5
SUBJECT
]
5.1    THE PROBLEM
In section 3.3, we mentioned some of the problems inherent in working with grammatical relations, including subject, and some of the possible approaches towards a solution to these problems. In the present chapter, we will look in considerably more detail at one particular aspect of this problem, namely the definition of subject cross-linguistically. Subject is an important notion, used frequently, both in traditional grammar and in more recent linguistic work, both in the descriptions of individual languages and in stating cross-linguistic generalizations. If linguists were invariably in agreement in stating which noun phrase, in each construction in each language, is the subject, then we could, perhaps, accept this inter-subjective agreement, and devote correspondingly less energy to trying to find an explicit definition of subject. However, it turns out that, in a wide range of cases, this inter-subjective agreement is lacking, so that the need does arise as a serious empirical problem to establish criteria for declaring a given noun phrase to be or not to be a subject.
One particular instance of lack of agreement among linguists on subject-hood is illustrated by competing analyses of the ergative construction. We shall return below, in sections 5.3 and 6.2.2 to a more detailed discussion of ergativity, and for present purposes we may simply give some illustrative examples of the kind of problem that arises, using Chukchi as our example:
Yam ta-yet-yPek. (1)
I -ABSOLUTivE came-1 SINGULAR
II came.'
Yam-nan   yat ta-lPu-yat. (2)
I-ERGATIVE thou-ABSOLUTIVE saw-1SINGULAR-2SINGULAR
' I saw thee.'
Analyses of English agree that, in the English versions of these two sentences, / is subject both of the intransitive construction of (1) and of the transitive construction of (2); moreover, English morphology, at least for pronouns, exactly mirrors this distribution: the subjects are in the nominative, the direct object in the accusative. In Chukchi, as in English, there are two cases used for these three noun phrases, but their distribution is quite different: the absolutive case is used to translate / (intransitive subject) of (1), and to translate thee (direct object) of (2), whereas a separate case, the ergative, is used to translate / (transitive subject) of (2). The question therefore arises whether, in Chukchi, one should not rather group together the absolutive noun phrases as subject, following the morphology, rather than simply following the distribution that turns out to be relevant for English. Although in early periods many linguists working on ergative languages tried to solve this problem a priori, by fiat - and in either direction, by relying on the morphology or by disregarding it - the question is in fact an empirical question, and in sections 5.3-4 we will see that its answer is much less simple than either of these solutions. For the moment, however, we may simply note that the problem exists.
Of course, in addition to criteria of case marking in establishing subject-hood, it will be clear from the discussion of section 3.3 that syntactic criteria are also important in establishing subjecthood. In English, for instance, we can note the following two syntactic criteria of subjecthood. First, verbs agree in person and number with their subject; although English verb morphology is fairly atrophied, this distinction is still maintained consistently in the difference in the present tense between third person singular and all other forms, and in a few other instances with irregular verbs, so that we have the third person singular form in he sees you but the non-third person singular form in / see you. Secondly, in the kinds of constructions called subject-to-object raising by many transformational-generative grammarians, we find that the subject of a rfcaf-clause, and only the subject, can, after certain verbs, appear in an alternative construction of type (4):
/ believe that Max is an accountant. (3) / believe Max to be an accountant. (4)
In the vast majority of sentence-types, these two syntactic criteria coincide, i.e. there is agreement between logically independent criteria as to the subject in English.There are, however, some sentence types where this agreement is not found, such as sentences introduced by there is/are:
There are unicorns in the garden. (5) There is a unicorn in the garden. (6)
98
ioo
SUBJECT
In such examples, at least in the standard language, verb agreement is determined by the noun phrase that follows there is/are. Subject-to-object raising, however, treats there as the subject, giving:
/ believe there to be a unicorn/unicorns in the garden.
(7)
And, indeed, in such instances we rind disagreement as to which noun phrase, in (5) and (6), should be considered the subject: different weighting of different criteria gives different results. So even in English there are some construction types where there is no agreement among linguists as to which noun phrase is subject.
Faced with such problems surrounding the characterization of the notion subject, there are two possible approaches. On the one hand, one could claim that the notion of subject is misleading from the outset, and should be banished from linguistic theory. On the other hand, one could try and work out a definition of subjecthood which, while corresponding to linguists' inter-subjective intuitions in the clear cases, would also make insightful claims about the unclear cases. In the present chapter, we follow the second of these paths. Before embarking on the details of the definition, however, we should make some further preliminary remarks. First, we are not committed a priori to the view that subject is a necessary descriptive category in the grammar of every language: there may well be languages where it is not appropriate, though equally there are languages (including English) where it is appropriate. Secondly, we are not committed to the view that, even in a language where subject is generally valid, every sentence will necessarily have a subject. Thirdly, we are not committed to the view that the translation of a sentence from language X where a certain noun phrase is subject will necessarily have that same noun phrase as subject in language Y. Examples of all of these points will occur below.
Finally, although we will argue that the notions of topic and agent must play a role in the definition of subject, we argue that, even in English, it is clear that the notion of subject cannot be identified with either of these notions. If we take, for instance, our criterion of verb-agreement, then it is clear that in the passive sentence the men were hit by the boy, the plural verb were does not agree with the agent; and it is equally clear that in the topicalized sentence John I know the non-third person singular verb is not in agreement with the topic. However close the connection may be among grammatical relations, semantic roles, and pragmatic roles, they cannot be identified with one another.
5.2    ON DEFINITIONS AND CATEGORIES
Before turning specifically to the definition of subject, it is necessary for us to make some preliminary remarks on the nature of definitions, in par-
r
SUBJECT
ioi
ticular on the nature of definitions of linguistic categories, in order to avoid certain later misunderstandings. The kind of definition of subject towards which we will be working is the following: the prototype of subject represents the intersection of agent and topic, i.e. the clearest instances of subjects, cross-linguistically, are agents which are also topics. There are two important characteristics of this definition: first, it is multi-factor; second, it is stated in terms of prototypes, rather than in terms of necessary and sufficient criteria for the identification of subjects. The second point is particularly important, given that many subjects in many constructions in many languages are not topic, or are not agent, or are neither.
The use of a multi-factor definition is unlikely to raise any eyebrows, since such definitions are quite widespread in linguistics and other areas, as for instance if we define preposition in terms of the intersection of adposition and position in front of the governed noun phrase. However, the attempt to use definitions in terms of prototypes for linguistic categories has met with an inordinate amount of opposition and prejudice, so that it is worth spending some time on discussion of this issue. Rather than discussing the problem directly in terms of subject properties, we will use some other examples, where the use of prototypes is much more clearly justified. Note that the use of these analogies does not in itself justify the use of a prototype-based definition of subject, but it does demonstrate that we cannot a priori reject this kind of definition, but must rather weigh up the pros and cons in terms of their fit with the data and their evaluation relative to alternative definitions.
In chapter 2, we illustrated one very clear area where definitions of categories in terms of prototype seem to be required, namely with colour terms, where humans seem to recognize a central, focal value for a colour term, rather than clear-cut boundaries. What this means is that there is no set of necessary and sufficient conditions that an object must satisfy in order to be called, for instance, red. But equally, this does not mean that we can state no restrictions on the use of the term red: this term is most appropriate for the focal value, and less and less appropriate as one moves away from this focal area and approaches the foci of other colour terms. This example thus establishes that there is at least one area where humans do categorize in terms of prototypes, thus opening up this kind of definition as a real possibility.
Similar examples can also be found using more clearly linguistic categories, and the example we will use here concerns the distinction between nouns and adjectives in Russian, in particular the relation of numerals to these two. In Russian, in general, the distinction between nouns and adjectives is clear-cut, so that we can establish criteria that correlate with the focal values (prototypes) of noun and adjective. Numerals, however, fall in between these two prototypes, in a way that makes impossible any estab-
102
subject
subject
103
lishment of non-arbitrary cut-off points. In distinguishing adjectives from nouns, we may take two comparable construction types, the first being a noun phrase consisting of an attributive adjective and head noun (e.g. xorosij mal'cik ' good boy'), the second being a quantity phrase consisting of a head noun defining the quantity and a dependent genitive denning the entity being measured (e.g. stado ovec' flock of sheep').
The following criteria characterize the adjective in the attributive construction : (a) the adjective agrees in gender with its head noun, following a three-way masculine/feminine/neuter distinction (though only in the singular), e.g. xorosij mal'cik 'good boy', xorosaja devocka 'good girl', xoroseje okno 'good window'; (b) the adjective agrees in number with its head noun, on a singular/plural opposition, e.g. xorosij mal'cik ' good boy', xorosije maPciki ' good boys(c) the adjective agrees in case with its head noun, e.g. nominative xorosij maFcik, but dative xorosemu maPciku, instrumental xorosim mal'cikom; (d) many nouns have distinct accusative forms depending on whether or not they are animate, and adjectives agree with their head noun in terms of this distinction, e.g. inanimate accusative xorosij stol' good table', animate accusative xorosego mal'cika ' good boy', even though both stol and mal'cik are masculine singular. Head nouns in the quantitative construction have none of these properties. Thus we have stado ovec ' flock of sheep' where ovca ' sheep' is feminine, and stado gusej ' flock of geese' where gus' ' goose' is masculine. For number, we have tnassa benzina ' a mass of petrol' and massa ljudej ' a mass of people'. For case, we find that the head noun changes in case, but the dependent noun remains in the genitive, e.g. nominative stado ovec, dative stadu ovec, instrumental stadom ovec. Finally, the head noun does not change depending on the animacy of the dependent noun, cf. accusative massu ljudej ' mass of people' and massu karandasej' mass of pencils'.
On the other hand, the head noun of a quantitative construction has a number of properties that are not shared by the adjective in the attributive construction, as follows: (e) the head noun can vary in number independently of the dependent noun, e.g. stado ovec ' flock of sheep', stada ovec 'flocks of sheep'; (f) the head noun in the quantitative construction can take an attribute agreeing with it, e.g. xoroseje stado ovec 'good flock of sheep', where xoroseje is neuter singular nominative, agreeing with stado, while ovec is genitive plural; (g) the noun dependent on the head noun is invariably in the genitive, and if countable in the genitive plural - contrast the attributive construction under point (c), where adjective and head noun must be in the same case.
In terms of their adherence to the above seven criteria, we find that we can divide Russian numerals into several classes. First, the numeral' one' has all the properties of an adjective and none of those of a head noun: it
1
■1
can even agree in number, with pluralia tantum, e.g. odni (plural) noznicy 'one (pair of) scissors'. At the other extreme, the numeral million 'million ', and also all higher numerals, have all the properties of a noun and none of those of an adjective. Intermediate numbers have a varying number of adjectival and nominal properties, as illustrated in the table. In this table, A means that the numeral has the appropriate adjectival property, N that it has the appropriate substantival property; A/N means that either property can be used, A/(N) indicating that there is clear preference for adjectival behaviour; (A) means that the numeral has the adjectival property, but in a restricted form, in particular the numeral' two' has only a two-way gender opposition, distinguishing feminine dve from masculine-neuter dva; (N) indicates a similar restriction on a substantival property, as with the plural of sto ' hundred', which has only a few restricted uses. In the table, note that ' four' behaves like ' three', and that non-compound numerals between ' five' and' ninety' inclusive behave like' five'.
adjectival and substantival properties of russian numerals
(a)
(b)
(c)
(d)
(e)
(f)
(g)
odin	dva	tri	pjat'	sto	tysjaca	million
•r	'2'	'3'	'5'	'ioo'	'1000'	' 1,000,000
A	N	N	N	N	N	N
A	N	N	N	N	N	N
A	(A)	N	N	N	N	N
A	A/(N)	A/(N)	N	N	N	N
A	A	A	A	(N)	N	N
A	A	A	A	A	N	N
A	A	A	A	A	A/N	N
If we now ask the question whether Russian numerals are adjectives or nouns, it becomes clear that there is no straightforward answer, except in the case of'one' (adjective) and 'million' (noun): in particular, we cannot establish a cut-off point between adjectives and nouns, except arbitrarily, i.e. by deciding arbitrarily that we are going to take one, rather than another, of the seven criteria as definitive - and even then, some of the individual criteria are not definitive, as indicated by alternative entries separated by a slash or entries in parentheses. The situation is rather that we have clear prototypes, and a continuum separating those prototypes from one another, much as with colour terms, even though here we are clearly dealing with grammatical categories.
Actually, the continuum-like nature of the distinction between adjectival and substantival properties finds an even stronger manifestation in
104
SUBJECT
Russian numerals if we also take into account statistical preferences where alternatives are possible. For instance, after the numerals ' two', ' three and ' fouran adjective may be in either the nominative plural (as would be expected if these numerals were adjectives) or the genitive plural (as would be expected if these numerals were nouns). If one counts the occurrences of either possibility in text, it turns out that the preference for the adjectival type is greatest with 'two' and lowest with 'four', i.e. even as between adjacent numerals one can establish that the lower is more adjective-like than the higher.
In conclusion, definitions based on prototypes must be allowed as a possibility.
5-3 ERGATIVITY
In section 5.1, we posed a general problem for the syntactic analysis of any sentence, namely: what is the subject of the sentence? In view of the discussion of section 5.2, we can slightly reformulate that question. Implicit in the original question was that the question would have a clear-cut, discrete answer, i.e. a given noun phrase either would or would not be a subject. However, in terms of our characterization of subject as the intersection of agent and topic, and given that agent and topic are logically independent notions and need not coincide in a given sentence, it is clear that the answer to our question may well be less than clear-cut: it may be the case that a given noun phrase has certain subject properties, but not all, i.e. instead of simply saying that a noun is or is not a subject we will characterize it as being a subject to a certain degree. Similarly, it is possible that subject properties in a sentence will be distributed among several noun phrases, or at least between two, rather than all characterizing a single noun phrase. In many instances, then, it is as pointless to expect a clear-cut answer to the question * what is the subject of this sentence ?' as it is to expect a clear-cut answer to the question 'is Russian pjat' 'five' a noun or an adjective?' In the present section we will examine implications of this further, with particular regard to ergativity.
In section 5.1, we also posed the more specific question of identifying the subject of the ergative construction. In order to discuss this construction adequately, especially in terms of its similarities to and differences from the nominative-accusative construction, it is necessary to have a set of terms that is neutral between the two systems. The following is the set that we propose: The single argument of an intransitive predicate we will symbolize as S; this is clearly mnemonic for subject, and in general there is little or no controversy concerning the subject status in most intransitive
'í *
A h
SUBJECT
105
(single-argument) constructions across languages, so the mnemonically suitable symbol is also suitable in terms of its content. In the transitive construction, there are two arguments, and in order to avoid circularity we shall label neither of these with the symbol S. In the prototypical transitive situation, the participants are an agent and a patient, and this remains constant irrespective of the morphological or syntactic behaviour of the sentence in any individual language. We may therefore, starting originally with transitive predicates describing actions, label the agent as A, and the patient as P, so that in the sentence / hit you, or in its translation into Chukchi, irrespective of the case marking of the various noun phrases / will be A and you will be P. The labels are again clearly mnemonic, for agent and patient, respectively. However, the advantage of having arbitrary labels A and P rather than actually using agent and patient is that we can continue to use the arbitrary symbols even when we pass beyond prototypical transitive situations (i.e. actions) to other constructions in the language that have similar morphology and syntax. In English, for instance, the transitive verb see behaves morphologically and syntactically just like the action transitive verb hit, so that although in J saw you the pronoun / is not, in terms of semantic role, an agent, we can still symbolize it as A. A and P are thus syntactic terms, whose prototypes are defined in semantic terms.
In discussing examples (1) and (2) introduced at the beginning of this chapter, then, we can say that in (1) Chukchi yam and English / are Ss; in (2) Chukchi yamnan and English I are As, while Chukchi yat and English thee are Ps. Moreover, in English one case is used to encode S and A - a case of this kind is called nominative; and another case is used to encode P - a case of this kind is called accusative. In Chukchi, one case is used to encode S and P - a case of this kind is called absolutive; another case is used to encode A - a case of this kind is called ergative. The discussion thus far has related essentially to morphology, and we return to ergative-absolutive and nominative-accusative case marking in chapter 6. It is now time to turn to syntactic properties of subjects.
From the remarks made hitherto about subjects in English, it should be clear that English treats S and A alike as subjects for syntactic purposes, certainly for those syntactic points discussed so far, and indeed for most others. We can illustrate this by means of examples using coordination, in particular coordination of clauses that share a noun phrase in common and where that noun phrase is omitted in the second conjunct. If we try and conjoin sentences (8), (9), and (10), taking a transitive clause and an intransitive clause, in that order, then it is clear that we can conjoin, with omission of the second occurrence of the coreferential noun phrase, only (8) and (9), and not (8) and (10):
106 SUBJECT
The man hit the woman. The man came here. The woman came here.
The man hit the woman and came here. (= (8) + (9))
(8)
(9)
(10)
(11)
Even though sentence (11) contains no overt S for the intransitive predicate came here, it is absolutely clear to the native speaker of English that the only possible interpretation for this sentence is that the man came here, even though the alternative interpretation * the man hit the woman and the woman came here' would make perfect sense. In other words, in order to permit omission of a noun phrase from a second conjunct, English makes two requirements: (a) the semantic requirement that the two noun phrases be coreferential; (b) the syntactic requirement that the two noun phrases be either S or A. For syntactic purposes, English treats S and A alike, so subject in English means S or A.
We may contrast this situation with the situation that obtains in Dyirbal, with the translations of our three English sentences (8M10):
Balan cFugumbil bangul yararjgu balgan. (12)
woman-ABSOLUTiVE man-ERGATiVE hit
'The man hit the woman.'
Bayiyara baninyu. (13)
man-ABSOLUTiVE came-here 'The man came here.'
Balan dyugumbil baninyu. (14)
woman-ABSOLUTiVE came-here 'The woman came here.'
Balan dyugumbil baijgul yarangu balgan, baninyu.
( = (12)+ (14)) (15) ' The man hit the woman, and the woman came here.'
(In Dyirbal, nouns are usually accompanied by a classifier agreeing in class, including gender, and case with the noun; in the above examples, these are balan, bangul, and bayi.) Note in particular that (15) does not, and in Dyirbal cannot, have the meaning of English sentence (11): the two sentences in the two languages are crystal-clear in their interpretations to native speakers, though the interpretations happen to be different in the two languages. Dyirbal, like English, has two restrictions on coordination with omission of a noun phrase, but while the semantic restriction is as in
SUBJECT
107
■
English (the two noun phrases must be coreferential), the syntactic restriction is different: in Dyirbal, the coreferential noun phrases must be S or P. Thus for syntactic purposes, Dyirbal treats S and P alike, as opposed to A, so that in Dyirbal the appropriate grammatical relation is one that groups S and P together, in other words subject in Dyirbal means * S or P'.
Although it might seem that the syntactic difference follows the morphological difference between nominative-accusative morphology in English and ergative-absolutive morphology in Dyirbal (as can be seen by comparing examples (i2)-(i4))> »t is important to emphasize that this is not the case. In English, the syntactic identification of S and A proceeds even with non-pronominal noun phrases, which do not have a morphological nominative-accusative distinction. In Dyirbal, personal pronouns of the first and second persons happen to have nominative-accusative case marking, a fact to which we return in chapter 6, but this does not affect the ergative-absolutive basis of the coordination construction:
YJaďa rjinuna balgan.
I-NOMINATIVE yoU-ACCUSATIVE hit
' I hit you'
Yjaďa baninyu. I-NOMINATIVE came-here ' I came here.'
TJinda banin"u. you-NOMiNATlVE came-here ' You came here.'
Tjaďa ninuna balgan, baninyu.
' I hit you, and you/*I came here.'
(16)
(17)
(18)
(19)
We should also note that not all languages pattern either like English or like Dyirbal. In Chukchi, for instance, in coordinate constructions the omitted S of an intransitive verb can be interpreted as coreferential with either the A or the P of the preceding verb:
dittey -e talayvanen ekdk
father ergative he-beat-himson-ABSOLUTiVE
ankPam ekvetyPi-
and he-left 'The father beat the son, and the father/the son left.'
(20)
io8
SUBJECT
In Yidiny, as we saw in section 3.4, the preferred interpretation for an omitted S follows the morphology (coreferential with an absolutive or nominative noun phrase in the transitive clause), thus combining aspects of nominative-accusative and ergative-absolutive syntax, whereas Chukchi is completely neutral as between them, in this instance. One important point that the Yidiny material illustrates particularly clearly is that it is misleading to classify a language as being either ergative or not, rather one must ask: to what extent, and in what particular constructions is the language ergative, i.e. where does its syntax operate on a nominative-accusative basis and where does its syntax operate on an ergative-absolutive basis. In Yidiny, then, in the transitive construction, in some instances the A will have subject properties under coordination (example (44) of chapter 3), in other instances the P will have subject properties (example (43) of chapter 3), in yet other instances subject properties will be distributed between the two noun phrases (example (45) of chapter 3).
In common with many, but not all, languages, both English and Dyirbal have different syntactic means of encoding the same semantic roles, i.e. different voices. In English, for instance, we can take the transitive sentence (8), with the man as A and the woman as P, and rephrase it as a passive, an intransitive construction, in which the woman appears as S and the man as an oblique object (i.e. neither S, A, nor P):
The woman was hit by the man.
(21)
Since the woman is S of (21), and also S of the intransitive sentence (10), it is possible to coordinate these two sentences together, omitting the coreferential S from the second conjunct, to give (22), which has exactly the same meaning as Dyirbal sentence (15):
The woman was hit by the man and came here.
(22)
In Dyirbal, it is possible to take a transitive sentence like (12) (or, for that matter, (16)) and rephrase it so that 'the man' appears as an S, and 'the woman' as an oblique object, adding the suffix -nay to the verb. This kind of voice, whereby the A of the basic voice appears as an S, has in recent work on ergativity come to be called the antipassive voice:
Bayi yam bagun dyugumbilgu
man-ABSOLUTiVE woman-DATiVE balgalrjanyu. hit-ANTIPASSIVE
' The man hit the woman.'
(23)
SUBJECT
IO9
In Dyirbal, it is then possible to conjoin (23) with the intransitive sentence (13), of which' the man' is also S. For reasons that go beyond our concerns here, the only order in which this particular conjunction is possible is with the intransitive clause first:
Bayiyara baninyu, bagul d"t^umbilgu balgalrjanyu. ' The man came here and (he) hit the woman.'
(24)
Thus we see that one of the functions of different voices in languages is to redistribute subject properties: in English, to enable what would otherwise be a P noun phrase to have subject properties (as an S); in Dyirbal, to enable what would otherwise be an A noun phrase to have subject properties (as an S).
We may close the discussion of this section by recapitulating the main points, and driving them home with one further example. While the assignment of subject is clear in most intransitive constructions, especially those that are literally one-place predicate constructions, in transitive constructions we may find subject properties assigned either to the A, in which case we have nominative-accusative syntax, or to the P, in which case we have ergative-absolutive syntax. Some languages show strong preference for one or the other - e.g. English is largely nominative-accusative, Dyirbal largely ergative-absolutive - while other languages are more mixed. In Chukchi, the infinitive construction works on the nominative-accusative system, with omission of the S or A of the infinitive, with the suffix -(9)k:
a dmnan    yai tite
I-ergative you-ABSOLUTiVE sometime
mavinretyat ermetvi-k.
let-me-help-you to-grow-strong ' Let me help you to grow strong.'
Morysnan    ydt n&trevinretyat we-ERGATiVE you-ABSOLUTiVE we-will-help-you
rivl-zk  dmal?o yeceyot.
to-move all gathered-things-ABSOLUTivE 'We will help you move all the gathered items.'
(25)
(26)
In (25), the S of 'grow strong' is omitted; in (26), the A of 'move' is omitted. In the negative participial construction, with the suffix -IP on the
no
subject
subject
iii
verb in the participial form, the construction may be used to relativize either the S or the P of the participial clause, but not its A (unless the clause is antipassivized, as in (29), with relativization then effectively of the
S):
E -tipPeyne-ka -IP .{„
negative Sing        negative participle absolutive
rjevacqet raytayPi. (27)
woman-ABSOLUTiVE she-went-home 'The woman who was not singing went home.'
Iyara -yoP   -ka -IP -eta
now negative reach negative participle allative
enm -eta        manalqanmak. (28)
hill allative let-us-go 'Now let us go to the hill which (someone) didn't reach.'
En -aytat-ka -I? -a
antipassive chase negative participle ergative oaa       -k Paacek-a vinretarkaninet
reindeer locative youth ergative he-helps-them rjevacqetti. (29) women-absolutive ' The youth who does not chase the reindeer is helping the women.'
(Note that in (29) the object of the antipassive verb stands in the locative
case.)
5-4    SEMANTIC AND PRAGMATIC FACTORS
So far, we have not related splits between nominative-accusative and ergative-absolutive syntax to the distinction between those properties that are more properly correlated closely with agent, and those that are more closely correlated with topic, and it is to this discussion that we now proceed, although our discussion will necessarily involve only exemplification of a limited number of properties.
We may start off with subject properties that correlate more closely with agent properties. In many languages, in imperatives it is possible to omit reference to the addressee if that addressee is an A or an S, but not if it is a P; indeed, many languages have an even stricter requirement, namely that
1 b' 9 !
3
the S or A of an imperative construction must be second person (addressee), i.e. they only have second person imperatives. This can be illustrated for English by the examples come here! (i.e. you come here!) and hit the man! (i.e. you hit the man!), where it is possible to omit the addressee pronoun, in contrast to let/may the man hit you!, where it is not possible to do so. Interestingly enough, in Dyirbal, precisely the same constraint holds: despite the widespread prevalence in this language of syntactic constructions where S is identified with P, in imperative addressee deletion S is identified with A, as in English:
(Yjinda) bani. (3°)
you-NOMiNATlVEcome-here-iMPERATiVE
' Come here!'
(TJinda) bayiyara balga. (31)
you-NOMINATIVE        man-ABSOLUTIVE hit-IMPERATIVE
'Hit the man!'
The motivation for this distribution is not hard to find. For an instruction to be felicitous, the person to whom the instruction is addressed must have control over the resultant situation. In general, S and, especially, A are the participants who have most control over the situation, whereas P rarely has much control, so that it is more natural that the recipients of instructions should be encoded linguistically as an S or an A than as a P. Imperative addressee deletion simply provides a more compact means of expression for the more expected situation, i.e. addressees can be deleted when they are the more agentive S or A, but not when they are the less agentive P. This is thus a clear instance of a subject property that correlates with an agent property. Note that we are not saying that subject and agent are identical with respect to this property, or that the syntactic rule can be stated in terms of agents rather than in terms of subjects. For English, this is clearly untrue, since one can form passive imperatives where the addressee is not an agent but can be deleted, or where the agent is addressee but cannot be deleted (although the resultant sentences are very unnatural):
Be amazed by the world's greatest lion-tamer! Let/may this problem be solved by you!
(32)
(33)
What we are claiming is that this subject property has a high correlation with an agent property, and therefore the S/A identification is more natural, even in a language like Dyirbal where the S/A identification otherwise plays little or no role in the language.
112
SUBJECT
Moreover, we are not claiming that a language will necessarily have S/A identification for a subject property that correlates highly with an agent property, only that there will be a strong tendency for this to be the case (i.e. a universal tendency rather than an absolute universal). In Dyirbal, for instance, one might expect the same nominative-accusative syntax to carry over to indirect commands, deleting the S or A of the indirect command if coreferential with the recipient of the command. In fact, however, the A of an indirect command cannot be deleted in that form, rather the antipassive must be used, presenting that noun phrase as an S, which can then be deleted by the general rule allowing deletion of either an S or a P:
7]ana yabu gigan rjumagu
we-NOMiNATiVE mother-ABSOLUTiVE told father-DATivE buralrjaygu.
See-ANTIPASSIVE-INFINITIVE
' We told mother to watch father.'
(34)
(Note that the dative is one of the possible cases for the patient in the antipassive construction.) If the unmarked voice is used for a transitive verb in the infinitive, then only a coreferential P may be omitted, as in
(35):
VJad'a bayi yara gigan
I-NOMINATIVE man-ABSOLUTIVE told
gubirjgu mawali.
doctor-ERGATiVE examine-lNFlNlTlVE ' I told the man to be examined by the doctor.'
(35)
The example of imperative addressee deletion involved a natural identification of S and A, i.e. natural nominative-accusative syntax. We may now turn to an example of natural ergative-absolutive syntax. In Nivkh, there is a resultative construction, i.e. a construction referring to a state that has come about as the result of a previous event, using the suffix -yata. With intransitive verbs, this involves simply the addition of the suffix to the verb:
Anaqyo -a".
iron rust
'The iron rusted.'
Anaqyo   -yata -a". iron rust RESULTATIVE ' The iron has rusted.'
(36)
(37)
1
?
SUBJECT
113
(The verb-final suffix -a" is an indicator of finiteness.) If, however, we take a transitive verb, then a number of changes take place relative to the non-resultative form:
Umgu  t'us   t*a    -d\ (38)
woman meat roast
'The woman roasted the meat.'
Tus fa    -yata -d\ (39)
meat roast resultative ' The meat has been roasted.'
First, for the majority of transitive verbs in most circumstances, the A of the transitive verb must be omitted in the resultative construction. Secondly, the P of the transitive verb has the property that it conditions consonant-initial alternation in the verb (cf. the initial r*- of (38)), and the absence of such alternation in the resultative verb suggests that this noun phrase is no longer P. Whatever the precise details of the syntactic analysis, we can say that the resultative verb has a single argument, and that this argument corresponds to the S of a non-resultative intransitive verb, but to the P of a non-resultative transitive verb. In other words, S and P behave alike, as opposed to A.
The explanation this time is to be sought in the pragmatic structure of resultative constructions. Any such construction attributes a change of state to a certain entity. With intransitive predicates, the change of state is necessarily attributed to the S: in sentence (37), it is the iron that has undergone a change of state. With transitive predicates, although it is in principle possible for the change in state to characterize the A, as in John has climbed the mountain, it is more usual, especially with the prototypical transitive predicates describing an action involving a change of state, for the change of state to be attributed to the P. If we say the woman has roasted the meat, then we are necessarily talking about a change of state in the meat, and whether or not there is any change of state in the woman is simply left open. What Nivkh does is to grammaticalize this natural top-icalization of S or P in the resultative construction, by allowing only S or P to be expressed.
Again, we are not claiming that a language must make this identification in the syntax of resultative constructions. English, for instance, does not, so that the woman has roasted the meat is perfectly acceptable as the resultative of the woman roasted the meat. We are claiming, however, that languages will tend to show a bias towards ergative-absolutive syntax in resultative constructions.
114
SUBJECT
In many constructions, unlike imperatives and resultatives, there seems, a priori, to be no expected bias towards identifying S with either of A or P, for instance with coordination omission of noun phrases, and it is in these constructions that we find most variation across languages: with coordination, for instance, English has nominative-accusative syntax, Dyirbal has ergative—absolutive syntax, Yidiny has both, and Chukchi has neither. However, our present understanding of the cross-language distribution in such cases suggests that nominative-accusative syntax is in fact more widespread than ergative-absolutive syntax, and we might ask why this is so. Moreover, if we take a piece of natural nominative-accusative syntax like imperative addressee deletion, there are few or no languages that go against it by having ergative-absolutive syntax. However, if we take a piece of natural ergative-absolutive syntax, like resultative constructions, then we do find a wide range of languages that go against the natural syntax by having nominative-accusative syntax. In other words, there seems to be a general bias in language, interacting with naturalness of identification of S with A or P, towards nominative-accusative syntax. This general bias, in turn, has an explanation: as we shall see in a slightly different context in chapter 9, humans have a strong tendency to select more agentive entities as topics of discussion, which means that there is a natural correlation between agent and topic: other things being equal, one would expect agent and topic to coincide. The notion of subject then simply reflects the gram-maticalization of this expected coincidence, and explains why so many languages do have a grammatical relation of subject definable in its core as the intersection of agent and topic, whereas few languages similarly define grammatical relations reflecting the intersection of, say, patient and topic.
While preference for equating agent and topic does seem by far the most prevalent identification across languages, there are some languages that do not show this particular identification. In Dyirbal, for instance, subject properties that are not agent-bound, and even some of those that are (cf. indirect commands), adhere to the P rather than to the A. In Dyirbal, then, it seems that agentivity is virtually irrelevant to the establishment of subjecthood, preference being given to P, i.e. P is more natural as a topic than A. In a number of Austronesian languages, especially in Philippine languages, a similar, though somewhat less extreme, situation seems to obtain, with some syntactic processes being conditioned by the agentivity (semantic role) of the noun phrases involved, and other syntactic processes - including most of those where no bias would be expected a priori - being controlled by the topic, with preference for a patient rather than an agent to be topic. The following examples are from Tagalog.
If we take the situation of someone borrowing money from a bank, then in Tagalog if the P is definite, it must be topic, irrespective of the defi-
subject
115
niteness of the other noun phrases, as in (41). Only otherwise is it possible for the A to be topic, as in (40):
Humiram siya      ng pera   sa      bangko. (40)
borrowed-ACTiVE he-TOPic P money dative bank ' He borrowed money from the bank.'
Hiniram niya ang   pera   sa      bangko. (41)
borrowed-PASsrvE he-A topic money dative bank ' He borrowed the money from the bank.'
If we embed this construction under a verb meaning' to hesitate', then this is a construction which, a priori, favours S/A identification - one can only hesitate about something under one's own control - and here Tagalog allows deletion of the A, irrespective of whether it is topic or not:
Nagatubili        siya       -ng humiram ng pera sa bangko.
hesitated-ACTiVE he-TOPic
' He hesitated to borrow money from the bank.'
Nagatubili siya-ng hiramin ang pera sa bangko. ' He hesitated to borrow the money from the bank.'
(42)
(43)
(In the last two examples, the suffix -ng is a clause-linker; hiramin in (43) is the nonfinite equivalent of hiniram.)
If, however, we take a construction that is neutral as between identification of S with A or P, then Tagalog treats the topic as subject. For instance, in Tagalog relative clauses, the noun relativized can only be topic of the relative clause. Compare the following sentences:
Bumili ang   babae   ng baro. (44)
bought-ACTiVE topic woman P dress 'The woman bought a dress.'
Binili ng babae   ang   baro. (45)
bought-PASsiVE A woman topic dress 'A/the woman bought the dress.'
If we want to relativize ' woman', then the active construction of (44) must be used, but if we want to relativize ' the dress', then the passive construction of (45) must be used; no other alternatives are possible:
Ii6
SUBJECT
Iyon ang   babae-ng bumili ng baro. (46)
that topic woman  bought-ACTiVE P dress ' That is the woman who bought a dress.'
Iyon ang baro-ng binili ng babae. (47) 'That is the dress that the woman bought.'
To conclude this chapter, we note that treating subject as a diffuse, rather than a discrete, notion, while perhaps seeming at first to weaken the notion of subject, does in fact provide us with a powerful tool which, in conjunction with independently established correlations with agent and topic properties, enables us to describe in a unified way, with a large measure of explanation, disparate phenomena across a wide range of languages.
NOTES AND REFERENCES
The idea of defining prototypical subject as a multi-factor concept is developed initially by Keenan (1976b), although I do not use his classification of properties here. The strongest criticism of this approach comes from Johnson (1977a), but unfortunately Johnson begs the question by assuming that a definition must be in terms of necessary and sufficient conditions. My discussion of the continuum (* squish *) from adjective to noun in Russian numerals is based closely on Corbett (1978).
The discussion of ergativity in section 5.3 is based on Comrie (1978b). Very similar ideas, though with certain differences in terminology, emphasis, and concept, are given independently by Dixon (1979); note in particular that Dixon uses O for my P, uses subject for a natural grouping of S and A, and uses pivot for a grouping of S with A or S with P in a particular language. The D yirbal examples derive originally from D ixon (1972), though they reappear in almost every subsequent discussion of ergativity; example (35) was expressly collected for me by R. M. W. Dixon, who thereby earns a particular debt of gratitude, the more so since this example requires slight modification of pages 128-9 of Dixon (1979). Numerous studies on ergativity are gathered in Plank (1979); the Chukchi examples are from the contributions to this volume by Comrie (1979c, 226,227,229) and Nedjalkov (1979,242).
Splitting subject properties between agent (role) and topic (reference) properties is developed, especially for Philippine languages, by Schachter (1976,1977); the Tagalog examples are taken from the second of these. The Nivkh examples are from Nedjalkov et al. (1974). Factors controlling the distribution of nominative-accusative and ergative-absolutive syntax are discussed by Moravscik (1978b). The discussion of imperative addressee deletion is based on Dixon (1979,112-14), that of resultative constructions on Comrie (1981a).