs, one of which is zero (with a vowel-final stem of more than two syllables).
Construction 32a is used with statements of past fact and of future intention; but
32b is required in all other cases, i.e. for negative sentences in past tense ('He didn't
do it') and for future irrealis (e.g. wishing). We see that this split is not simply in
terms of past/future tense, but in terms of things that have happened (or are
promised to happen) vs. those that have not happened or might conceivably happen.
Construction 32b must also be used, whatever the tense/polarity choice of the
sentence, if (i) A is 3rd person and 0 is 1st or 2nd person, or if (ii) A is 2nd person
and 0 is non-singular 1st person.45This second conditioning factor appears to deal
with the relative positioning of A and 0 NP's on the hierarchy of ?3.25, somewhat
as in Algonquian; it is explainable in terms of the discussion there, except for the
odd specification of NON-SINGULAR 1st person under (ii).
3.5. 'MAIN' VS. 'SUBORDINATE'CLAUSESPLIT. The literature on 'ergativity'
contains some mention of a split conditioned in another way: morphological
marking may differ between 'main' and 'subordinate' clauses. At first this appears
to be a grammatically conditioned split, of a quite different type from the seman-
tically motivated splits dealt with above. However, deeper consideration shows that
this type of division can be related to tense/aspect-type splits, and that it does have a
semantic basis.
The term 'subordinate clause' covers a variety of phenomena with different
kinds of semantic implication. Thus 'purposive (= infinitival) complements'
45 Blake 1976 mentions similar phenomena in two other Queensland languages, Kalkatungu
and Pitta-Pitta.
ERGATIVITY 97
normally refer to some attempt at controlled action; clauses of this kind generally
have an A or S 'agent' NP that is coreferential with some NP in the main clause
('We went to pick fruit', 'We went to play', 'I told you to pick fruit' etc.) For this
type of subordinate construction, we would surely expect S and A to be treated in
the same way within the complement clause.
Then there are relative clauses, most commonly detailing some action that
takes place at the same time as (or previous to) the event referred to by the main
clause: 'I kissed the child who had fallen over', 'I heard the man cutting wood',
'I saw the child being spanked by its mother.' Here the relative clause simply
describes some event (usually an actual or completed event) that related to an NP
in the main clause: any syntactic orientation is possible.
Now if there were a split in morphological marking between main clause and
purposive complement clause, we should expect the subordinate clause to show
'accusative patterning'-while the main clause would, if it differed from the sub-
ordinate clause, require an 'ergative pattern'. But if it were relative clauses that
entered into a split, we might expect the subordinate clause to show 'ergative'
characteristics-and the main clause would, if it differed from subordinate clauses,
be of the 'accusative' type. (In the great majority of languages, of course, the same
marking conventions apply to all clauses, whatever their grammatical status.)
The point at issue here is that purposive complement clauses are like main
clauses in future tense (or imperfective aspect): they express some potential event
as a propensity of the (A or S) agent, and thus demand nominative/accusative
marking. However, relative clauses resemble past tense (or perfective) main clauses
in simply describing something that has happened or is happening; they can
appropriately be treated in an absolutive/ergative manner. (See also Dixon 1977c
for related discussion.) Main clauses, for which either accusative or ergative
marking is appropriate, must show the type of marking OPPOSITE to that of the
subordinate clause, IF there is a split.
Very few examples are known of main/subordinate clause splits; those that have
been mentioned are not clearly enough described for us to be quite certain of all the
relevant details. One of the clearest examples comes from Tsimshian (Boas; I am
grateful to Michael Silverstein for drawing this to my attention). Here subordinate
('subjunctive') clauses-e.g. 'Then he heard HIM COME AGAIN', 'His mother was
glad WHENSHESAWHIM '-consistently show an ergativepattern of cross-referencing;
a 'subjective' verbal prefix refers to A, and an 'objective' prefix to S or 0. But in
main ('indicative') clauses, 'objective' is used for A cross-reference,as for 0, when
1st or 2nd person is acting on 3rd person; subjective prefixes are used for the A NP
in a main clause only when 3rd person is acting on 1st or 2nd person (thus going
against the hierarchyof ?3.2). Tsimshian subordinateclauses clearly show 'ergative'
marking, while main clauses demonstrate a split between 'ergative' marking and no
marking at all, the split being conditioned by the nominal hierarchy.
Further examples, which appear on the information available to support the
hypothesis suggested above, are the Australian Ngarluma (Hale 1967; but see
Nash, MS) and Lardil (McConvell 1977, following Klokeid 1976).46
46 Craig 1975, 1976 reports a split according to clause type in Jacaltec, a Mayan language of
Guatemala. In main clauses, Set 1 of pronominal affixescross-referencesA, while Set 2 refers
It is clear that to talk simply of 'main clause' vs. 'subordinate clause' split is
misleading. Any analysis of this phenomenon will have to take account of the types
of subordinate clause involved and their semantic function. At present the need is
for detailed and reliable data on this type of split, in a number of languages, to see
how well it fits with the a-priori explanation presented above.
3.6. SUMMARY. In ?3, each type of split of morphological marking (whether
realized by case inflections or particles or cross-referencing affixes) has been
explained in semantic terms. Some splits are motivated by the semantic content of
intransitive verbs, others by the semantic contents of A and O NP's within a
transitive sentence, and others by the aspect or tense choice of the sentence.
Differences of marking between main and subordinate clauses have been related
to tense/aspect-type distinctions, while differencesbetween case-markingand bound
affixes have been related to the hierarchy that underlies NP conditioning. Case-
marking is also, of course, syntactically motivated. In ?8 I will summarize the
varying pressures on morphological marking-universal and language-particular
syntactic requirements (discussed in ??5-7), as well as the semantic preferences
described above.
The logical possibilities for each type of conditioning are given in the Appendix.
Note that the largest number of possibilities belongs to NP-conditioned splits;
these are typically realized by case inflections (and most commonly encountered in
Australian languages, although examples are found in other continents). Verb-
conditioned splits are typically realized by bound affixes (here the largest con-
centration of examples is in language families from the central USA). Tense/aspect
conditioning can be associated with cases or with bound affixes, and is sporadically
attested from several parts of the world.
We could, in conclusion, enquire which type of split is most superficiallyergative
-and, at the other extreme, which type is most likely to correlate with ergativity
at the syntactic level. Any answer to this question must at present be quite ten-
tative: detailed investigation of the morphology and syntax of a large number of
'ergative' languages (in terms of the distinctions stressed in this paper) would
be required before we could put forward any firm hypothesis. It is, however, tempt-
ing to indulge in a-priori speculation. Syntactic processes typically operate with
NP's as pivots, and a particular case-marking is perhaps more likely to corre-
late with some syntactic property than is a corresponding pattern among
bound affixes. Since case-marking is most strongly associated with splits condi-
tioned by the semantic nature of the core NP's, it may be among languages of this
type (e.g. Dyirbal, ?1.1) that we will get the strongest examples of 'ergativity' at
the syntactic level.
to S or 0. But in subordinate 'aspectless' clauses, Set 1 refers to A or S, and Set 2 to 0. On the

data available, it looks as if Set 2 could be considered 'unmarked' with respect to Set 1 (3rd
person singular has zero realization in Set 2, but there are no zero forms in Set 1). In view of this,
we should be cautious about describing aspectless clauses in terms of a nominative/accusative
pattern; it may be more appropriate to refer to an 'extended ergative-restricted absolutive'
arrangement, as in ?2.33. The way in which this type of split fits (or does not fit) into the
hypothesis presented in this section awaits further study.
ERGATIVTY 99
EVOLUTION OF ERGATIVE MORPHOLOGIES

4. A fair amount has been written about the diachronic development of ergative
morphologies. I attempt here only to summarize some of the suggestions that have
been made, as they touch on the theme of the present paper.
Bound affixes (to the verb or some other sentential constituent), cross-referencing
semantic features on core NP's, are in almost every instance derived from free-form
pronouns and demonstratives. Splits of the 'bound' vs. 'free' variety can usually
be explained in diachronic terms. Thus, in ?3.3, I mentioned the strong evidence
(from comparison with other Australian languages) that Walbiri pronouns once
had a nominative/accusative pattern. They gave rise to accusative-type bound
suffixes on a verbal auxiliary; then the absolutive/ergative system of nominal
inflection was extended to hpply to free-form pronouns.
Splits between 'main' and 'subordinate' clauses may have come about through
syntactic reinterpretation in just one of these grammatical areas. Thus there is
clear evidence that, in the not too distant past, Ngarluma had an entirely absolutive/
ergative paradigm; by recent reanalysis, the stem form of a noun is now used for S
and A functions (previously S and 0), while O function is marked by -gu -yu -
-yi (the erstwhile dative inflection). It appears that ergative-type marking may have
simply been retained in certain types of subordinate clause. In Jacaltec it seems that
many subordinate clauses derive from nominalizations; this leads to an explanation
of why subordinate complements show a type of accusative patterning, while main
clauses are absolutive/ergative in marking (Comrie, MS a; Craig 1976).
Ergativity can be an uncomfortable phenomenon for anyone who speaks a fully
accusative language, or who tries to fit all languages into a linguistic theory induced
from the structures of accusative languages. Some 19th-centurycomparativists and
some recent transformational grammarians have sought to 'explain' away ergative
constructions as being basically passives. Little can be said in support of this as a
synchronic explanation; the difficultyof explaining the facts catalogued here in terms
of 'passive' should be sufficient to demonstrate this (cf. Kurylowicz 1960). But it
does appear that some (though by no means all) instances of ergative constructions
have developed from an original passive; e.g., one group of scholars has argued for
a passive-to-ergative shift in Polynesian languages (see Chung 1976, 1977, Hale
1968b, and Hohepa 1969).47
A passive origin is especially common for ergativity within a tense/aspect split.
Anderson 1977 suggests that one major way in which ergative developed for the
47 It must be remarked that an equally vigorous case has been put forward for the opposite
hypothesis-that Proto-Polynesian was essentially ergative, and that some of the modern
languages have developed an accusative pattern by historical change from this (see Clark 1973,
1976, Foley 1976). It seems to me that neither side in this argument has proved its case, and that
the matter requires further consideration. (See ?5.43 for comment on Chung's evidence.)
Hale 1970 suggested an accusative basis for Proto-Australian, although most modern
languages have a good deal of ergativity. This hypothesis was at one time supported by a number
of scholars, but the emerging factual evidence is increasingly against it. McConvell 1977 shows
that Lardil, a critical accusative language, has developed from an ergative ancestor. It is likely
that Proto-Australian had a split-ergative morphology, and was typologically similar to
Arabana (?3.24); see Dixon 1977c. Nothing can be said with certainty about the syntactic type
of Proto-Australian.
originally accusative Indic family, in perfect aspect, was through the loss of the
inflectional perfect, and replacement by a periphrastic construction based on a
participle that was passive in form. He makes the important observation that
'passive constructions are semantically close to perfect in that they generally pre-
sent a state resulting from a completed action' (cf. the semantic discussion in ?3.4,
above, and Comrie 1976a: 85-6).
Benveniste 1952 and Alien 1964 have drawn attention to the widespreadtendency
to construct perfect-tense forms for transitiveverbs by employing the verb 'to have'
as a verbal auxiliary. Some periphrastic perfects that involve 'to have' and a
participial form of the verb are found to show ergative marking (Anderson 1977).
After considering a number of instances where ergativity has developed through
restructuringof the perfect, Anderson notes that the ergative case tends to coincide
with instrumental when it has a passive origin, and to coincide with genitive or
dative when it evolves through a possessive route.
It is far from being the case that all instances of ergative morphology are caused
by recent developments within an accusative ancestor. First, some semi-ergative
morphologies may be quite ancient, going back to a distant proto-language (as in
Australia; Dixon 1977c). Second, the change ergative-to-accusative is just as
plausible as accusative-to-ergative; i.e., there is certainly no unilinear development.
Just as passives can give rise to ergative structures, so antipassives can be the basis
for accusative reinterpretation. (A specific example is described in some detail in
Dixon, MSa.)
Less can be said about the development of splits conditioned by the semantic
content of NP's or of verbs. Pronouns typically show a slightly different pattern
of inflection from nominals, often involving irregularitiesand archaic features; this
could well interrelate with semantic explanations for some NP-conditioned splits.
For verb-conditioned splits, we might speculatively posit an original ergative
system, and then suggest that some transitive verbs underwent a semantic shift that
led to the elimination of previously obligatory object NP's: they would be re-
analysed as intransitive verbs with the S NP marked by ergative case.
Note that the facts available do not support any suggestion that one type of split
might be the origin for all modern varieties of ergativity. Consider the hypothesis
that a Dakota-type system, split according to the semantic content of verbs, was
the genesis for all ergativity. But if this were so, we should surely expect some trace
of a verb-conditioned split among the 200-odd languages of Australia, almost all of
which show some degree of ergativity; but all ergativity splits in Australia are
conditioned by NP-content or by aspect/tense. There seems no doubt that the
morphological phenomena grouped together under 'ergativity' have evolved in a
variety of ways from diverse beginnings.48
48 Many other factors, beyond those mentioned here, play a role in developments to or from
ergative patterning (e.g. causatives and other transitivizing processes; cf. ?5.4). And sometimes
changes that are motivated by a phonological restructuring may lead to realignment of case-
marking. Dixon 1977c mentions that Proto-Australian had yay, nyun - yin for 1st and 2nd
person singular pronouns in S function, with ergative forms vadauand nYundu pindu in A
function. Then a restriction that each word must consist of at least two syllables was introduced,
and in many languages the A form was generalized to cover S; this produced a nominative!
accusative paradigm for singular pronouns, at the left of the hierarchy in ?3.2.
ERGATIVITY 101
There is no reason to suppose that an out-and-out ergative case system could not
be as stable as a fully accusative arrangement,from a syntactic-semanticviewpoint.
But there is always a tendency for analogy to extend generalizations, effectively
eliminating 'splits' in morphological marking. I have cited ergative inflection in
Walbiri being generalized to free-form pronouns. Data on Zan, a Kartvelian
language, suggest that ergative case was originally used to mark the A NP just in
the aorist (as in the related Georgian). In Mingrelian, one dialect of Zan, the
original ergative marker is now used to mark A and S in the aorist; there is now
plainly a nominative/accusative system, in aorist as in non-aorist.49 But in Chan,
another dialect of Zan, ergative marks A (but never S) in ALLTENSES;Chan has an
absolutive/ergative syntax, in aorist as in non-aorist (Comrie 1973:252). Both
dialects have eliminated the original aspect-conditioned split; one has a fully
nominative/accusative case system, while the other shows consistent absolutive/
ergative marking, through different directions of generalization.
THE CATEGORY OF 'SUBJECT'
5.1. PRELIMINARY
REMARKS.Turning now to syntax (and beginning at the
deepest level) we can first of all note the confusion concerning the identity of the
'subject' in ergative languages. The confusion results simply from the fact that
linguistic theory evolved in the context of the better known languages of Europe,
which have a predominantly accusative character at every level. For languages of
this type, certain semantic and grammaticalproperties coincide to give a two-sided
definition of subject. The 'subject' of a sentence is that NP whose referent COULDBE
the 'agent' that initiates and controls an activity; the subject NP is normally
obligatory in a sentence, receives the unmarked case, may be cross-referenced in
the verb, and is the pivot for operations of coordination and subordination.
For ergative languages, these semantic and grammatical criteria for 'subject' do
not coincide ;50 to employ the notion of subject in such languages, one must decide,
in effect, which of these two kinds of criteria should take precedence. Some
49 The marking on A NP's not infrequentlyis generalized to apply to S NP's; here the
syntactic-semanticnotion of 'subject' is being marked at the morphologicallevel (as happens
in languagesfrom the Cushiticand Yuman families; see ?2.33). But as Anderson 1977 remarks,
there are no attested instances of an accusative marker (for 0 function) being generalizedto
apply to S over all types of NP constituent; see fn. 24. Thus, while an ergative system may
develop into an accusativeone (but an accusativepatternwith 'marked nominative'; see ?2.33)
simply through generalizationof the A case-markeralso to cover S function, an accusative
system will seldom (or never?) develop into an ergative one through the 0-marker's being
generalized to cover S. Rather, deep A may become surface S in some (antipassive-like)
derivation; and this may then become, through re-analysis,the regulartransitiveconstruction.
50 There is
an indirect analogy to the unit 'word'. Every language has a unit '(grammatical)
word', with considerable psychological reality for the speaker (cf. Sapir 1921:33-5); and
grammaticalcriteriacan always be given to define this unit, although the nature of the criteria
differ from language to language. It appears also to be the case that every language has a
phonological unit larger than the syllable, which can be called '(phonological) word'; and
phonological criteria, usually involving considerations of stress etc., can always be given to
definethis unit. Now in most languages,'grammaticalword' and 'phonological word' coincide;
but this is not so in EVERY language. To mention one exception, in YidinYa grammaticalword
consists of ONE OR MORE phonological words (see Dixon 1977a,b).
linguists emphasize semantic criteria, but encounter severe difficultiesin explaining

all types of grammatical processes in terms of semantically defined 'subject' for
ergative languages. (In ?6.3, I describe the difficulties which Relational Grammar
has in accounting for antipassive derivations.) Other linguists take syntactic/
morphological criteriaas basic; this facilitates statements of grammatical derivation,
but is bound to complicate any attempt to provide semantic interpretation for the
grammar. (In the case of Dyirbal, Keenan takes 'absolutive NP' as subject: it
bears unmarked case, is obligatory, is the pivot for most syntactic operations, etc.
But this 'subject' relates to S and 0, not S and A functions!)
Although all languages have reasonably complex grammars, with comparable
sets of parts of speech etc., the details of syntactic and morphological patterning
do show wide variation. Classes of 'noun' and 'verb' can, it seems, be recognized
for every language, on internal grammaticalcriteria. But these grammaticalcriteria
differ from language to language; recognition of 'noun' and 'verb' between
languages must invoke semantic criteria (that class which includes words referring
to concrete objects is called 'noun', etc.; see Dixon 1977d). In fact, any attempt to
establish true universals must be semantically based. 'Subject' is surely more likely
to be definable as a universal category, playing a productive role in the grammar of
each language, if viewed from a semantic angle.
I will show that 'subject' is a universalcategory, having a vital role in the grammar
of every natural language, be it 'accusative', 'ergative', or any mixture. But-and
this is a vital point-it is not the MOSTfundamental category. 'Subject' links
functions from intransitive and transitive sentence types; it effectively involves a
grouping of S and A, out of the basic semantic-syntactic relations,S, A, and 0-a
grouping that is made entirely on semantic grounds.
The basic thesis of this paper is that A,S, and 0 are the universal core categories,
and that syntactic rules in every grammarare framed in terms of them. There is, as a
further stage, a universal grouping of A and S as 'subject''-a category that plays
an important role in every grammar. Some types of syntactic processes will always
be statable in terms of 'subject', in every language. Other types of process may
relate to 'subject' in some languages, but in other languages they may involve some
other combination of the core functions.
5.2. UNIVERSALDEFINITIONOF 'SUBJECT'. People observe events, of many
different sorts, happening in the world; the participant roles show a great deal of
diversity, and the effects of the activities on the participants vary a great deal. Yet
all human languages classify actions into two basic types: those involving one
obligatory participant, which are described by intransitive sentences, and those
involving two obligatory participants,which are dealt with by transitive sentences.51
In some languages, verbs are strictly classified into transitive and intransitive
subsets, and these are mutually exclusive (derivations from one type to another are
possible, but in this kind of language are usually marked morphologically). Other
51
Activities involving three obligatory participants are also dealt with by a type of transitive
sentence; i.e., the 'subject' of a verb like 'give' or
'show' or 'tell' is always dealt with in the
same way as the 'subject' of 'hit' or 'see' (whereas the subject of 'run' or 'fall' may be treated
quite differently). We should, more accurately, talk of events with one vs. more than one
obligatory participant. (See also fn. 52.)
ERGATIVITY 103
languages may have SOMEverbs that can function either transitivelyor intransitively;
even here, though, MOSTverbs belong unambiguously to the transitive or to the
intransitive set. In each language, there is explicit grammatical marking of each
transitivity type-in terms of case-marking on nouns and pronouns, occurrence of
particles, pronominal affixes on verbs, inflectional allomorphs on verbs, and so on.
The details of grammatical marking vary from language to language, but the SAME
TYPES of criteria recur.
Thus all languages treat 'cut' and 'give', 'rub' and 'carry', 'take' and 'cook' as
transitive verbs. In addition, very nearly every language classifies 'see' and 'hear'
in the same way. This is really a surprising fact-these verbs refer to totally
different kinds of events; but they all involve two basic participants, and are dealt
with by verbs belonging to the semantic-syntactic class 'transitive' in all types of
language. What is even more surprising is that all languages consistently identify
participants between these different verbs. By this I mean that the participant who
makes the incision (for 'cut') is equated with the person who transfers possession
of something he has had (for 'give'), with the participant who receives the sense
impression (for 'see'), and so on. I denote this participant by the functional label
'A': the A NP's for 'cut', 'give', 'see' etc. are consistently treated in exactly the
same way, in all aspects of morphology and syntax, across every type of human
language. The other participant in these transitive activities-that which is incised
(for 'cut'), whose sense impression is noticed (for 'see'), etc. is also treated in the
same way;52 here I use the functional label '0'. It is these facts which lead me to
suggest that the syntactic-semantic functions A and 0 are universal linguistic
primitives.
On a-priori grounds, the various participants that occur with different transitive
verbs would be classified in quite diverse ways. Fillmore's original suggestion
(1968) that 'cut' involves NP's in Agentive and Dative cases, whereas 'see' requires
Dative and Objective (and Dative is used to describe 'what is cut', but 'who sees')
has strong appeal, as a semantically-based description. Why is it that no language
(or, at least, none of the many tongues investigated by me, or by colleagues I have
asked about this) treats the NP's associated with 'cut' and 'see' in this manner?
The only language that does not treat 'see' and 'hear' in exactly the same way as
'cut', 'give' etc. is the northeast Caucasian Avar (Cerny). But here both 'that which
is seen' and 'that which is cut' are in absolutive case, the 'cutter' is marked by
ergative/instrumental inflection, and the 'see-er' is in locative case. With this one
exception, A and 0 NP's are consistently identified between 'cut', 'see', 'give',
'carry' etc., over languages of every typological sort.
It seems that the basic reason for this identification lies in the idea of 'agency' or
52
I mentioned (in fn. 51) that verbs like 'give' involve three obligatoryparticipants.Here the
'giver' is always equated with the 'cutter', 'see-er' etc. (as the A NP); but the 0 NP (which is
given the same grammaticaltreatmentas 'that which is cut', 'that which is seen' etc.) is in some
languages the recipient, in others the gift. (In still others, there are two construction types
involving 'give': in one, the recipientis 0 NP; in the other, the gift is. This applies in English
and in Dyirbal; Dixon 1972:300.) Positive criteria are employed to pick out the A NP. The
'other' obligatory participantis then identified as being in 0 function; if there is more than
one additional participant, then either could conceivably be taken as the 0 NP (see Kuipers
1968:614).
'control'. For most multi-participant events, there is just one participant who
potentially initiates or controls the activity. It is the NP referringto this participant
that is identified as being in A function (on negative grounds, the other obligatory
O
participant is then taken to be in function in each instance). In Mary cut John,
it is Mary (if anyone) who controls what is happening. It may be that Mary cut
John accidentally, so that no one can be identified as the controller, for some
particular token of this sentence; but it is clear that John can never be taken as
'agent'.
Some activities require the cooperation of two participants: both are, in a sense
'agents'. It is generally possible to focus on either participant (as the A NP), either
through employing two different but semantically-relatedverbs (e.g. Mary sold it to
John, John bought it O from Mary) or through a single verb that effectively allows
interchange of A and NP's (Mary shook hands with John, John shook hands with
Mary).53 In such an event, it is often the case that one participant does play a
leading role in initiating the transaction/salutation; this can be indicated by
assigning him to A function. But where double agency is required by the nature of
the event, either participant could conceivably be shown in A function; which is
chosen will depend upon who the speaker wishes to focus on as being, in terms of
his total discourse, 'the protagonist ... at the center of events' (Schachter 1977:
283).
'See' and 'hear' do not describe actions; indeed, these verbs cannot-except in
quite marked circumstances-be used in imperative form in English or in most
other languages. They contrast with 'look at' and 'listen to', which more clearly
involve the idea of volition and effort on the part of the observer, and which can
appear in imperative constructions (like almost every other transitive verb). But all
verbs that involve some further specification-hyponyms of 'see', such as 'watch',
'observe', 'scan', 'ogle''-plainly involve the referents of the A NP's initiating or
controlling the event (they also occur as imperatives). 'See' and 'hear' themselves,
as the most neutral verbs describing visual and aural reception, scarcely accord
with our criterion for why one particular participant is, in almost every language,
marked as A, on a par grammaticallywith the A NP's for 'cut', 'give' etc. But with
all other more specific verbs of seeing and hearing, the referent of one NP does
initiate or control the event, satisfying the criterion for recognition of this NP as
being in A function; 'see' and 'hear' as treated in the same way as their hyponyms,
'the one who sees' being assigned the same grammatical marking as 'the one who
(purposefully)
63
watches'.54 Mary saw John can describe an event where John just
Verbs of this kind can be symmetrical-like 'marry', 'meet', 'shake hands', and 'kiss'-
or directional; compare Mary rented the house to John, John rented the house from Mary with the
sell/buyexamples (see Dixon 1973). The matter of whethersymmetricverbs should be regarded
as underlying intransitives with plural subject (John and Mary married), from which the
transitivesare derived by syntactic operations, is too complex a question to go into here. But
note
64 that such a 'solution' is not available for buy/sell, rent to/rent from etc. (I am gratefulto
Paul Schachterand Rodney Huddleston for discussion of 'double agent' verbs.)

Support for this line of argumentcomes from Australian languages, which have a single
verb covering both 'see' and 'look at', and another for 'hear' and 'listen to'. That is, a single
lexical root is employed to describechance or involuntaryperception, and also for purposeful
ERGATIVITY 105
came into Mary's field of view, and no agency was involved; but there is the
potentiality that Mary looked for and sought out John-then Mary, but not John,
could be something like an 'agent'. (If John TRIEDto be seen, then some other verb
like 'show', would be appropriate).
We begin with our perception of the world: we see many activities, of many
different kinds. Then a grammatical classification is imposed on these events: they
are divided into those described by intransitive and those described by transitive
sentences. The grammatical requirement is that an intransitive sentence has a verb
and one core NP, whereas a transitive sentence has a verb and two obligatory
NP's.55 There is then a semantic identification of one transitive NP as being in A
function-this is consistent across languages, for transitive verbs of all semantic
types-and complementary identification of the other transitive core NP as in.O
function (but see fn. 52). This A NP refers to the actual or potential 'agent', who
could (if anything could) initiate and control the activity. Note that, in some
languages, the A NP must be animate (e.g. Jacaltec; Craig 1976:108-9);56it is then
likely that, for every transitive sentence, the A NP could be agent. Most languages
have some transitive verbs whose major occurrence is with an animate agent, but
which can also be used in an extended sense with an inanimate noun in the agent
slot-e.g., The wind closed the door; Sorrow is eating at my heart.57The central
directingof attention; in the latter sense, these verbs can of course be used in imperativeform.
Almost all Australianlanguagesshow this pattern(cf. Dixon 1972:41). In the Dyirbal 'mother-
in-law' style, a single verb root covers 'watch', 'stare at', 'look sneakily', 'look up at' etc., in
addition to 'see' and 'look at' (Dixon 1972:294-6).
&ernyimplies that in Avar all verbs of perception take absolutive 0 and locative A. In the
related Lesghian, there is a more subtle and interesting pattern of case assignment; here the
main class of transitiveverbs ('cut' etc.) has the A NP in ergative and the 0 NP in absolutive
case. There is a root akun that can occur in this regulartransitiveframework,and then means
'look at'; OR the 0 NP can still be in the absolutive case, but the A NP in a dative inflection,
being now translatableas 'see'. Similarly,the form van akun can occur in either framework,
meaning 'listen to' and 'hear' respectively.(The Lesghiandata were supplied by David Kilby,
from Mejlanova 1960.) This system of case assignment shows quite precisely the semantic
differenceinvolved, in the same sort of spirit that a few languages employ fluid S-markingto
show whetheror not agency is involved with some token of an intransitiveverb (?3.1). But just
as MOST languages have one marking for the S NP in all types of intransitivesentence, so
55
ALMOST every language treats 'see' in the same way as 'watch' and 'observe'-as a regular
transitiveverb.
Certain NP's are termed 'obligatory' not because they must necessarily occur in the
surfacestructureof every sentence involving a certain verb, but because the speaker and hearer
must have some understandingof them if the sentence is to form a conceptual whole-with
potentiality of referringto some actual, possible, or habitualevent.
I refer to 'intransitiveverb and S NP' or 'transitiveverb plus A and 0 NP's' as the 'core' of
a sentence. Any core may of course be augmented by peripheralcomponents: locational or
temporal qualifiers, adverbial specifications, NP's in dative case, etc. Peripheralcomponents
5 as a rule, occur with a core of either transitivity.(Fuller discussion of these points is
can, in
Dixon 1977a:401 if.)
6 Craig notes that, correspondingto 'He closed the door', we cannot have 'The wind closed
the door', involving the same transitiveverb speba 'close'; instead, a sentence translatableas
57 door closed by the wind' must be used, involvingxpehi 'closed' and with 'wind' expressed
'The
throughan agentiveprepositionalphrase.
Note here the inclusion of at (cf. He is eating the meat and He is eating at the meat.) We
meanings of close and eat require animate agency; but the physical action of the
wind can create the same impression as an animate agent, so that the windis clearly
regarded as an A NP in The wind closed the door. And a language-particularmeta-
phorical extension views the effects of sorrow as akin to 'eating' with respect to the
institutionalized symbol 'heart'. (Note that this is a fairly general metaphorical
extension, applying over a wide semantic field in English: e.g., I am consumedby
sorrow/withenvy; She was devouredby anxiety, etc.)
Certain transitive verbs occur in all languages: 'cut', 'throw', 'give', 'see', 'eat',
and a few score more. All these describe actions controllable by a human or animate
agent. But individual languages allow different types of semantic extensions from
the recurring 'central meaning'. Some extensions may retain the idea of animate
agency, but enlarge the class of actions the verb can refer to (e.g., We cut our losses);
other extensions may apply the verb to events that do not have a controlling agent,
when there is some culturally-perceived similarity to the central reference of the
verb (e.g. That interruptionthrew me off track; Rock music gives me a headache.)
Metaphorical uses of these verbs are always outnumbered by occurrences in the
'central meaning', where there is a human or other animate agent who could (and
most often does) control the activity.
Beyond this universal set of transitive verbs (and their more precise articulations
and hyponyms),58individual languages include further verbs in the transitive class
which have more-or-less idiosyncratic and language-particularmeanings. Most of
these will again demand an animate 'agent' in their central use. But for some, no
core NP need be animate (e.g. attract in A magnet will attract iron, Wealth attracts
robbers; Lyons, 359); here one NP is recognized to be in A function, through a
perceived similarity of this event to activities that are controllable (e.g. pull).
I am suggesting that all languages have a class of 'transitive verbs' whose
semantic effect is defined in terms of the universally occurring 'controllable' verbs
like 'cut' and 'give'. But verbs describing other activities may then also be included
in this class, with a participant recognized as being in A function because of
culturally-perceivedsimilaritiesto some variety of controllable event. Extensions of
the transitive class to essentially non-controllable events differ from language to
language (and could perhaps be taken as evidence for difference in Whorfianworld
view). Some languages have 'like' as a transitive verb (as in 'I like tea'); others
must use an intransitive or adjectival construction (something like 'Tea is likeable
could alternatively have Sorrow is eating my heart out, but scarcely *Sorrow is eating my heart.
This illustrates typical grammaticalrestrictionson metaphoricalextensions of common verbs.
I am grateful to W. S. Allen for drawing my attention to Homer's description(Iliad 6:202)
of Bellerophon 'eating his heart' (hon thumonkatedon)with sorrow; it may be significant that
he here uses the verb (with kata 'down') which is elsewheregenerallytranslatableas 'devour,
eat up' rather than just 'eat'.
58 Of course some
languages lack a single verb 'cut' or 'carry'. In Dyirbal one must choose
between nudi-l 'cut right through, sever' and gunba-l 'cut partway through, cut a piece out';
and in Indonesian it is usually necessaryto specify pikul 'carry on the shoulder', jinjing'carry
by the tips of the fingers', kepit 'carry under one's arms', galas 'carry with a carrying pole',
genggam'carry in fist or claws', or junjung'carry on the head', etc. This in no way affects my
argumentation.The point is that each language has one OR MOREverbs 'cut', 'carry' etc.; the
Dyirbal and Indonesianexamples could be taken to indicate a gap where a generalverb would
ERGATIVITY 107
to me.') English has annoy and endure as transitive verbs.59 These can take an
animate A NP, but the referent could not be said to control the activity in the way
that a 'cutter' or a 'carry-er' does. Note, though, that someone CAN purposely
annoy, and that endurance implies a measure of will-power; the A NP's for these
verbs do NOTinitiate or control the activity, but the role they play can be LIKENED
to that of an agent. Verbs of this nature tend to be language-specific,60and should
be regarded as idiosyncratic extensions to the universally occurring set of control-
lable verbs that make up the core of the class of transitiveverbs in every language.61
Intransitive verbs take a single obligatory NP which is in S function. With some
verbs (e.g. 'run', 'jump'), the referent of the S NP will be unequivocal agent,
controlling or initiating the activity (this was referredto as subtype Sa in ?2.2). For
other verbs (e.g. 'yawn', 'break'), the S NP is unlikely to be able to exercise any
measure of control (it can be termed SO).In ?3.1 I described fluid S-marking and
split S-marking languages, which sometimes treat an S NP like a transitive A
constituent and sometimes like 0. It seems, though, that no language carries this
distinction between Sa and SOinto its shallow syntax. All types of language treat S
NP's of all kinds in the same way for operations of coordination, subordination
and the like.
In fact, most languages that distinguish between 'active' and 'stative' intransitive
be expected, culturalreasons dictating more detailed specificationby use of a hyponym. (Note
that the Dyirbal 'mother-in-law' speech style does just have one verb dYalUga-l, whose central
meaning exactly correspondsto that of Eng. cut; cf. Dixon 1971).
59Schachter 1976, 1977 mentions the Tagalog verb -tiis 'endure' as evidence that an actor
nominal (the A NP for -tiis) is not necessarilythe 'perceived instigator of the action' (this is
Fillmore's criterion for 'agent', p. 24). It appears, however, that the arguments from this
section do apply to Tagalog: MOST membersof the class of transitiveverbs will have an animate
actor that satisfies Fillmore's criterion for 'agent' (which is a part of my criterion for A
function). Thus -tiis is just an extensional member of the transitiveclass in Tagalog, as endure
is in English(and as the examplesof fn. 60 are for YidinY).
Note that the definition of an A NP in terms of a participantwho 'initiates and/or controls
the activity' is not vitiated by odd verbs like endure; the test of any such generalizationis
whetherit describesthe majoritypatternof a language.Idiosyncraticverbs in any languagecan
be dealt with as institutionalizedextensions to the universaldefinition,or they can be dealt with
simply as 'exceptions' that have to be learnt by heart. (Exceptions are recognized as a valid
category in phonology and morphology; the idea is also applicable within syntax and even
within semantics.)
60 I have gathered around 300 verbs for YidinY,about 200 of them transitive. All but three
occur predominantlywith animate A NP's (although many have metaphoricalextensions, e.g.
'The fever is eating my body.') The three exceptions-which are the only transitive verbs that
do not occur in imperative form-are guba-n 'burn', whose A NP must be 'fire', 'sun', or
something burning (there is another verb wadYu-l 'burn, cook' which must have a human A
NP); dvadva-l'(sacredwater)rises against (someone who has broken a taboo)'; and wigi-l'(too
rich food) makes (a person) feel sick' (cf. Dixon 1977a:257-8).
61Some putative counter-examplesto the definition of the A function in terms of potential
agency demand a different explanation. Consider John underwent torture/an operation/an
examination(cf. Lakoff & Ross 1976:161). Here the underlyingsemantic representationcould
be taken as (Someone) tortured/operated on/examined John, from which the sentence with
underwentcan be derived by a passive-likeoperation(but note that, unlike passive, this DOES
change meaning). It appearsthat, in most instances of use, the 'object' of undergois a deverbal
nominal, as in the exampleshere.
verbs do not follow a strict semantic principle. There are generally some volitional
verbs in the stative class (whose S takes O-type marking), and a few non-volitional
items in the active class (S taking A-type marking): see, e.g., the sketch of Hidatsa
in ?3.1 (and Van Valin on Lakhota). We cannot precisely identify Sa and SO-
defined on semantic grounds as agentive and non-agentive S-with the two
morphological subclasses of intransitive verbs in split S-marking languages. And
even in fluid S-marking languages, the distinction between Sa and SOdoes not
continue into the syntax (or, often, into other parts of the morphology). Thus
Choctaw cross-referencesthe S NP by prefixes typical of an A NP, O NP, or dative
NP, depending on the verb involved; but it does treat all S NP's in the same
manner for nominal case-marking and for purposes of syntactic rules (see ?3.1,
above, and Heath 1977).
We may recognize S as the third basic function, defined simply as the only
obligatory NP in an intransitive sentence. S can, as a later step, be subclassifiedinto
agentive Sa and non-agentive SO,4 Sa and SO,help explain some morphological splits,
but this distinction plays virtually no role in syntax.62
It will be seen that the universal syntactic-semantic functions A, S, and O are
defined on rather different principles: the ONLYOBLIGATORY NP in an INTRANSITIVE
clause, for S; the NP in a TRANSITIVE clause which CAN BE AGENT,for A; and the
OTHER OBLIGATORYNP in a TRANSITIVEclause, for 0. These functions appear to be
valid for all natural languages and to be the basis for all grammatical operations.
We surveyed in ?3 the possible splits and variations in morphological marking; these
can all be explained in terms of A, S, and 0, sometimes referring also to the
semantic content of an NP or of a verb, or to the aspect or tense of a sentence.
We can now define a further universal category: 'subject'. A and S functions are
grouped together as 'subject .63 These are the NP's which refer to participantsthat
can be the initiating/controlling agents. There is a difference: an A NP almost
always has the potentiality of being agent for any transitive verb (the 'almost' may
be omittable for languages like Jacaltec which limit A NP's to animates); and an S
NP could conceivably be agent only for certain verbs. Ideally, we should define
' subject' as {A,Sa}, linking A with the subtype of S which can be agent. But we have
noted that no language consistently distinguishes Sa from SOin all aspects of its
grammar. It is the category S that functions as a syntactic prime, in the syntax of
62
Although the extremesof Saverbs and SOverbs are clear enough, there is a large fuzzy area
in between, where SOMEMEASUREof agency is possible (see ?2.2). This must be one factor in the
failure of most 'split S' languagesto maintainan exact semantic basis, and in the treatmentof
all S NP's in the same way in syntactic rules.
63 The status of
A, S, and 0 as universalprimitives,having priority over the recognition of
'subject', is vital to the thesis of this paper. Some linguists use symbols like St, S5,and 0 (or P
'patient'). But this suggeststhat 'subject' is the FIRSTcategoryto be recognized-and that it can
then be subclassifiedinto transitiveand intransitivevarieties, according to the sentence type it
occurs in. The use of such symbolisms(see Anderson 1977, Heath 1976a, and Woodbury 1975,
1977) has misleading implications. The symbols A, S, and 0 used here (the choice of letters is
immaterial-the main point is to choose different symbols for different primitive functions)
emphasizes the syntactic and semanticDIFFERENCESamong these threefunctions; once these are
established, A and S can as a next step be grouped into the category 'subject' on the basis of
PARTIAL SIMILARITIES.
ERGATIVITY 109
everylanguage.ThesemanticlinkbetweenA andthe subtypeSa of S is generalized,

as it were,so that A and S are groupedtogetherto makeup the universalsyntactic-
semanticcategory'subject'.No matterthatnot everyintransitiveverbcanplausibly
appearin the (positive)imperative;everyS NP countsas a 'subject'.64
Although I have used the semanticcriterionof 'potential agency' to define
'subject',this is a categorythat plays an importantrole in the grammarof every
language.Certainuniversalsyntacticphenomenafollow from the propertiesof
'subject'.We will surveysome of these in ?5.4,afterfirstreviewingotherattempts
to clarifythe notionof'subject'.
Throughoutthis section,I have beentalkingin termsof the 'deepstructures'of
sentences,65and I have defined'subject' as a universaldeep-structurecategory.
Some linguistshave also used the notion 'surfacesubject';e.g., they may say that
the passivetransformation places'deep object'in 'surfacesubject'function.I shall
show in ?6.2that, while'surfacesubject'maybe definablefor some languages,it is
by no means a workableuniversalcategory.It is useful and valid to speak of
'derivedA/S/O' functionalslots, but not in the sameway of'derived subject'.As
a universalsemantic-syntacticcategory, 'subject' must be definedat the deep-
structurelevel, and alwaysrelatedto that level.
Everylanguagecould be said to have a measureof 'accusativity'at the level of
deep structure,in that the universalcategoryof 'subject'-a groupingtogetherof
A and S-plays some role in its syntax.In the same way, everylanguagemingles
accusativitywithergativityin the structureof its lexicon(compareHe ate it/He ate
with He broke it/It broke; ?2.2). But it really makes little sense to attemptto
characterizean individuallanguageas 'accusative'or 'ergative'at this level.Deep
structuredeals with the way in which syntax codes the semanticdescriptionof
events.Thereare threebasicsyntactic-semantic categories-A, S, and 0; theseare
true universals,beingapplicableto everytype of sentencein everylanguage.Then
A and S are groupedtogetheras 'subject',a deep-structure categoryby virtueof
which certain universalsyntacticphenomenafollow. It is only at the level of
'shallowstructure',afteroperationslike passiveand antipassivehaveapplied,that
languagesdifferin the way they group syntacticfunctions;it is at this level that
typological characterizationsof languages as being syntactically'ergative' or
'accusative'have significance.We returnto this topic in ??6-7.
64 Our discussion thus far has focused on sentences containing a lexical verb. Of course,
there are also sentences involving 'have' and 'be' (for languagesthat have a copula)-and, in
many languages, minor sentencesthat involve no verb at all but just, say, a noun as topic and
an adjectiveas comment. Now it is a fact that a conceptwhich is dealt with througha verb in one
language may be renderedby an adjective in another (Dixon 1977d); this suggests that, in any
universalcategorization, the function S should be extended to apply to the topic of adjectival
comment sentences and to the 'subjects' of 'to be' and 'to have'. I mentioned in fn. 15 that it
is always the S form of a nominal or pronominalwhich is used in minor sentencetypes like these.
65'Deep structure'is used in a fairlyinformalway throughoutthis paper. It is not necessarily
intended in the sense of Chomsky 1965, which definesa grammaticallevel largelyin terms of its
syntactic usefulness; it is perhaps closer to the 'semantic representation' of Generative
Semantics. The notion is purposely left vague here; detailed research into the syntax and
semantics of a wide sample of languages, within the terms suggested in this paper, is a pre-
requisite to a full characterizationof this (surely universal)level.
5.3. KEENAN'S DISCUSSION OF'SUBJECT'. In an important paper, Keenan seeks 'to

provide a definition of the notion "subject of" which will enable us to identify the
subject phrase(s), if any, of any sentence in any language' (1976:305). The moti-
vation for this attempt is the reliance on 'subject' in a number of currenttheoretical
enquiries (e.g. the Accessibility Hierarchy of Keenan & Comrie, and the Relational
Grammar of Perlmutter & Postal).
Keenan's procedure seems to be to survey the properties of what have been
recognized as 'subjects' in a wide selection of languages. If the linguists working
on a language have not used the term 'subject', then Keenan brings to bear his
own criteria, which tend to emphasize surface grammaticalfactors (cf. ?5.1, above).
He presents a list of 30-odd properties characteristically possessed by subjects.
Almost all the properties are qualified by 'usually', 'normally', or 'in general';
there is no attempt at a universal 'definition' (in the logical sense of 'Every A
which shows X is a subject'); rather, 'what is subject' is based on a statistical
assessment of which NP satisfies the largest number of the 30-odd properties.
Keenan gives a miscellany of syntactic and semantic 'properties' with no
priority among them (so that the whole argument runs the risk of being circular).
Some are characteristicsof shallow structures;some involve morphological marking
conventions; and some follow from universal semantic arrangements. As an
example of the last point, Keenan mentions that the entity to which a subject refers
'exists independently of the action or property expressed by the predicate. This is
less true for non-subjects.' (He cites as example A student wrote a poem.) But this
is surely a consequence of the fact that the A NP tends to come further to the left
than the 0 NP in Silverstein's hierarchy(see ?3.2, above); abstract NP's like a poem
belong at the right of the hierarchy.66
Keenan does include, among his 30-odd properties, that 'Subjects normally
express the agent of the action, if there is one'; this is the universal defining
criterion which I adopted for 'subject' above (but note Keenan's use of 'normally').
In Keenan's presentation, this semantic criterion follows a number of surface
grammatical properties with which it appears to be accorded equal weighting: the
subject is usually indispensable(i.e. non-deletable); the subject is usuallythe leftmost
NP; if anything has zero marking, it will be the subject of an intransitive verb-
and so on.
In ?2.3,I showed the variety of types of case-marking that occur. That NP which
is in the unmarked syntactic case is the most likely candidate to be' indispensable'-
but the unmarked case can be (a) absolutive, covering S and 0 functions; or (b)
nominative, covering S and A functions; or (c) accusative, with just 0 function. The
property that 'Subject is usually indispensable' follows from the facts that the
majority of languages are of type (b), and that the NP in unmarked nominative
case is not normally deletable. The propertythat' If anything has zero case-marking,
it will be intransitive subject' follows from the facts that the great majority of
languages are of types (a) or (b), and that the unmarked case is likely to have zero
66 Keenan intends this not as an absolute rule, but as a tendency-i.e.
'subject is more likely
to have independentexistence than object.' It is of course not hard to find odd exampleswhere
the A NP does not have independent existence: e.g., His wife married him on the rebound; My
new book pretty well wrote itself.
ERGATIVITY Ill
realization, if anything does. (Note that this last property says nothing about the
category of 'subject', but just about S-one of the two basic functions grouped
together to form 'subject' under my definition).
I have already mentioned that 'subject' has been used by linguists in a variety of
ways. It is illegitimate to compare the properties of 'subjects' without first con-
sidering the criteria used for recognizing 'subject' in each grammar. Plainly, there
is no uniquely correct notion of 'subject'; the category has to be carefully defined
in any universal or language-particular enquiry. In ?5.2, I took A, S, and O as
universal functions, in terms of which all grammatical phenomena may be de-
scribed; I then DEFINED subject as the class {A,S}: thus every sentence will have a
subject. For transitive sentences, A is distinguished from O in terms of potential
agency, the criterion that is later taken to underlie 'subject'. Intransitive sentences
have only one obligatory NP (which I label as in S function), and so this is linked
with A as 'subject', whether or not it could be 'agent' for any particularverb.
Certain of Keenan's properties automatically follow from my definition of
subject. He notes that 'Subjects normally express the addressee phrase of impera-
tives'; if subject is defined as (potential) agent, the addresseephrase of an imperative
MUST always be subject.67The 'normally' in 'subject = agent' and the 'normally'
in 'subject = addressee of imperative' are linked (although this is not noted by
Keenan) in that we must have 'addressee of imperative = agent'.
Most of Keenan's criteria effectively define the shallow-structure category
'pivot' which I discuss in ?6.2. 'Pivot' is a language-particularcategory: in some
languages it links (derived) S and A, in others S and 0-and there are languages
which employ both types. But Keenan also includes some criteria that relate to
'deep subject', the universal category {S,A}. There can be serious conflict between
'subject' and 'pivot' in the most ergative languages. For Dyirbal, the S and 0
NP's (marked by absolutive case, on nominals) show more of the 30-odd properties
-e.g. indispensable, leftmost, zero marking, syntactic pivot-than do S and A;
thus Keenan takes {S,0} to be 'subject',68 letting grammatical criteria override
semantic considerations. He has effectively recognized that S/0 is the shallow-
structure pivot in Dyirbal syntax. But with this definition of 'subject', Dyirbal must
be noted as an exception to 'subject = agent', and it is also an exception to
'subject = addressee of imperative'.
As Blake has shown, there can be even more serious difficulties attached to
defining 'subject' on purely grammatical criteria. Pronouns in Dyirbal show a
nominative/accusative paradigm, and it is the 'nominative' which is leftmost,
unmarked, and non-deletable (though it is not the syntactic pivot; see ?1.1). So
67 Keenan's reference to Maori and Malagasy as counter-examples
indicates confusion
between'deep subject' and'surface subject'; see ?5.41, below.
68 Postal (1977:278) mentions'an analysis of ergativity phenomena
which takes "patient"
nominals to be initial subjects of transitiveclauses ... the analysis of Dyirbal in Dixon (1972:
128-30) is of this type ... 'In fact, 'subject' is used in a semanticsense throughout the grammar
of Dyirbal, never in the 'grammatical'sense suggestedby Postal. Tree structuresof an'ergative
type' are used, but they do not imply that the notion of'subject' is differentin Dyirbal from
any other language. The tree structuresmake syntactic, not semantic, claims. The notion of
'subject' is effectively defined through the feature [+actorl (Dixon 1972:199-205); this is
needed to deal with the 'accusative' syntactic properties of Dyirbal (see ?5.4, below).
the label 'subject' should, on this type of criteria, be applied to a pronoun in S or A

function, and also to a noun in S or O function. A sentence with pronominal A and
nominal O, like 33 (cf. ?1.1) would have two subjects, whereas a sentence with
nominal A and pronominal 0, like 34, would have none!
(33) yana yuma bura+n 'We saw father.'
we-NOM father-ABS see + PAST
(34) yana+ na yuma+-gu bura+ n 'Father saw us.'
we + ACC father + ERG see + PAST
Further difficultieswith this type of approach are discussed by Blake.
5.4. UNIVERSAL SYNTACTIC PHENOMENADEPENDENT ON 'SUBJECT'. Classifying a
language as 'ergative' in terms of morphological marking is a relatively straight-
forward matter when compared with making a decision between 'ergativity' (S
treated in the same way as 0) and 'accusativity' (S treated like A) at the syntactic
level. Different kinds of syntactic evidence can be brought to bear, and these by no
means all give the same result. It is, in fact, necessary to distinguish among (a)
universal syntactic behavior, that recurs in all languages-or which has the same
form in every language in which it does occur; (b) transformations needed to place
an NP in 'pivotal function'-usually corresponding to unmarked case-for a
variety of syntactic and discourse purposes; and (c) language-particularoperations
that provide genuine evidence for syntactic 'ergativity'. Possibilities (b) and (c) are
interrelated, and are discussed in ??6-7.
The syntactic operations that involve the SAME identifications, among A, S, and
0-in EVERY language in which they occur-as in (a) above, are all corollaries of
the universal category of subject. I will now give three critical examples of such
operations, and then discuss 'causatives' which have a universal basis that is
dependent upon A, not on {A,S}
5.41. IMPERATIVES. In an imperative sentence, the speaker requests the addressee
to do something-to act as agent in initiating/controllingsome activity. Imperatives
in every language have a 2nd person pronoun as (stated or understood) S or A
NP.69 This is a universal property; thus the fact that S and A have the same
possibilities of reference for the imperative constructions of some particular
language (and the fact that, say, either can be deleted from surface structure) is no
evidence at all for the placement of that language on a continuum of syntactic
'ergativity' vs. 'accusativity'. Even the most ergative language will treat S and A
NP's of imperatives the same. This follows from the meaning of imperatives
(addressee is told to be agent) and the definition of 'subject' (the NP whose
referent can be agent, if anything can).70
69 Most languages restrict the subjects of imperatives to 2nd person pronouns. In a few
languages, there is extension to 1st or even to 3rd person subjects (although these are always
greatly outnumberedby 2nd person subjects).The discussionhere can naturallybe extended to
these additional cases (e.g., lst/3rd person possibilities always apply equally to A and to S
structuralslots).
70 Generally, an imperativewill have a 2nd person pronoun as
the S or A NP in both deep
and surface structure; this condition is ideally satisfied in reflexives.Passives cannot normally
occur in imperativeform. It is possible to devise examples of the type (i) Come to Palm Court
and be entertained by Joe Loss and his Orchestra! and (ii) Be impressed by his stamp collection
ERGATIVITY 113
Whether any particular verb can occur in imperative form (marked by special
inflection in many languages, and almost always characterized by distinctive
intonation) depends upon its semantic type-whether it describes a state or activity
that is controllable. Most (though not all) transitive verbs are controllable, but
only that subset of intransitiveswhich take an Sa NP fall into this category. Indeed,
some languages that show split S-marking permit imperatives only of Sa verbs, not
of the SOset. Such a restriction is explicitly stated by Gregores & Suarez for
Guarani (see ?3.1, above); it is an isolated instance where the Sa/So distinction has
some syntactic relevance.
Sa and SOare not distinguished for later syntactic rules such as coordination and
subordination, in any language; 'S' is treated as a single homogeneous category.
Similarly, all types of S are linked with A as being POTENTIALLY the addressee of an
imperative, in almost every language (Guarani being a fairly rare exception).
Imperatives usually involve verbs that demand a fair level of control; but the
grammatical construction can be extended to verbs where the level of control is
minimal or non-existent. We can conceive of Endureitfor afew weeks longer(and I'll
arrangea transfer)!or a whispered wish Slip downthereand breakyour leg! Negative
imperatives are more plausiblewith barelycontrollable verbs-Don't yawn!-though
even here the limits of possibility can be crossed, e.g. Don't die! Since some A and
S NP's function naturally as addressees of imperatives, this property is potentially
extendable to all members of the grammatical classes covered by A and S.71
It is of course possible that, IN ADDITION
TO this universal S/A linkage, imperatives
in particular languages may also in some way treat S and 0 alike. In Nass-Gitksan
if you wanthim to like you! But note that, for these to be acceptable,the passive must be linked
to another clause with which it sharesa subjectNP. The acceptabilityof these sentencesappears
to stem from formal analogy to imperativecopular sentences(e.g. Be satisfiedwithit!; Be happy
to go!-see fn. 71), and from the demands of discoursestructure.It is noteworthy that ex. (i) is
recognizable as a compelling advertisingslogan; most speakers would prefer to use Come to
Palm Court and let Joe Loss entertain you! (the addressee has no control over whether he is
entertainedor not, only in whetherhe allows Joe Loss to try to entertainhim). Ex (ii) is felt to
be an elliptical version of Try to be impressed ... ! or Appear to be impressed ... !
For those linguists who accept that the deep structureof Appearto be asleep! has [You be
asleep] as subject, and that the deep structureof Be easy to please! has [PRO to please you] as
subject, there are compelling reasons to specify shallow structureas the level of application of
the requirementthat the (here, derived) S or A NP of an imperative should be 2nd person.
I would not accept these analyses, for reasons which it would be inappropriateto explore here.
(Rodney Huddleston and Geoff Pullum provided these examples and useful discussion of this
topic; they do not agree with the position adopted here.)
71 The principle that a grammatical property which is semantically plausible for some
membersof a grammaticallydefinedclass CAN be extendedto all membersof the class, essentially
going beyond the limits of semantic plausibility (and then producing sentences which may be
semantically bizarre but grammatically acceptable), can be illustrated from the difference
between languages that include a copula in adjectival predications and those that do not.
Languages like English, of the former type, can use imperatives like Be happy!, Be thin!, Be
hungry!-which, although not 'normal', can be contextualized,e.g. Be thin! 'Pull your stomach
in!', Be hungry!'Act hungry!' Adjectival sentences in Dyirbal involve no verbal form, and so
CANNOT be marked as grammatical imperatives (which in this language are indicated by
distinct verbal inflections). Note the gradual restriction on imperativesfrom English through
Dyirbal to Guarani, whereonly Saverbscan form imperatives.
(Rigsby), one pronominal suffix cross-references S or 0, and another suffix cross-

references A-basically an absolutive/ergative pattern. But in an imperative
construction, just the A suffix, not S orO forms, must be deleted. Thus, for Nass-
Gitksan imperatives, both S and A must be the 2nd person (the universal rule);
but A has to be deleted and S retained (this is a fact about the structure of verbs in
Nass-Gitksan).
Sometimes an obligatory transformation must apply to an imperative. Keenan
(321) gives, as an exception to his property that 'Subjects normally express the
addressee phrase of imperatives', the fact that 'in many Malayo-Polynesian lan-
guages, e.g. Maori and Malagasy, imperatives are frequently in non-active forms,
and the addressee phrase, if present, appears as a passive (or other type of non-
active) agent phrase'. Keenan is here noting the non-application of a semantic
criterion, which is valid for deep subjects, to derived 'surface subjects' (on this,
see ?6.2, below). Maori and Malagasy conform perfectly to the universal pattern;
it is just that the passive transformation, which is optional for other construction
types, is obligatory or nearly so for imperatives,72so that the (deep) subject of an
imperative is always realized with oblique marking.73
5.42. JUSSIVE COMPLEMENTS. A number of construction types, occurring in all or
almost all languages, are semantically close to imperatives. In many languages,
jussive complement constructions-with a main-clause verb like 'tell' or 'order'-
could be regarded as indirect imperatives; examples from English would be I
orderedhim to go; I told him to bring the water. Here the object of the main-clause
verb must be, at the level of deep structure, coreferential with the S or A NP of
the verb in the subordinate clause. The subordinate clause describes an instruction
that has been given to someone to do something: plainly this participant must be
agent, and thus subject, for the subordinate clause.74
Since this requirement on identification of S and A NP's in jussive complement
constructions is a universal consequence of the definition of subject in terms of
agency, and of the meanings of jussives, it cannot be taken as evidence for the
syntactic type of any individual language. In an early unpublished paper showing
that Walbiri, despite some morphological ergativity, is basically 'accusative' at
72 For cultural reasons of 'politeness' etc. (see Keenan & Keenan, MS).
73 Malagasy, Maori, and English all satisfy the universalrequirementthat 'deep subject' of
an imperativebe 2nd person. They differin that English also requiresthis NP to come through
into surface structurein S or A function (*Mary be watchedby you is unacceptable),whereas
Malagasy and Maori do not impose this extra condition. In an ergative languagelike Dyirbal,
the same universalrequirementholds. But here an imperativecan, optionally, be antipassivized
(cf. ?1.1):
(n"ura) yabu bura 'You look at mother!'
(n9ura) bural+oa yabu+gu 'You look at mother!'
It is impossible to decide whether Dyirbal should be classified with Malagasy and Maori or
with English. Whereasa passive transformationplaces a deep 0 NP in surface S function, the
antipassive places a deep A in surface S function. Since A and S are subject functions, an
antipassive imperative will, on almost any definition of 'subject', NECESSARILYstill have the
addresseephrasein subjectfunction.
74For a useful discussion of jussives in Rembarnga (an Australian language) and their
universalsyntacticform, see McKay.
ERGATIVITY 115
the syntactic level, Hale (1968a:36-7) mentions that 'the well-formedness of the
jussive complements also requires that the subject of the embedding be identical to
the object of the matrix'. But this is so for all languages; it certainly holds, at the
level of deep structure, for jussives in Dyirbal. (But the 'ergative syntax' of Dyirbal
does demand a different arrangement in surface structure; see ?7.2.)
5.43. 'CAN' AND SIMILARVERBS.In certain languages, some or all of the group
'begin', 'finish', 'can', 'must', and 'try' are inflected like lexical verbs; in other
languages, they are verbal auxiliaries-separate words that must occur in con-
junction with a lexical verb; and in still another group of languages, they are
realized as derivational affixes on verbs. Whatever their surface grammaticalstatus,
these items are essentially modifiers of a lexical verb; when they have some of the
surface characteristics of main verbs, they must share the subject of a lexical verb
(and sometimes object as well). Thus, in sentences like I began to paint the wall and
I began to laugh,the subject of begin is identical to the A or S NP of the complement
clause; but this does not constitute evidence for English having an accusative-type
syntax. This property follows from the semantics of begin, and is shown by every
language that has a distinct verb 'begin'.
Great care must be taken to ensure that constructions of this type-in which A
and S are treated alike, as a natural corollary of the universal definition of subject
(the auxiliary element commenting on the subject's 'control' of an event)-are not
taken as evidence about the syntactic accusativity of a language. Thus Woodbury
(1975:66-70) mentions, among other pieces of evidence, that Eskimo has con-
structions of the form NP1 V1 [SJ], where V1 is one of a restricted class of verbs
that includes 'can', 'must', 'begin', and 'want'. Sentences of this form have a
syntactic constraint that NP1 must be coreferential with the A or S NP of the
embedded clause 22; Equi-NP Deletion then takes place. But 'can', 'must' and
'begin' always behave in this way, and 'want' almost always does.75 This syntactic
constraint is a natural consequence of the meanings of V1 and the universal
definition of subject; it cannot be taken as any indication of the syntactic type of
Eskimo.76
75
'Want' varies in its behavior from language to language, in a way that 'can', 'begin', and
'must' do not; and its behavior in Eskimo and Khinalug could be taken as very weak evidence
for syntactic accusativity. There are languages where 'want' appears to behave as a straight-
forwardmodifierto a lexical verb, on a par with 'can'. But in other languagesthere need be no
NP in common to the main and complement clauses; Comrie points out that there is a literal
translation of 'I want John to hit Bill' in Khinalug (see also Anderson 1976:8).
76 Woodbury (1975:118-19, 131) recognizes that 'the accusativity of
EQUI can be best
explained in terms of the semantic class of EQUI type verbs, a subclass of which requires the
"like-subject" constraint discussed in Perlmutter 1971, which limits certain verbs and their
complement clauses to coreferentialsubjects, e.g. Eng. begin, try, can etc.' Recognizing that
coreferentialityand deletion in Dyirbal depend on the S/O pivot, Woodbury then mentions
that 'it would be interestingto see whether there are verbs in Dyirbal to which the like-subject
constraint applies.'
If there were any verbs 'must', 'can', 'try' in Dyirbal, then-as a consequence of their
semantic import and the universalcategory of subject-they would have to involve 'deep S/A'
identification. It would not be impossible to accommodate this within Dyirbal, or for it to be
consistent with the 'S/O pivot' shallow structureof Dyirbal (see the discussion of jussives in
There are other examples of this nature in the literature on 'ergativity'. Comrie
(MSa) mentions that many languages which are ergative at the morphological level
are accusative in their syntax. Exemplifying for Khinalug, a northeast Caucasian
language, he simply quotes constructions with 'want to V' and 'can V'; but these
constructions would be EXPECTED to equate A and S, as a corollary of the universal
definition of subject.
Chung 1976 argues for an 'accusative' syntax in modern Polynesian languages,
entirely on the basis of a handful of verbs-'can', 'begin', 'must'-whose subject
must coincide with the subject of the complement clause, triggering a 'raising
rule' which applies only for these verbs (see also Anderson 1976:13). This is
insufficientbasis for typological classification of the syntax of Polynesian languages,
let alone as a major step in the argument that Proto-Polynesian had accusative
morphology and syntax.
I shall show in ?7 that there are syntactic operations whose identificationswithin
the set A, S, O do vary from language to language; these operations enable us to
place languages along a typological continuum ranging from 'syntactically
ergative' to 'syntactically accusative'. But we must be careful to distinguish them
from universal syntactic phenomena of the type described here-which always
equate S and A-as a consequence of the universal category of 'subject' and its
semantic implications.
5.44. CAUSATIVES. One of Keenan's properties (321) is that 'Subjects normally
exhibit the same position, case-marking, and verb agreements as does the causer
NP in the most basic type of causative sentence.' Now a construction type is
recognized as 'causative' partly on the semantic grounds that the referent of the
'causer NP' MAKESTHEEVENTHAPPEN.We saw in ?5.2 how all languages assign one
syntactic-semantic function (that we are calling A) to that NP in a transitive
sentence which could be 'agent'. The causer must plainly be the A NP (one of the
two basic functions covered by' subject').
Many languages have a productive mechanism for deriving a transitive causative
verb from an intransitive verb (The door opened-> The man opened the door) or
from an adjective (The wall is black -> I blackenedthe wall). In addition, a number
of pairs of lexical roots usually show the same relation (e.g. in English,fell = make
fall, kill = make dead).77Here the S NP of the intransitive sentence (e.g. The tree
?7.2). But in fact Dyirbal does not have lexical verbs of this type: 'try' is renderedby a non-
inflecting particle, while 'must' and 'can' are expressedby the use of purposiveinflectionon a
main-clause verb. There is an adverbial 'begin'-an item that takes the same inflection as a
verb, and occurs with a lexical verb in a 'verb complex' (just as an adjectiveoccurs with noun
in an NP); thus, 'begin' naturally has the same A and 0 (or S) as the main verb it modifies
(Dixon 1972:67-9, 119, 301-2).
77 Sapir (1917:84) suggested that kill be related to cause to die, and this has been repeated
many times since. In fact cause is a quite uncommon verb in English, with fairly restricted
meaning; causatives are more revealingly dealt with in terms of the very common verb make.
Make takes an adjectival or participial complement, not a to-clause, and analyses such as
blacken = make black, kill = make dead are a considerable improvement on cause to become
black and cause to die (or cause to become dead). For instance, none of Fodor's arguments
(1970) against a derivationof kill from cause to die apply against make dead.
ERGATIVITY 117
is falling) corresponds to the O NP of the corresponding causative (The woodmanis

felling the tree); this has been taken as evidence for 'ergative syntax'.
Keenan's comment relates not to 'subject' but to A function: the causer NP in
a causative construction is in A function. For lexical verb pairs (fall/fell), there is
an underlying semantic relation between intransitive and transitive construction.
Various proposals have been put forwardfor derivedcausatives, e.g. that I blackened
the wall is derived from I CAUS+ PASTthe wall and The wall (be) black; the under-
lying A NP can, of course, be the addressee of an imperative. But the identification
of 'S = 0' and 'causer = A' is a universal feature of causative constructions; it
does not provide evidence as to the syntactic 'ergativity' or 'accusativity' of any
language, and it does not relate directly to the universal category of 'subject'
(defined as a class consisting of the primitives A and S).
Note, though, that some languages have more complex and more interesting
syntactic phenomena connected with 'causatives'. First we can remind ourselves
(cf. ?2.2) that, although transitive/intransitive pairs in some areas of the lexicon
show S = O identification ('lexical ergativity'), those in other parts have S = A
identification-e.g. He is talking, He is telling me ('lexical accusativity').78
Now the productive process of forming transitive stems from intransitive roots
(with some explicit morphological marking) is syntactically homogeneous in many
languages, corresponding to the 'causative' S = O type I have been discussing-
e.g. Swahili (Loogman 1965:104ff.) and Turkish (Lewis 1953:106ff.) But in other
languages the transitivizing process may be of type S = 0 for some intransitive
verbs, but S = A for others: in YidinY,wanda-n is 'fall down' and wanda-ya-l is
'make fall over (e.g. by pushing)', whereas badi-n is 'cry' and badi-ya-l is 'cry over
(e.g. a lost child)'.79 The derivational suffix -ya-l always signals a transitive verb;
but in some cases it is genuinely causative (S = 0), and in other cases it signals that
the intransitive S NP has become A, and that an originally oblique NP (in, say,
dative case) from the intransitive sentence has moved into 0 function in the
derived transitive clause (Dixon 1977a:302-22).
There is in fact some semantic explanation for which verb behaves in which way.
Lexical pairs of type S = 0 are mostly found among verbs of rest or motion
('stand' vs. 'make stand', 'come out' vs. 'take out'; Dixon 1972:297), with S = A
pairs being predominantly found in other semantic areas-'eat (transitive)' vs. 'eat
intransitive)' etc. If a derivational affix has varying effect, it will most frequently
be of type S = 0 for verbs of rest or motion, and of type S = A for other semantic
classes of verb.80Rembarnga, of Arnhem Land, shows an interestingrationalization
of this tendency: here the derivational suffix-wa- derives S = A transitives from
intransitive roots of conjugations 1-2 (kaluk 'play', kaluk-wa 'play with'; wak
'laugh', wak-wa 'laugh at' etc.), but causative S = 0 forms from roots of
78 Examples from Dyirbal, where the intransitiveand transitivemembers of each S = 0 and
S = A pair are totally differentin form, are given in Dixon 1972:297 and 1977c.
79
'Make cry' can only be renderedby a two-clause sentence, normally specifying what was
done to bring on the tears-e.g., 'The man teased me and I cried' (Dixon 1977a:313-14).
80 This is basicallythe situation in YidinY,but here there is no strict categorization:transitive
forms of SOME intransitive verb roots can be understood in either S = 0 or S = A sense,
depending on context. (A full account is in Dixon 1977a:302-22.)
conjugations 4-6 (pariyana 'be hanging up', pariyana-P-wa 'hang-up' etc.)81 The
important point is that most (but not quite all) of the intransitive verbs in con-
jugations 4-6 are concerned with motion or rest; but in fact -wa- has causative
sense even with those verbs from these conjugations that belong to another
semantic domain (e.g. 'yawn' -* 'make yawn'). Which kind of syntactic effect the
derivational affix -wa- has with any verb may have originally been semantically
conditioned in Rembarnga, as in YidinY;but this process has been grammaticalized
in terms of the predominant semantic field associated with each conjugation, so
that it is now conjugationally determined. (Rembarnga data are from McKay 1975
and p.c.)
The point of these examples is to stress that 'causative' can be one sense of a
more general transitivizing process, and that whether it is S = A or S = O can be
semantically conditioned; this is surely reminiscent of the semantic conditioning of
morphological splits, exemplified at length in ?2. Languages with a transitivizing
process that is always causative in nature could be said to have generalized from
one semantic type to all intransitive verbs (as Rembarnga appears to have general-
ized in a more restrictive way, to all intransitive verbs in conjugations 4-6).82
Data of this sort are surely not indicative of whether a language is 'ergative' or
'accusative' at the syntactic level.
5.45. We have seen that certain constructions-imperatives, jussives, verbs like
'can' -must involve identification of S and A at the deep syntactic level, purely
because of their semantic content and the semantic natures of S and A functions
(it is these semantic factors that lead to the grouping of S and A as the universal
deep-structure category 'subject'). There do not appear, in the same way, to be
any universal phenomena that must link S and0.83 All languages work in terms of
'subject'. For some languages, almost every syntactic operation appears to be
based on this category. But there are languages which require 'subject' only for
the type of syntactic operation exemplified here. In other areas, they may work in
terms of S and 0, rather than S and A. This will be taken up again in ?7.
DERIVEDSTRUCTURES
6.1. PASSIVE AND ANTIPASSIVE-FORMAND FUNCTION.Like 'ergativity' and
'subject', the terms 'passive' and 'antipassive' have been employed with a range of
81 There are no intransitiveroots in conjugations 3 and 7.
82 Dyirbal has a derivationalaffix -ma-l -mba-l that derives a transitiveverbal stem from
an intransitive root; it is uniformly S = A in effect, e.g. dYana-y 'stand' -> dYanayma-l 'stand
with/on'. But there is also a derivationalaffix -ma-l - -m(b)a-lthat derives a transitiveverbal
stem from an adjective; this DOES have causative meaning, e.g. guyi 'dead' -* guyima-l 'kill'.
These facts are particularlyinterestingin view of the'ergative' syntax of Dyirbal (?7.2).
83 Heath 1976a, McKay 1975, and Dixon 1977c have suggested that some syntactic
phenom-
ena would involve the same identificationfrom S, A, and 0 for all languagesin which they
occurred, implying S = A sometimesand S = 0 other times. It has proved difficult,however,
to find examples of universalS = 0 phenomena. The best that can be cited is noun incorpor-
ation, which involves S or 0 more often than it does A NP's. But this is a matter of word
formation-surely a phenomenon of a quite differenttype from those discussedhere. And it is
not so cut-and-dried:Comrie (MSa) mentions that 0 NP's are easiest to incorporate, followed
byS, with A most resistant to incorporation.
ERGATIVITY 119
meanings. It is thus necessary to explain the sense in which I employ these labels
in the present discussion. Both passive and antipassive are taken to be syntactic
derivations (transformations) that derive an intransitive sentence,84at what we can
call 'shallow structure' level, from an underlying transitive sentence:
(a) PASSIVEplaces the deep O NP in surface S function, and marks the deep A
NP with an oblique case/preposition/etc. (this NP can then be deleted).
(b) ANTIPASSIVE places the deep A NP in surface S function, and marks the deep
O NP with an oblique caselpreposition/etc. (this NP can then be deleted).
It will be seen that passive and antipassive are not mutually exclusive; both
could occur in a single language (thus Woodbury 1975:26-7, 88, 1977:322-5
reports both for Eskimo). But passive is normally encountered in languages with
nominative/accusative marking; one major use is to bring an O NP (which would
otherwise be in marked accusative case) into a surface function where it bears the
unmarked nominative inflection. Similarly, antipassive occurs in languages with a
predominantly ergative morphology, and brings an A NP (that would otherwise
bear the marked ergative inflection) into surface S function, where it is in unmarked
absolutive case.85
It is, of course, putting the cart before the horse to explain that, e.g., anti-
passivization is needed to put an A NP in the unmarkedcase. Case-markingis plainly
a late rule of the grammar, applying to surface structures after all syntactic
operations (including passive/antipassive) have applied. Passivization (/anti-
passivization) is largely motivated by a need to place a deep 0(/A) NP in derived
S function, in order to meet certain syntactic requirements(and/or preferences of
discourse structuring; cf. Kalmar 1976). Now if 0(/A) does not satisfy some
syntactic constraint, then in most cases A(/0) does, and in addition S will do so.
It is the class of NP's that can be 'syntactic pivot', in this way, which generally
takes the unmarked case inflection. It is thus generally true (but as a conclusion,
not as a premiss) that passive operates in languages that are morphologically and
syntactically nominative/accusative, and that antipassive will be found predomin-
antly in languages that have some measure of ergativity at the syntactic and
morphological levels.
84 For all languages that have strict marking of transitivity(with which I am acquainted),
passives and antipassivesare clearly intransitive;cf. Langacker& Munro 1975. For languages
with more fluid transitivity(e.g. English),it has not been the custom to comment on the transi-
tivity of passives. However, I do not believe that the treatment here is inconsistent with any
properties of English passives (or with other treatmentsof them).
85 'Passive' is of course a well-establishedgrammaticalterm, and its application is fairly
consistent. The label 'antipassive' was coined about 1968 by Michael Silverstein to refer to
such phenomenaas the -ray transformationin Dyirbal (see ?1.1, above, and Dixon 1972:65-7).
The term has recently been extended to phenomena of a widely differentnature; thus Heath
(1976b:203) mentions, among other examples, that 'indefinite object deletion in the type He
drinkscan be considered an antipassiverule'. By the criteriaI am employing, He drinksis not
an antipassive;the original0 NP (referringto that whichis consumed)cannot occur in oblique
form. Many of Heath's examples, and some of those in Postal 1977, fall outside the scope
of 'antipassive' as the term is employed here.
Comrie (1973:244 ff., MS b) describeswhat appears to be an antipassive transformationfor
Chukchee. See also Woodbury (1975:26-7) on the antipassivein Eskimo.
I shall not here attempt a full account of the reasons for the existence and the
uses of antipassives and passives. A basic reason is, as just indicated, to bring A
into S function when coordination or subordination operations take O or S as
pivot, or to bring O into S function when A or S is taken as pivot (see ?6.2, ?7).
That type of NP which functions as the syntactic pivot is usually obligatory in a
sentence. Woodbury 1975 reports that, in Eskimo, each sentence must involve an
absolutive NP. If the O NP is not to be stated, for some reason, antipassive must be
applied; the deep A NP is now in absolutive case, and the deep O NP-taking
instrumental inflection-can be freely deleted. Identical remarks apply to Dyirbal.
Similarly for passives: Kurylowicz notes that there is no language with a passive
transformation which does not then permit agent deletion.
Most conditions on the use of passive also apply to antipassive. Comrie (MSc)
has shown that, in many languages, 'subjects' (i.e. S and A NP's) are typically
'definite' and 'animate'. In these cases, passive may be used to bring a definite/
animate O NP into surface S function. Dyirbal shows a similar phenomenon from
the ergative angle: demonstratives exist only in absolutive case, for S and O
functions. Thus, if a demonstrative is to be used to mark a deep A NP, it must be
brought into surface S function through antipassivization (e.g. line 9 of text XXV
in Dixon 1972:388).
6.2. SYNTACTIC'PIVOTS'. We recognize a level of 'shallow structure', derived

from deep structure by the application of singulary transformations such as passive,
antipassive, reflexive, and reciprocal. For the majority of sentences, shallow
structure will be identical to deep structure: i.e. passive, antipassive etc. apply only
in a minority of instances. (A passive construction is, in a number of ways,
'marked' with respect to its active counterpart, and so on.) I am saying, then, that
a considerable number of the sentences of a language are 'kernel sentences', in the
sense of Chomsky 1957.
Each language has a morphological (or word-order) marking convention that
enables one immediately to recognize which of the core functions an NP (S, A, or O)
is in within a kernel sentence. Thus, in Dyirbal, an S NP will have absolutive case on
nouns and nominative on pronouns (and it cannot involve a noun in ergative or a
pronoun in accusative inflection); in English, an S NP occurs before the verb (and
triggers verb concord), and there is no prepositionless NP after the verb.
Passive, antipassive, reflexive etc. derive a shallow structure that differs from
the underlying deep configuration. There will be some marking, usually on the verb,
indicating the construction type, and the NP's will be in DERIVED
FUNCTION.Thus,
in a passive, the deep 0 NP receives morphological marking that characterizes an
S NP in a kernel sentence; we say that it is in derived S function. Similarly, in an
antipassive construction, it is the deep A NP that is, on morphological/surface-
syntactic grounds, said to be in derived S function. We can thus speak of that NP
which is in 'deep A/S/0 function' and that NP which is in 'derived A/S/0 functions
in a given sentence. (For a kernel sentence, these will coincide; if passive, reflexive
etc. has applied, then they will differ.)
It is at the level of shallow structure that operations of coordination and sub-
ordination apply ('generalized transformations' in the sense of Chomsky 1957).
ERGATIVITY 121
Typically, each operation of this type may impose a syntactic condition on the
coreferential NP's in the clauses of the construction. We can describe these
possibilities in terms of a 'syntactic pivot'. There are basically two varieties of
pivot (some languages show just one type, others have a mixture of the two):86
(i) S/A PIVOT: the coreferential NP must be in DERIVED S or A function in one
(or both) clauses.
(ii) S/O PIVOT: the coreferential NP must be in DERIVED S or O function in one
(or both) clauses.
Consider two clauses that are related in a coordinate or subordinateconstruction,
and have an NP in common. In some languages, for the construction to be well-
formed, the coreferential NP can be in virtually any syntactic function in either
clause; but there may be syntactic conditions on deletability of one occurrence of
this NP. Other languages have definite conditions on the function of the common
NP in each clause, if the construction is to be well-formed.In ?7I give examplesof S/A
and S/O pivots in coordinate and subordinate constructions for two Australian
languages; Keenan & Comrie 1977 discuss conditions on relative clause formation
in a number of languages.
There are several quite different ways of marking coreferentiality and meeting
coreferentiality conditions; each language tends to specialize in just one method.
We can distinguish the following:
(a) Some languages allow a basically transitive verb to occur in an intransitive
construction in order to meet syntactic conditions on coordination and sub-
ordination (an example is given in Dixon, MS a).
(b) Some languages employ a system of switch-referencemarkers. Thus, in the
South Australian language Diyari, there are two forms of verbal inflection for each
type of subordinate clause-one indicating that the coreferential NP's are both in
(derived) S/A function, and the other indicating that this is not the case.87 The
second occurrence of the coreferential NP can be freely deleted without any
possibility of ambiguity or confusion (Austin 1978). Languages with switch-
reference systems typically lack passive and antipassive transformations (or if they
do have them, they are used rather sparingly); it can be argued that these are not
needed, since it is not necessary to bring a deep 0 or A NP into surface S function,
to facilitate NP deletion or the like.
(c) The third and commonest type of language uses passive or antipassive to feed
'pivot conditions' on the formation of subordinate/coordinateconstructions and/or
deletability conditions; these operations involve putting a critical NP into derived
S function.
It is vitally important to distinguish between 'subject' and 'pivot'. 'Subject' is a
86
'Pivot' is here used in very much the sense attached to 'topic' in Dixon 1972. 'Topic' has
recently been accepted with a differentsense (e.g. Li & Thompson 1976), and it now seems
advisable to employ a differentterm.
87 Switch-reference
systems have been recognized and investigated only quite recently (e.g.
Jacobsen 1967, Keenan 1976:315-16). All those so far described equate S and A. There is no
reason to suppose that a switch-referencesystem could not work in terms of an S/0 pivot; it
would certainly be worthwhile instituting a thorough search (e.g. among North Caucasian
languages) to see whether such a system could be uncovered.
universal category, defined on semantic-syntactic criteria. 'Pivot' is a language-

particular category that is entirely syntactic in nature and application. Whereas
'subject' is applicable only at the level of deep structure, 'pivot' refers to derived
functions. Thus in a kernel sentence, for a language with S/O pivot, an A NP will
be 'subject' but not 'pivot'; in the antipassive version of this sentence, the same
NP will be 'subject' (at deep-structurelevel) and also 'pivot' (in shallow structure).
We can identify (underlying) S, A, and O functions at the deep level, and
(derived) S, A, andO functions in shallow structure. The categories 'subject' and
'pivot' cannot be related to two distinct levels: the criteria for defining 'subject'
and 'pivot' are different, and the types of grammatical behavior determined by
these categories are quite dissimilar.
However, when the syntactic pivot is exclusively (or almost exclusively) S/A,
there is a temptation to use just one term. Linguists typically speak of 'deep
subject' (corresponding to our universally defined category of 'subject') and
'surface subject' (meaning 'pivot'). If it is realized that 'deep subject' is essentially
a semantic notion (determining universal conditions on imperatives, jussive con-
structions and the like), whereas 'surface subject' is a syntactic category (in terms of
which the rules for coordination and subordination in any particularlanguage may
be described), then this terminology need not be too confusing.
But for a language that has a significant S/O pivot, the two terms must be kept
apart. S and A will be related at the deep-structurelevel, but have quite different
relevance in shallow structure: it would be pointless to group together 'derived S'
and 'derived A' functions as 'surface subject' (and to take 'derived S' and
'derived 0' as 'surface subject' would be totally confusing). That is, for a language
that is syntactically ergative, it is meaningless to talk of 'surface subject', although
the usual category of 'subject' is quite applicable at the deep-structurelevel.
Within a universal enquiry, it is best always to distinguish between (deep)
subject and (surface) pivot, even when discussing a language where both involve
the same functional groupings from A, S, and 0. We can talk about 'deep' and
'derived' types of the three primitive functions; 'subject' involves a grouping at the
deep level, and 'pivot' a grouping of derived functions, after any singulary trans-
formations have applied.
6.3. RECENTPROPOSALS CONCERNING ANTIPASSIVES ETC. Postal and Perlmutter
have recently put forward a functionally-based universal theory called ' Relational
Grammar' (RG).88 This recognizes a hierarchy of 'grammatical relations':
(35) a. Subject (i.e. S, A)
b. Object (i.e. 0)
c. Indirect object...
There are a number of 'laws'. The Relational Annihilation Law states that if NPi
88
The discussion here is based on the statement of rules by Postal 1974 and by Woodbury
(1977:335). Further information concerning RG has been gathered from Bell 1976, Postal 1977,
and Perlmutter & Postal 1977; from papers in Cole & Sadock 1977; and from discussion with
Ken Hale, Geoff Pullum, and David Nash. It is not easy to comment on RG, inasmuch as it
has not yet been described in any systematic way; but from the lack of any correction from
Postal or Perlmutter on a draft of this paper, I infer that the account here does not involve any
basic misunderstanding.
ERGATIVlTY 123
assumes the grammatical relation borne by NPj, then NPj ceases to bear any
grammatical relation at all (it is said to be a chomeur or en chomage). The
Motivated Chomage Law states that chomeurs can arise only as a result of the
Relational Annihilation Law. The Reranking Law states that an NP can only be
moved up the hierarchy.
This scheme works perfectly for most phenomena in accusative languages (where
the sole syntactic pivot is S/A). Thus the passive rule raises an 'object' (O) NP
from 2 to 1, the original 1 (A NP) goes into chomage, and the original O becomes
new 1 (with S function, since there is now no O and the sentence is intransitive).89
However, substantial difficultiesarise with 'ergative' languages and antipassives.
The antipassive transformation involves deep A becoming surface S, and O going
into chomage. RG cannot handle this in terms of the hierarchy given above. O
cannot go into chomage unless something replaces it, and only something from
below can replace it. Obviously, some law must be dropped. The simplest change
would be to relax the 'motivated chomage' requirement; if O is allowed to go
spontaneously into limbo, as it were, the absence of a term 2 would lead auto-
matically to the deep A NP being interpreted as surface S (still remaining 1) in
what is now a derived intransitive antipassive construction.90
The difficulty with the RG hierarchy is that it conflates semantic 'subject' with
syntactic 'pivot'. Although it is set up to explain syntactic relations (which relate
to pivot, not subject), Perlmutter and Postal require term 1 always to be subject
(i.e. {S,A})-thus imposing a valid universal category of the DEEP-structure level
onto EVERYlevel of syntactic derivation. A more appropriate course would be to
explain syntactic phenomena in syntactic terms, and to recognize two distinct types
of hierarchy:
(36) HIERARCHY
A HIERARCHY
B
1. Pivot (S, A) 1. Pivot (S, 0)
2. Non-pivot core NP (0) ... 2. Non-pivot core NP (A) ...
Now antipassive works in terms of B as neatly as passive does in terms of A, and
does conform to Postal and Perlmutter's three laws.91
89
Comments in parentheses are supplied by myself. There is apparently no mention of
transitivitynor any distinction between S and A in Postal and Perlmutter'swork.
90There appearsto be no a-priorisemantic or syntactic motivation for any of the three laws.
They have been put forward as putative components of a hypothetical 'grammatical theory'
(in much the same way that mathematiciansdefine a new system in terms of a set of axioms),
and appear to work well for some languages.Attempts have then been made to fit the facts of
other languages into this (in essence, arbitraryand unmotivated) framework.
Postal 1977 has recently suggested a different derivation for ' antipassive'. Preferring
apparentlyto relax the Reranking Law and retain the Motivated Chomage Law (although no
reason is given as to why Motivated Chomage should be considered more important than
Reranking),he suggests that the A NP replace 0, pushing it en chomage; since term 1 is then
vacant, the term 2 NP (deep A) ascends to it. But this implies a derivation A-0O-S; there is
no justification of any sort for the intermediatestep.
91The two-hierarchyapproach was suggested by Johnson 1974, but appears to have been
abandoned by him in 1976. Woodbury 1977 shows the most sophisticatedapproach to dealing
with ergative phenomena in RG terms; his detailed discussion, with particularreference to
Eskimo, has considerable similaritiesto the approach suggested here.
An approach which has some similarities to this has been followed by Keenan
and by Keenan & Comrie. In order to explain antipassivization and various wide-
ranging syntactic operations in Dyirbal (see ??1.1, 7.2), they take 'S and 0' as the
critical pairing.92 But, like Postal and Perlmutter, they also favor a universal
hierarchy which involves 'subject' and 'object', and which conflates semantic and
syntactic information. Keenan takes 'subject' in Dyirbal as the class {S,O0-see the
discussion in ?5.3 above.93
The basic fault with both these schemes is the attempt universally to relate S to
A (in RG) or to relate S to EITHER A ORO for each particular language (in Keenan
& Comrie's treatment); in each case a single grouping is taken to cover (deep-
structure)'subject' and (shallow-structure)'pivot'. There are in fact three primitives
-S, A, and 0-not just two. (In order to deal with a language that has both
passive and antipassive, it is necessary clearly to distinguish S from A and 0, at
every syntactic stage. There are three core terms in the hierarchy, not two, even
though no more than two can occur in any one sentence.)94
SYNTACTIC ERGATIVITY
7. We have already seen that a miscellany of syntactic features have typically
been quoted as evidence for or against syntactic 'ergativity'. In ?5.4 I discussed
phenomena that have the same deep-syntactic orientation in every language where
92 Dixon 1972 groups S and 0 together as 'pivot' (there called 'topic'), and
adopts a tree-
structureconvention which mirrors this. However, no claim is made that this 'pivot' in any
way corresponds to subject (see fn. 68, above). Note that, in Dixon 1972, the term 'deep
structure'is used for an entirely syntactic level, nearerthe surface than the level referredto by
'deep structure'in the presentpaper.
93
The only workable universalhierarchywould be to recognize each primitive as a distinct
term. One ordering would be:
1, S;
2, A;
3, 0.
Then passive would involve 3-1; 2 en chomage. Antipassive would be21-\; 3 en chomage.
Subject would comprise terms 1 and 2 at the deep level (before any derivationshave occurred).
A hierarchyof this type would provide a less concise model for accusativelanguages,but also
a less inappropriateand tortuous one for languages with ergative syntax.
It is not in fact necessaryto set up any sort of hierarchy.Passive, antipassive,and all sorts of
generalizedtransformationscan be handled perfectly well in terms of a model that recognizes
basic configurations [AOV] and [SV] for transitive and intransitive sentences respectively;
transformationaloperations now move terms into and out of the [ ...] core, and so on. For
discussion and exemplificationof this model, see Dixon 1977a.
94
Postal and Perlmutter,as well as Keenan & Comrie, fail to recognize a basic distinction
between transitive and intransitivesentence types. MOSTlanguages have clear morphological
marking of these two sentence types; English (and other modern European languages) are at
one typological extreme in that they do not. Some English sentences are clearly transitive(e.g.
John hit Bill) and some intransitive(John laughed),but there is a large fuzzy area in between
where decisions are not easy: what is the transitivityvalue of Johnshot thepigeon,Johnshot at
the pigeon; John banged the door, John banged on the door? RG, in particular, enables one to
avoid making decisions about transitivity. But this simply obscures important and clear
distinctionsin other languages,and evades the importantquestion (virtuallyavoided by modern
linguists) of investigating transitivity in English.
ERGATIVITY 125
they occur, following from the universal category of' subject'. The way is now clear
for us to examine language-particular syntactic operations: coordination, sub-
ordination, and related phenomena. These normally apply at the level of' shallow
structure', and are fed by passive, antipassive, reflexive, and other singulary
transformations.
It is undoubtedly the case that MOSTsyntactic operations of this type, across the
4,000 or so currently spoken languages, equate S and A functions. That is, (derived)
S and A are grouped together as the syntactic pivot-just as, in all languages,
(deep) S and A functions constitute the category of subject. But it is incontrovertibly
the case that NOT ALLsyntactic operations, in all languages, work in terms of an
S/A pivot. Some languages employ an S/O pivot; others use S/A for certain types of
operation, and S/O for other types.
All languages which use an S/O pivot, to any degree, show some 'ergativity' in
morphological marking. The reverse does not hold: perhaps the majority of
languages which mark S and O in the same way in some part of the morphology
(while marking S and A in the same way in some other part, according to one of the
'splits' documented in ?3) rely exclusively on S/A as syntactic pivot.
In this section we examine two languages, both from Australia, that show some
ergativity at the morphological level; and we investigate the type of syntactic pivot
they use.
7.1. WALMATJARI is spoken in Western Australia (and is a close genetic relative
of Walbiri). Data come from Joyce Hudson (1976a,b, 1978, and p.c.) As in Walbiri,
nouns and free-form pronouns show an absolutive/ergative paradigm:
(37) absolutive (S and 0 functions): 0
ergative (A function): -yu -lu etc.
Each sentence involves an 'auxiliary' (usually as second word). The auxiliary begins
with a modal root (the three possibilities are indicative pa-, interrogative na-, or
imperative/hortative 0) and then four orders of person/number markers.95The
first and fourth order suffixesrefer to (derived) S or A NP's. Second and third order
suffixes refer to an 'accessory NP' if there is one (e.g. 'you' in 'The boys were
talking with you', or 'The boys sat with you', or 'I blew the grass out of your eye');
otherwise they refer to 'dative NP' if there is one (e.g. 'them' in 'I told them about
my dream', 'I work for them'). If the sentence involves neither accessory nor dative
NP's, then a surface 0 NP will be cross-referenced by second and third order
suffixes (cf. fn. 44).96 First and second order suffixes essentially indicate person of
subject and accessory/dative/object, while third and fourth order suffixes show
their number.
Thus, at the morphological level, case-marking on NP's is ergative, but cross-
referencing suffixes in the auxiliary are quite accusative. It is noteworthy that,
95 Some details have been omitted here; they do not affect the point under discussion.
Complete informationis in Hudson 1978.
96
Accessory and dative NP's can occur in transitive and in intransitive clauses. Note that
the auxiliarywill always cross-referencetwo NP's for a transitivesentence (there will always be
an 0 NP, which gets cross-referencedif accessory and dative are lacking). A single NP is cross-
referencedonly in an intransitivesentence that involves no accessory or dative NP.
The case markingon dative NP's is -ku - -wu; on accessory NP's, it is -la - -ya etc.
despite the ergative case-marking conventions, Walmatjari does not have any
transformation of the antipassive variety.
As in all languages, imperatives operate on an S/A principle at the deep-structure
level. S or A must be 2nd person. The auxiliary root is 0, and the first order suffix
(marking person of S or A) is absent; note, though, that the fourth order suffix
(showing number of S or A) is retained. Suffixes of orders two through four are
removed from the domination of AUX and attached to the verb.
Hudson (1976a:9-12) describes three transformations, each of which links two
clauses to form a complex sentence. First, -ula is added as a suffix to the verb of the
subordinate clause:
(38) tikiryan+ ula ma+ na+ 0 + nja + lu mana+ wati + 0 patjani
return + ula INDIC+ 1(EXCL)+ 3 + PL + PL tree + PL + ABS chopped
'Having returned, we chopped trees.'
An -ula construction describes the activity of the subordinate clause as completed
before that referred to by the main clause is begun.
Second, the addition of-u to the subordinate verb indicates that the action of the
main clause was performed so that the subordinateclause activity would be possible
(e.g. 'We will go to search for pigs', where mu:puy 'search for' bears the suffix -u).
Third, simple coordination is shown by the addition of -t]a: to the last word of
the first of two coordinate clauses (the clauses must have the same mood and
tense).
There is a syntactic condition common to -ula, -u, and -tja: constructions: there
must be an NP common to the clauses, and it must be in surface S or A function in
each clause. That is, S/A is the syntactic pivot for Walmatjari.On the data presented
by Hudson, the language has an entirely accusative syntax, despite the split
ergative/accusative morphology.
This is in fact what we could expect from the absence of an antipassive trans-
formation. I noted in ?6.1 that the antipassive derivation is essentially needed to
put an A NP into S function to satisfy constraints that operate in terms of an S/0
pivot. The only constraints encountered in the grammar of Walmatjariare of type
S/A; there is thus no strong motivation for antipassive.
Since Walmatjari has an S/A pivot, we might expect a passive transformation.
Note, though, that only SOMElanguages of this type show a passive. The reason for
this interrelates with the reason why S/A is by far the most commonly encountered
type of pivot: basically, it is because A and S NP's refer to the participant that
'controls' an event, if anyone does. This property determines the universal category
of 'subject', at the level of deep structure, and the properties of imperatives,
jussives etc. that depend upon it.
A multi-clause sentence links descriptions of events which, in the great majority
of instances, have some common participant. The most common linkage involves
a chain of control,97 e.g. John went home because he wanted to rest; Mary left the
office early and bought a chicken to cook for supper. It is because S and A may be
controllers that they are the NP's which speakers tend to focus on, as protagonists;
97 Of course, some types of clause linkage typically involve 0 NP's, e.g. complementclauses
to verbs of attention, such as John heard Mary laugh.These are outnumbered(in terms of text
occurrences) by coordinations, purposive complements etc.
ERGATIVITY 127
and it is because they may be controllers, who will often be doing something just so
that they may then be able to do something else, that they tend to be the shared
NP's that link a sequence of clauses. All of this tends to favor S/A pivots as the
basis for generalized transformations and general constraints on discourse
structure. A passive transformation can be useful for bringing a deep O NP into
surface S function, and for linking it with another S or A NP; but it may not be
essential. In Walmatjari, a number of other clause-linking operations appear to
work in terms of a common NP that can have ANY function in each of the two
clauses involved; the S/A pivot applies only over part of Walmatjari syntax, a
part that is highly controller-oriented.
The arguments just presented suggest that a language with S/0Opivot would
REQUIRE an antipassive transformation. An NP in deep A function is a likely
controller, and some means would be needed to link it with other NP's across a
clause boundary. If there is a strong S/O pivot, then an antipassive transformation
-putting a deep A NP into surface S function-would be needed, and would
surely be much used. It seems in fact that ALLlanguages which make any use of an
S/O pivot DO have an antipassive transformation.98
7.2. DYIRBAL. In ?1.1 we saw how Dyirbal has absolutive/ergative inflection on
nouns and 3rd person pronouns, but nominative/accusative inflection on 1st and
2nd person pronouns; unlike Walmatjari, there are no cross-referencing bound
affixes. Imperatives treat S and A in the same way, both in selectional restrictions
and deletability.
An antipassive transformation is used quite frequently, largely to meet syntactic
constraints of the S/0 type. We saw in ?1.1 that any two clauses can be coordinated
if they have a common NP that is in surface S or 0 function in each clause. If the
common NP is in S or 0 function in the first clause, but in A function in the second,
then the antipassive transformation must be applied to bring it into derived S
function in the second clause.99 An antipassive construction will not normally be
used in the initial clause of a coordination; but it may be needed to bring a deep A
NP into pivotal S function to continue a coordination.
TheS/0 pivot is pervasive in Dyirbal syntax. A relative clause must have an NP
in common with the main clause, and it must be in S or 0 function in the relative
clause. The verb in the relative clause bears a suffix-eu, followed by a case inflection
agreeing with the case of the common NP in the main clause. Compare the following
examples with 1-20 in ?1.1:
(39)yuma+ygu yabu+0 [du,gara+-yu?0] bura+n
father + ERG mother + ABS cry + REL + ABS see + PAST
'Father saw mother, who was crying.'
98 Comrie (MS b) mentions that the negative participial construction in Chukchee
involves
relativization on S or 0, not directly on A. But there is an antipassive transformation, marked
by verbal suffix -tku, that can put an underlying A NP into derived S function just to satisfy this
syntactic condition. Comrie describes this operation as serving 'to reduce the degree of erga-
tivity of a construction', i.e. to overcome the constraint imposed by an S/O pivot in this part of
the grammar. For a further example of an antipassive operation used to meet a syntactic
condition on subordinate clauses, see Dixon (1977a:277-80, 323-7 etc.) on YidinY.
99
If the common NP is in A function in the first clause, and in S or 0 in the second, then a
different construction, involving verbal affix -yura, can be used (see Dixon 1972:77-9).
(40) yuma+ ygu [duygara+ rju+ ru] yabu+ 0 bura+ n

father + ERG cry + REL+ ERG mother + ABS see + PAST
'Father, who was crying, saw mother.'
(41) yuma+( [yabu+ygu bura+yu+0] duggara?+nYu
father + ABS mother + ERG see + REL+ 0 cry + PAST
'Father, who was seen by mother, was crying.'
Where the common NP is in deep A function in the relative clause, it must be anti-
passivized:
(42) yuma+ -yu- 0
[bural+ ya + yabu+ gu] duygara+ nYu
father + ABS see +ANTIPASS+REL+ABS mother + DAT cry + PAST
'Father, who saw mother, was crying.'
Purposive complement constructions show even stronger S/O characteristics.
(This is somewhat surprising: I remarked in ?3.5 that, on semantic grounds, these
are strong candidates for 'accusative identification', at both morphological and
syntactic levels.) Here the common NP must be in S or O function in both main and
subordinate clauses:
(43) yuma+ banaga+ nYu [yabu+gu bural+ yay +gu]
father + ABS return + PAST mother + DAT see + ANTIPASS+ PURP
'Father returned to watch mother.'
In 43 the complement clause must be antipassivized to put the deep A NP, puma
'father', into surface S function.
I have mentioned that the universal category of 'subject' plays an important role
in the grammar of any language. 'Subject' is a deep-structure category, and
syntactic constraints that depend on it are exclusively at the level of deep structure.
The language-particularcategory of 'pivot' is, however, applicable at the level I am
calling 'shallow structure'. Study of jussive constructions-a class of purposive
complement constructions-in Dyirbal shows how a sentence will simultaneously
(a) meet a deep-structure constraint which links S and A, a corollary of the
universal category of 'subject'; and (b) meet a shallow-structureconstraint which
involves S and 0-i.e., which works in terms of the language-specific syntactic
pivot S/0.
The transitive verb giga-l 'tell to do' can be followed by either (i) a clause
directly quoting the order that was given, whose verb must be in imperative
inflection; or (ii) a purposive complement clause. In the latter instance, the 0 NP of
giga-l must be coreferential with the DEEP S or A NP of the complement clause.
Note, though, the general condition on complement constructions in Dyirbal: that
the (derived) S or 0 NP of the complement clause must be coreferential with the
(derived) S or 0 NP of the main clause. To achieve this, the complement clause
must be antipassivized. Thus, from underlying 44, shallow-structure45 is derived:
(44) eyana yabu giga+n [yabu yuma bura? li]}00
we-A mother-0 tell to do + PAST mother-A father 0 see + PURP
(45) yana yabu giga+n [yabu yuma bural+yay+gu]
we-A mother-0 tell to do + PAST mother-S father-DAT see + ANTIPASS +
PURP
100 Case endings are not specified in these deep and shallow structure representations. Verbal
ERGATIVITY 129
Here 44 satisfies deep-structure constraint (a), and 45 satisfies shallow-structure

constraint (b). An Equi rule deletes the occurrence of the common NP in the sub-
ordinate clause, yielding the surface form:
(46) yana yabu giga+n yuma+gu bural+ yay+gu
we-NOM mother-O tell to do + PAST father + DAT see + ANTIPASS+ PURP
'We told mother to watch father.'
In an intransitivejussive complement clause, of course, there is no difficulty.The
S NP will be coreferential with the O NP of the main clause, thus satisfying the
'subject' constraint (main clause O = complement S or A) and also the 'pivot
condition' (main S/0 = complement S/O):
(47) yana yabu gigan banagaygu 'We told mother to return.'
7.3. It is thus clear that S/A is the pivot for Walmatjari, and S/O the pivot for
Dyirbal. At the syntactic level, Walmatjarican be described as a mildly 'accusative
language', Dyirbal is highly 'ergative'; this is in contrast to their morphological
marking, where both languages show a split ergative/accusative system. Dyirbal
appears to be more ergative at the syntactic than at the morphological level,
whereas ergativity in Walmatjari is exclusively morphological.10l
Some languages mix S/A and S/O as syntactic pivots: YidinYfalls into this cate-
gory (Dixon 1977a,c), and so probably do Eskimo (Woodbury 1975, 1977),
Chukchee (Comrie, MS b), and Tongan (my own field work).
It may be that some languages cannot clearly be characterized, at the syntactic
level, in terms of the ergative/accusative continuum. That is, processes such as
coordination may not operate in terms of well-defined constraints like those
applicable to Walmatjari and Dyirbal. It is difficult to be sure on this point:
although many languages have been adequately dealt with at the level of morphol-
ogy, few have been properly described at the syntactic level (it is sometimes hard to
know whether this indicates something about the language or something about the
linguist!) Certainly, some languages have a considerable set of well-defined
syntactic constraints, which facilitate a clear judgment of their position on the
ergative/accusative syntactic scale; but others have more fluid conditions that
provide slimmer evidence for judgment. For instance, coordination may largely
follow semantic, stylistic, or discourse-organization preferences, rather than
conforming to any strict syntactic matrix.
Generally, languages which have strong syntactic conditions at the shallow-
structure level, in terms of either type of pivot, will have a number of singulary
transformations (such as passive, antipassive, or reflexive) which feed these
conditions (as in Dyirbal). Languages which have more fluid syntax may have
inflection is shown at each level. Note that the purposive form of the transitive root bura-l
(belonging to the predominantly transitive -l conjugation) is burali, whereas the purposive form
of the derived intransitive stem bural+ya-y (belonging to the -y conjugation) is buralyaygu.
101 Some measure of syntactic ergativity is certainly rarer than morphological ergativity.
Anderson (1976:11-13) shows that, although Basque has a predominantly ergative morphology,
its syntax works at least in part in terms of an S/A pivot. (But note the invalidity of Anderson's
argument, on the next page, that Tongan has accusative syntax, despite ergative morphology,
since it deals only with the universal 'subject' syntax of lava 'be possible, can'; see ?5.43, above).
fewer operations which put an NP of a certain underlying functional status into

a different derived status. (Walmatjariis of this type; its S/A pivot is less pervasive
than the S/O pivot in Dyirbal, and in almost all cases NP's have the same function
in deep and in shallow structure.) Further discussion on this topic can be found in
Van Valin (ch. 4).
MORPHOLOGICAL MARKING
8. In progressing from the speaker's ideas to their expression and communi-

cation to some addressee, we begin with 'deep structures'; we optionally apply
operations like passive, antipassive, or reflexive to yield 'shallow structures'; and
then we coordinate and subordinate clauses to yield 'surface structures'. It is at
the last level that morphological marking operates. The sentences that result from
the various syntactic operations will have their NP's marked for case, and/or placed
in certain ordering relations, and/or cross-referenced on the verb or some other
constituent, and so on. All these marking operations depend on the derived
function of each NP: an NP that is deep transitive subject (A function) may be
placed in S function by an antipassive transformation, and then perhaps in O
function within a causative clause. It will be marked in surface structure as an O
NP.102
What motivates morphological marking? In particular, what motivates split
marking, as surveyed in ?2? There are, it seems, at least five possible factors.
8.1. REFLECTING THESYNTACTIC 'PIVOT'. Every language with an S/0 pivot has
some morphological ergativity (e.g. nominal inflection in Dyirbal); and languages
with S/A pivots are partially (and most of them wholly) accusative in morphology.
In ?2.3,I mentioned that ergative is always the marked member of a case system
in which it occurs. The 0 NP is then morphologically unmarked, as is S NP (both
can be said to be in absolutive case). I stressed that such a residual identification of
0 with S is quite negative in character; it could not, on its own, be invested with
much grammatical significance. When there is an S/0 pivot, however, the syntactic
identification indicates that the language does establish a positive connection
between S and 0 functions; the morphological marking (assigning both S and 0
to the unmarked 'absolutive' case) is in such a language one realization of this
positive connection.
There is, then, a considerable difference in the grammatical status of absolutive
case for the following.
(a) Languages with some S/0 pivot: here absolutive realizes the pivot category.
It is reasonable that the morphologically unmarked case should relate to syntactic
functions which are unmarked for critical syntactic operations.
(b) Languages with some ergative morphology, but an entirely S/A pivot: here
ergative involves positive SEMANTIC MARKINGof A function; 0 function is left
unmarked (and of course no marking is needed for S function). Absolutive case, in
this circumstance, merely indicates lack of semantic marking.
8.2. REFLECTING 'SUBJECT'.I have been careful to
THE UNIVERSALCATEGORY
102 For a specific example of this derivation and surface marking, see
buriin ex. 614 of Dixon
(1977a:317).
ERGATIVITY 131
insist that 'subject' is applicable only at the level of deep structure (indeed, it seems
impossible to establish 'subject' as a universal category if this restriction is not
made). Morphological marking, however, applies to surface structures (for which
'pivot' and not 'subject' is most relevant). Note, however, that many deep
structures (the kernel sentences) reach the surface without undergoing any syntactic
transformation; they are plainly the 'unmarked sentences' of the language, since
they bear no suffixes or particles indicating passivization, relativization or the like.
It is quite reasonable that even a syntacticallyergative language like Dyirbal should
have an accusative paradigm for some part of its morphology, reflecting the deep
identification of S and A.
UNDERLYING
8.3. REFLECTING SEMANTIC
TENDENCIES.
My original explanation of
split case systems etc. was that NP's of a certain semantic type are more appropriate
in A than in O function; that activities described by certain verbs are less likely to
be controlled than those described by other verbs; and that prospective events are
most normally viewed in terms of a propensity of some 'agent'. These types of
semantic information may be incorporated into morphological marking, so long
as they do not obscure the recognition of the surface function of an NP. Thus
Walmatjari uses ergative to mark a noun used in A function (whereas pronominal
cross-referencing in the auxiliary is entirely nominative/accusative in nature).
Languages that employ nominal case inflections in addition to cross-referencing
verbal affixes have the most opportunity to reflect more than one of the three
factors listed so far. Thus Choctaw case inflections are entirely nominative/oblique
-indicating the syntactic pivot S/A, and universal subject category {S, A}-while
bound prefixes are oriented to the semantic content of the verb.
8.4. DIACHRONICREASONS.Surface syntax may be more easily affected by
contact with some other language than is morphology. A coordination rule might
shift from S/0 to S/A pivot (or vice versa), leaving a morphology that still reflected
the old pivot. (However, these remarks are quite speculative, in the absence of any
detailed studies of syntactic diffusion of this type.) In ?4, I outlined the diachronic
reasons for ergativity in perfect aspect within the Indic family.
8.5. INTERNALGRAMMATICAL
REASONS.Walmatjari, like Walbiri, has extended
absolutive/ergative marking from nominals to free pronouns-just, it seems, to
simplify this aspect of the grammar. The S/A pivot is reflected in bound pronouns;
free pronouns are used sparingly (mainly for emphasis), and can most economically
be inflected like nouns. Ergative marking on nominals is present in Walmatjari for
semantic or perhaps diachronic reasons; it has been extended to pronouns-
apparently going against the S/A pivot and the universal {S, A} subject-in order
to rationalize the grammar.
Morphological marking is likely often to be caused by a combination of all these
five factors. It is often difficult to separate out a single 'reason' for some type of
marking. Indeed, these five parameters are not independent of each other. For
instance, the accusative paradigm of pronouns in Dyirbal could be described in
terms of the semantic hierarchy(?3.2), where pronouns are prototypical agents-or
in terms of the universal category of 'subject' which underlies imperatives,jussives
etc. There is a connection between these: imperatives,and to a lesser extent jussives,

typically involve pronouns.
8.6. Synchronically, morphological marking must be viewed as interpretive,
serving to indicate underlying semantic and surface-syntacticrelations. We cannot
say that a certain syntactic operation is needed to put the NP in a certain case;
rather, the operation is needed to put the NP in a certain derived function, and this
function is marked by a certain case (as part of the overall case-marking system of
the language, motivated as in ??8.1-8.5). However, in DIACHRONIC development, a
certain morphological system may well motivate a particular syntactic pattern.103
SUMMARY
9. At the level of deep structure, we recognize three universal syntactic-

semantic primitives: A and 0, occurringin transitive sentences, and S in intransitive
sentences. 'Subject' is defined as a universal deep-structure category, involving
functions A and S. Languages cannot be characterized as either 'accusative' or
'ergative' in deep structure.
The operation of optional singulary transformations on deep structures yields
shallow structures. It is at this level that generalized transformations operate,
forming coordinate and subordinate constructions. These rules may treat (derived)
S and A in the same way, or they may treat (derived) S and O in the same way; we
refer to S/A and S/O pivots respectively. If a language has an S/O pivot, it can be
said to have 'ergative' syntax.
Passives and antipassives derive shallow structures(effectively putting underlying
0 or A, respectively, into S function). The existence of an antipassive trans-
formation (in the strict sense, defined in ?6.1) is not, per se, evidence for ergative
syntax. An antipassive 'feeds' shallow-structure constraints, and is normally
needed only when there is an S/0 pivot.
Morphological marking-whether involving case inflections/particles or cross-
referencing bound affixes-can reflect subject, and/or pivot, and/or the semantic
nature of NP's, verbs and aspect/tense (in terms of expectations of syntactic-
semantic correspondence). Many languages have some degree of ergativity at this
103
It is interestingto contrastDyirbal, wherecoordination depends on a common NP in S/0
function in each clause (whateverthe case-markingwithin the NP: ?1.1) with YidinY,where the
common NP must be in S/A function in each clause if it is a pronoun, but in S/0 form if non-
pronominal (Dixon 1977a:388-92). Both Dyirbal and YidinYhave accusativecase-markingfor
pronouns, and ergativemarkingon nouns. Whereascoordination in Dyirbal is SlO, regardless
of the morphological marking on the common NP, the coordination constraint in YidinY
exactly reflectsmorphologicalmarking.
The morphologies can be traced back to a split case system in a common ancestor language,
whereas patterns of coordination (and of subordination) in Dyirbal and YidinYare plainly
recent innovations. It seems that we would originally have had morphological marking deter-
mined largely by factors of ??8.2-8.3; at this stage the languages would perhaps have had a
fairly fluid syntax, without the coreferentialityconstraints that characterizemodern Dyirbal
and YidinY.The syntactic systems of the two modern languageshave been partlymotivatedby
the alreadyexisting morphological paradigms.(See Dixon 1977c for furtherdiscussion of these
examples, and for putative reasons for the different developments of coordination and sub-
ordinationrestrictionsin the two languages.)
ERGATIVITY 133
level; but some accusativity always exists (the split being conditioned as described
in ?3).
A language can be characterized as 'ergative' at two distinct grammatical levels:
shallow structure, and surface morphology. Once universal 'accusative' phenomena
are discounted, some languages are almost entirely ergative in their syntax, working
exclusively in terms of an S/O pivot. Others may show partial ergativity at this
level, using an S/O and also an S/A pivot. These types are greatly outnumbered by
languages that have a fully accusative syntax. But although languages with ergative
syntax are in a minority, they do exist, and they cannot be glibly explained away in
terms of an entirely 'accusative' universal theory (which at the least assigns
impossibly complex grammars to ergative languages).
Every language that is syntactically ergative shows a degree of morphological
ergativity. But many languages relate S and O in some way at the morphological
level, without showing ergative syntax. For some, there are strict syntactic con-
straints, and they work entirely through an S/A pivot. In others, the rules of
coordination and subordination may be quite fluid, without any strict syntactic
constraints (so that there is no clearly recognizable pivot of EITHER type).
Much more work is required on the description of ergative languages before all
the suggestions made here can be validated. This paper has attempted to put
forward a framework for such investigations, attempting to distinguish what is
universally valid from what is language-particular;what is semantically motivated
from what is syntactically determined; and so on. It has been purposefully wide-
ranging and programmatic, glossing over important theoretical issues for the sake
of achieving a broad perspective. I hope it will elicit detailed information on crucial
languages, which may confirm or disconfirm the present conclusions, and which
will lead the way to a fuller understanding of ergativity.
APPENDIX: INVENTORY OF TYPES OF SPLIT
I give here the logical possibilities for types of split, according to the parametersdiscussed
in ?3. Examples are quoted where known. Some of these possibilities may not occur in any
language; a full searchhas not yet been carriedout. (See also Silverstein, 124.)
The following notation is used: if two function-letters X and Y are juxtaposed with no
intervening symbol, this indicates that they are markedin the same way; a hyphen separates
cases. Thus XY-Z will show that one case marksX and Y functionswhile a second case is used
for Z function. A slash separatestwo kinds of morphologicalmarking;i.e., it indicatesthe split.
(1) SPLIT CONDITIONED BY SEMANTIC NATURE OF NP'S.
(a) Partly accusative:AS-O/ASO. E.g. Latin, wherethereis nominative/accusativeinflection
for masculine and feminine declensions, but one inflection covering A, S, and 0, for neuter
nouns.
(b) Partly ergative: ASO/A-SO. E.g. Burushaski.
(c) Partly accusative, the remainderergative(complementarydistributionof accusativeand
ergative case-marking):AS-0/A-SO. E.g. Gugu-Yalanji.
(d) Partly accusative, partly ergative (overlapping distribution of accusative and ergative
case-marking):AS-0/A-S-O/A-SO. E.g. Cashinawa, YidinY,Arabana.
(e) All accusative, partly ergative (distribution of ergative included entirely within distri-
butional scope of accusative): AS-O/A-S-0. No example to hand.
(f) All ergative, partly accusative (distributionof accusative included entirely within distri-
butional scope of ergative): A-S-0/A-SO. E.g. Waga-Waga.
(g) Partly accusative, partly ergative, partly neither (middle area where neither accusative
nor ergative applies): AS-O/ASO/A-SO. E.g. Gurindji, an Australian language (data from
Patrick McConvell).
(2) SPLIT CONDITIONEDBY SEMANTICNATURE OF VERB.
(a) Fluid S-marking(S NP can be marked by either A or O case, depending on semantic
effect of verb in this instance of use): AS-SO. E.g. Bats, Crow.
(b) Split S-marking(S NP is markedlike A for certain verbs and like 0 for other verbs, with
the conditioning largely semantic): AS-O/A-SO. E.g. Dakota etc.
No example is known of a markingused only for S in this type of split (an exception is where
there is unanalyzable'portmanteau' representationof A and 0). Choctaw is unusualin having
a third type of S-marking,identical to the normal markingfor dative NP's: see ?3.1 and fn. 28.
(3) SPLITCONDITIONED
BY TENSE/ASPECT.
There appear never to be more than two systems of
markingwith this type of split; this contrastswith the above sections, where therecan be three
kinds (i.e. AS-O/A-S-O/A-SO).
(a) Partly accusative: AS-O/ASO. No example to hand (Lardil shows an AS-O/ASO split
conditioned by mood, basicallyimperativevs. non-imperative).
(b) Partly ergative: ASO/A-SO. E.g. Burushaski.
(c) All accusative, partly ergative: AS-O/A-S-O. E.g. Pitta-Pitta (Blake 1979).
(d) All ergative, partly accusative: A-S-O/A-SO. No example known.
(e) Partly accusative,remainderergative: AS-O/A-SO. No example known.
REFERENCES
ALLEN, W. S. 1951. A study in the analysis of Hindi sentence structure. Acta Linguistica
Hafniensia 6.68-86.
. 1956. Structure and system in the Abaza verbal complex. Transactions of the
Philological Society, 127-76.
-. 1960. Notes on the Rajasthani verb. Indian Linguistics 21.1-10.
. 1964. Transitivity and possession. Lg. 40.337-43.
ANDERSON, S. R. 1976. On the notion of subject in ergative languages. In Li 1976:1-23.
-. 1977. On mechanisms by which languages become ergative. In Li 1977:317-63.
AUSTIN,P. 1978. A grammar of the Diyari language of north-east South Australia.
Australian National University dissertation.
BELL,S. J. 1976. Cebuano subjects in two frameworks. MIT dissertation.
BENDER,M. L. 1976. The non-Semitic languages of Ethiopia. East Lansing: Michigan
State University.
BENVENISTE, E. 1952. La construction passive du parfait transitif. Bulletin de la Societe
de Linguistique 48.52-62. [Translated in his Problems in general linguistics, 153-61.
Coral Gables: University of Miami Press, 1971.]
BLAKE,B. J. 1976. On ergativity and the notion of subject: some Australian cases.
Lingua 39.281-300.
. 1979. Pitta-Pitta. Handbook of Australian languages, I, ed. R. M. W. Dixon &
B. J. Blake, 183-242. Canberra: ANU Press.
BOAS,F. 1911. Tsimshian. Handbook of American Indian languages 1.283-422.
, and E. DELORIA.1939. Dakota grammar. (Mem. Nat. Ac. Sci., 23.) Washington,
DC: Government Printing Office.
BREEN, J. G. 1976. Ergative, locative and instrumental case inflections: Wangkumara.
In Dixon 1976:336-9.
BRICKER, V. R. MS. Antipassive constructions in Yucatec Maya.
BYINGTON, C. 1870. Grammar of the Choctaw language. Proc. Am. Philos. Soc. 11.317-
67.
CAPELL,A. 1962. Some linguistic types in Australia. (Oceania linguistic monographs, 7.)
Sydney.
. 1969. A survey of New Guinea languages. Sydney: University Press.
CATFORD, J. C. 1975. Ergativity in Caucasian languages. [Mimeo.]
ERGATIVITY 135
CERNY,V. 1971. Some remarks on syntax and morphology of verb in Avar. Archiv
Orientalni 39.46-56.
CHAFE,W. L. 1970. A semantically based sketch of Onondaga. (Indiana University
publications in anthropology and linguistics, Memoir 25). Bloomington.
CHOMSKY, N. 1957. Syntactic structures. The Hague: Mouton.
-- . 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press.
CHUNG, S. 1976. Case-marking and grammatical relations in Polynesian. Harvard
University dissertation.
-. 1977. On the gradual nature of syntactic change. In Li 1977:3-55.
CHURCHWARD, C. M. 1953. Tongan grammar. Oxford: University Press.
CLARK,R. 1973. Transitivity and case in eastern Oceanic languages. Oceanic Linguistics
12.559-605.
-- . 1976. Aspects of Proto-Polynesian syntax. (Te Reo monographs.) Auckland:
Linguistic Society of New Zealand.
COLE,P., and J. M. SADOCK(eds.) 1977. Grammatical relations. (Syntax and semantics,
8.) New York: Academic Press.
COLLINDER,B. 1965. An introduction to the Uralic languages. Berkeley & Los Angeles:
University of California Press.
COMRIE,B. 1973. The ergative: variations on a theme. Lingua 32.239-53.
-- . 1975. The antiergative: Finland's answer to Basque. CLS 11.112-21.
. 1976a. Aspect. Cambridge: University Press.
. 1976b. Review of Klimov 1973. Lingua 39.252-60.
. MS a. Ergativity. To appear in a volume of lectures given at the Linguistic
Institute, Oswego, 1976, ed. by W. P. Lehmann.
--. MSb. Degrees of ergativity: some Chukchee evidence. To appear in Ergativity:
towards a theory of grammatical relations, ed. by F. Plank.
. MSc. 'Definite' and 'animate': a natural class? To appear in Linguistica Silen-
siana 3.
CRAIG,C. 1975. Jacaltec syntax: a study of complex sentences. Harvard University
dissertation. [Now published as The structure of Jacaltec. Austin: University of
Texas Press, 1977.]
. 1976. Properties of basic and derived subjects in Jacaltec. In Li 1976:99-123.
DESHERIEV,Ju. D. 1953. Batsbijskij jazyk. Moscow: Akademia Nauk.
DIXON,R. M. W. 1970. Proto-Australian laminals. Oceanic Linguistics 9.78-103.
--. 1971. A method of semantic description. Semantics, an inter-disciplinary reader
in philosophy, linguistics and psychology, ed. by D. D. Steinberg & L. A. Jako-
bovits, 436-71. Cambridge: University Press.
--. 1972. The Dyirbal language of North Queensland. Cambridge: University Press.
. 1973. The semantics of giving. The formal analysis of natural languages, ed. by
M. Gross et al., 205-23. The Hague: Mouton.
-- (ed.) 1976. Grammatical categories in Australian languages. Canberra: Australian
Institute of Aboriginal Studies.
. 1977a. A Grammar of YidinY. Cambridge: University Press.
. 1977b. Some phonological rules in YidinY. LI 8.1-34.
-- . 1977c. The syntactic development of Australian languages. In Li 1977:365-415.
. 1977d. Where have all the adjectives gone? Studies in Language 1.19-80.
. MSa. Grammatical re-analysis: an example of linguistic change in Wargamay,
North Queensland.
--. MSb. The languages of Australia. Cambridge: University Press.
EADES,D. 1979. Gumbaynggir. Handbook of Australian languages, I, ed. by R. M. W.
Dixon & B. J. Blake, 245-361. Canberra: ANU Press.
FILLMORE, C. J. 1968. The case for case. Universals of linguistic theory, ed. by E. Bach
& R. T. Harms, 1-88. New York: Holt, Rinehart & Winston.
FODOR,J. A. 1970. Three reasons for not deriving 'kill' from 'cause to die'. LI
1.429-38.
FOLEY, W. A. 1976. Comparative syntax in Austronesian. Berkeley: University of

California dissertation.
GODDARD, I. 1967. The Algonquian independent indicative. Contributions to anthro-
pology: Linguistics 1, 66-106. Ottawa: National Museum of Canada.
GREGORES, E. and J. A. SUAREZ.1967. A description of colloquial Guarani. The Hague:
Mouton.
HAAS, M. 1940. Tunica. New York: Augustin.
HAIMAN, J. M. MS. Hua: a Papuan language of New Guinea. Languages and their
status, ed. by T. A. Shopen. Cambridge, MA: Winthrop.
HALE, K. L. 1967. Case and voice in some Australian languages. [Mimeo.]
--. 1968a. Preliminary remarks on Walbiri grammar, II. [Mimeo.]
--. 1968b. Review of A profile-generative grammar of Maori, by P. W. Hohepa.
Journal of the Polynesian Society 77.83-99.
--. 1970. The passive and ergative in language change: the Australian case. Pacific
linguistics studies in honour of Arthur Capell, ed. by S. A. Wurm & D. C. Laycock,
757-81. Canberra: Pacific Linguistics.
. 1973. Person marking in Walbiri. A Festschrift for Morris Halle, ed. by S. R.
Anderson & P. Kiparsky, 308-44. New York: Holt, Rinehart & Winston.
HEATH,J. 1976a. Ergative/accusative typologies in morphology and syntax. In Dixon
1976: 599-611.
. 1976b. Antipassivization: a functional typology. Berkeley Linguistics Society
2.202-11.
. 1977. Choctaw cases. Berkeley Linguistics Society 3.204-13.
HERSHBERGER, H. 1964. Case-marking affixes in Gugu-Yalanji. In Pittman & Kerr, 73-82.
HERSHBERGER, R. 1964. Personal pronouns in Gugu-Yalanji. In Pittman & Kerr, 55-68.
HETZRON, R. 1976. The Agaw languages. (Afroasiatic Linguistics, 3:3.) Malibu: Undena.
HILL,J. H. 1969. Volitional and non-volitional verbs in Cupeno. CLS 5.348-56.
HINTON,L., and M. LANGDON.1976. Object-subject pronominal prefixes in La
Huerta Diegueno. Hokan studies, ed. by M. Langdon & S. Silver, 113-28. The
Hague: Mouton.
HOHEPA,P. W. 1969. The accusative-to-ergative drift in Polynesian languages. Journal
of the Polynesian Society 78.297-329.
HUDSON,J. 1976a. Walmatjari: nominative-ergative or nominative-accusative? Papers
in Australian Linguistics 9.1-30. Canberra: Pacific Linguistics.
. 1976b. Simple and compound verbs: conjugation by auxiliaries in Australian
verbal systems-Walmatjari. In Dixon 1976:653-67.
. 1978. The core grammar of Walmatjari. Canberra: Australian Institute of
Aboriginal Studies.
HYMES,D. 1975. From space to time in tenses in Kiksht. IJAL 41.313-29.
JACOBSEN, W. H. 1967. Switch-reference in Hokan-Coahuiltecan. Studies in south-
western ethnolinguistics, ed. by Dell Hymes & W. Bittle, 238-63. The Hague:
Mouton.
JOHNSON,D. 1974. On the role of grammatical relations in linguistic theory. CLS
10.269-83.
. 1976. Ergativity in relational grammars. [Mimeo.]
KACHRU,Y. 1965. A transformational treatment of Hindi verbal syntax. University of
London doctoral thesis.
KALMAR,I. 1976. The antipassive in Inuktitut. [Mimeo.]
KASCHUBE, D. V. 1967. Structural elements in the language of the Crow Indians of
Montana. Boulder: University of Colorado Press.
KEEN, S. 1972. A description of the Yukulta language. Monash University master's
thesis.
KEENAN,E. L. 1976. Towards a universal definition of 'subject'. In Li 1976:303-33.
--, and B. COMRIE.1977. Noun phrase accessibility and universal grammar. LI
8.63-100.
ERGATIVITY 137
, and E. O. KEENAN.MS.Becoming a competent speaker of Malagasy. Languages

and their speakers, ed. by T. A. Shopen. Cambridge, MA: Winthrop.
KENNARD,E. 1936. Mandan grammar. IJAL 9.1-43.
KLIMOV,G. A. 1973. Ocerk ob?cej teorii ergativnosti. [Outline of a general theory of
ergativity.] Moscow: Nauka.
KLOKEID, T. J. 1976. Topics in Lardil grammar. MIT dissertation.
KUIPERS,A. H. 1968. The categories verb-noun and transitive-intransitive in English
and Squamish. Lingua 21.610-26.
KURYLOWICZ, J. 1960. Ergativnost' i stadial'nost' v jazyke. [Ergativity and the stadial
theory of language.] In his Esquisses linguistiques, 95-103. Wroclaw.
LAKOFF,G., and J. R. Ross. 1976. Is deep structure necessary? Notes from the lin-
guistic underground, ed. by J. D. McCawley (Syntax and semantics, 7), 159-64.
New York: Academic Press.
LANGACKER, R. W., and P. MUNRO.1975. Passives and their meaning. Lg. 51.789-830.
LEWIS,G. L. 1953. Teach yourself Turkish. London: English Universities Press.
Li, C. (ed.) 1976. Subject and topic. New York: Academic Press.
-- (ed.) 1977. Mechanisms of syntactic change. Austin: University of Texas
Press.
, and S. A. THOMPSON. 1976. Subject and topic: a new typology of language. In Li
1976:457-89.
-; -; and J. 0. SAWYER. 1977. Subject and word order in Wappo. IJAL 43.85-100.
LISTER-TURNER, R., and J. B. CLARK.1930. Revised Motu grammar and vocabulary.
Port Moresby.
LOOGMAN, A. 1965. Swahili grammar and syntax. Pittsburgh: Duquesne University
Press.
LORIMER, D. L. R. 1935. The Burushaski language, I: introduction and grammar. Oslo:
Aschehoug.
LYONS,J. 1968. Introduction to theoretical linguistics. Cambridge: University Press.
MCCONVELL, P. 1977. How Lardil became accusative. [Mimeo.]
MCDONALD,M. A., and S. A. WURM.MS.Basic materials in Galali: grammar, sentences
and vocabulary. Canberra: Pacific Linguistics.
MCKAY, G. R. 1975. Rembarnga: a language of central Arnhem Land. Australian
National University dissertation.
MCLENDON,S. 1978. Ergativity, case and transitivity in Eastern Pomo. IJAL 44.1-9.
MEILLET,A. 1917. Caracteres generaux des langues germaniques. Paris: Hachette.
[Translation: General characteristics of the Germanic languages. Coral Gables:
University of Miami Press, 1970.]
MEJLANOVA, U. A. 1960. Morfologiceskaja i sintaksiceskaja xarakteristika padezej
lezginskogo jazyka [Morphological and syntactic characteristics of Lesghian cases.]
Maxackala: Dagestanskij Filial Akademii Nauk SSSR.
MERRIFIELD, W. R., et al. 1965. Laboratory manual for morphology and syntax. Santa
Ana: SIL.
MILNER,G. 1973. It is aspect (not voice) which is marked in Samoan. Oceanic Lin-
guistics 12.621-39.
MOREAU,J.-L. 1972. La correlation du sujet et de l'objet en finnois. Etudes Finno-
Ougriennes 8.193-202.
NASH, D. MS.Is Ngarluma split-ergative? To appear in Ergativity: towards a theory of
grammatical relations, ed. by F. Plank.
N'DIAYE, G. 1970. Structure du dialecte basque de Maya. The Hague: Mouton.
O'GRADY,G. N.; C. F. VOEGELIN; and F. M. VOEGELIN. 1966. Languages of the world:
Indo-Pacific fascicle Six. Anthropological Linguistics 8:2.
PERLMUTTER, D. M. 1971. Deep and surface structure constraints in syntax. New York:
Holt, Rinehart & Winston.
- and P. M. POSTAL.1977. Towards a universal characterization of passivization.
Berkeley Linguistics Society 3.394-417.
PITTMAN,R., and H. KERR(eds.) 1964. Papers on the languages of the Australian

Aborigines. Canberra: Australian Institute of Aboriginal Studies.
POSTAL,P. M. 1974. Report of work done as part of MSSB Workshop on Constraints
on Grammars. [Mimeo.]
--. 1977. Antipassive in French. NELS 7.273-313.
REGAMEY, C. 1954. A propos de la 'construction ergative' en indo-aryen moderne.
Sprachgeschichte und Wortbedeutung: Festschrift Albert Debrunner, 363-84.
Bern: Francke.
RIGSBY,B. 1975. Nass-Gitksan: an analytic ergative syntax. IJAL 41.346-54.
ROBINETT, F. M. 1955. Hidatsa. IJAL 21.1-7, 160-77, 210-6.
ROOD,D. S. 1971. Agent and object in Wichita. Lingua 28.100-107.
RUMSEY,A. MS.Prolegomena to a theory of Australian grammatical case systems. To
appear in Papers in linguistics from the 1976 AIAS Biennial meeting, ed. by B. J.
Rigsby. Canberra: Australian Institute of Aboriginal Studies.
SAPIR, E. 1917. Review of 'Het passieve karakter ...', by C. C. Uhlenbeck. IJAL 1.82-6.
--. 1921. Language. New York: Harcourt Brace.
--. 1930. Southern Paiute, a Shoshonean language. (Proc. Am. Acad. Arts &
Sci., 65:1.) Boston.
SCHACHTER, P. 1976. The subject in Philippine languages: topic, actor, actor-topic, or
none of the above. In Li 1976:491-518.
. 1977. Reference-related and role-related properties of subjects. In Cole &
Sadock, 279-306.
SHIPLEY,W. F. 1964. Maidu grammar. Berkeley & Los Angeles: University of Cali-
fornia Press.
SILVERSTEIN, M. 1976. Hierarchy of features and ergativity. In Dixon 1976:112-71.
TAYLOR,A. J. 1970. Syntax and phonology of Motu (Papua): a transformational
approach. ANU dissertation.
THALBITZER, W. 1911. Eskimo. Handbook of American Indian languages 1.967-1069.
TIMBERLAKE, A. 1974. The nominative object in North Russian. Slavic transformational
syntax, ed. by R. Brecht & C. Chvany, 219-43. Ann Arbor: University of Michigan.
VAN VALIN,R. D. 1977. Aspects of Lakhota syntax. Berkeley: University of California
dissertation.
VOGT,H. 1971. Grammaire de la langue georgienne. Oslo: Universitetsforlaget.
WALSH,M. 1976a. Ergative, locative and instrumental case inflections: Murinjpata. In
Dixon 1976: 405-8.
. 1976b. The Murinjpata language of north-west Australia. ANU dissertation.
WHITMAN, W. 1947. Descriptive grammar of Ioway-Oto. IJAL 13.233-48.
WOODBURY, A. C. 1975. Ergativity of grammatical processes: a study of Greenlandic
Eskimo. University of Chicago master's essay.
. 1977. Greenlandic Eskimo, ergativity and relational grammar. In Cole & Sadock,
307-36.
WURM,S. 1976. Accusative marking in Duungidjawu (Waga-Waga). In Dixon 1976:
106-11.
[Received 27 January 1978.]

Ergativity R. M. W. Dixon

Uploaded by

Copyright:

Available Formats

Ergativity R. M. W. Dixon

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ergativity R. M. W. Dixon

Uploaded by

Copyright:

Available Formats

Linguistic Society of America

-ya-y. (Reasons for regarding an antipassive construction in Dyirbal as a derived

A function yabu + gu auma + ygu

For these reasons, 'accusativelergative' characterization based on word order

Similar remarks apply for cross-referencingpronominal affixes: either the A or the

2.3. MARKEDNESS. We have seen that case inflectionsexist primarilyto distinguish

16 Other examples of 'optional ergative' include the Australian

(27) wagu-da + 0 gunda+ dYi+r dyugi+ :1 (galba: n + da)

A similar situation prevails in other Balto-Fennic languages (Collinder 1965:

This terminology could appropriately be taken further; the name 'extended

generation of speakers has extended formal markedness to apply also at the

with morphological markedness (the Germanic example quoted in ?2.33 is one of

on an S NP appears to be semantically determined: instead of having to recognize

distinctionscross transitivityboundaries(cf. volitional -wecdxnen 'throw down' with voluntary

is entirely conditioned by the referential token of that instance of the intransitive

3.22. We have been viewing the hierarchy as an explanation of what is likely to

(c) Walbiri morphology was simplified, in that the absolutivelergative case

In this perspective, there is continual movement of time from the right-hand

43 Lorimer mentions two furtherpeculiarities:ergative MUST be used on the A NP of henWs

to S or 0. But in subordinate 'aspectless' clauses, Set 1 refers to A or S, and Set 2 to 0. On the

EVOLUTION OF ERGATIVE MORPHOLOGIES

linguists emphasize semantic criteria, but encounter severe difficultiesin explaining

Paul Schachterand Rodney Huddleston for discussion of 'double agent' verbs.)

everylanguage.ThesemanticlinkbetweenA andthe subtypeSa of S is generalized,

5.3. KEENAN'S DISCUSSION OF'SUBJECT'. In an important paper, Keenan seeks 'to

the label 'subject' should, on this type of criteria, be applied to a pronoun in S or A

(Rigsby), one pronominal suffix cross-references S or 0, and another suffix cross-

is falling) corresponds to the O NP of the corresponding causative (The woodmanis

6.2. SYNTACTIC'PIVOTS'. We recognize a level of 'shallow structure', derived

universal category, defined on semantic-syntactic criteria. 'Pivot' is a language-

(40) yuma+ ygu [duygara+ rju+ ru] yabu+ 0 bura+ n

Here 44 satisfies deep-structure constraint (a), and 45 satisfies shallow-structure

fewer operations which put an NP of a certain underlying functional status into

8. In progressing from the speaker's ideas to their expression and communi-

etc. There is a connection between these: imperatives,and to a lesser extent jussives,

9. At the level of deep structure, we recognize three universal syntactic-

APPENDIX: INVENTORY OF TYPES OF SPLIT

FOLEY, W. A. 1976. Comparative syntax in Austronesian. Berkeley: University of

, and E. O. KEENAN.MS.Becoming a competent speaker of Malagasy. Languages

PITTMAN,R., and H. KERR(eds.) 1964. Papers on the languages of the Australian

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.