You are on page 1of 37

Linguistic Society of America

The Social and Structural Dimensions of a Syntactic Change


Author(s): Anthony J. Naro
Source: Language, Vol. 57, No. 1 (Mar., 1981), pp. 63-98
Published by: Linguistic Society of America
Stable URL: http://www.jstor.org/stable/414287
Accessed: 21-12-2015 14:10 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

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

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
THE SOCIAL AND STRUCTURAL DIMENSIONS OF A SYNTACTIC
CHANGE

ANTHONYJ. NARO
Pontificia Universidade Catolica and Universidade Federal,
Rio de Janeiro
This paperproposesa model of syntacticchangebased on a quantitativestudy of the
rule of subject/verbagreementin spoken BrazilianPortuguese.Among speakersof the
lower socio-economiclevels, this rule is currentlyundergoinga process of elimination
from the grammar.The model postulatesthat syntactic change starts at a point where
surface differentiationbetween the old and new systems is zero (or nearly so); later it
spreadsthroughoutthe languagein inverse proportionto the degree of saliency of the
surfacedifferencesbetweenthese systemsin each particularenvironment.Thus 'natural'
clusters of linguisticfeaturesarise only gradually.
1. INTRODUCTION. Since syntactic change produces a finite number of dis-
crete variants, it cannot be a process of gradual replacement of minutely dif-
fering forms, operating unobserved over time.' In general, there is no surface
continuum of realizations to be dealt with; rather, a certain (usually small)
number of non-overlapping distinguishable variants can be identified and cat-
egorized. This circumstance, which makes syntactic change an ideal site for
observation of linguistic change-in-progress, has not been utilized in diachronic
research. Even in recent times, historical syntax has been heavily inclined to
a change-over-the-centuries methodology which makes it impossible to study
the social and structural dimensions that shape the initiation and diffusion of
a change; this has led to a distorted view of the process, in which only the
endpoints are seen. Furthermore, since the usual endpoints are relatively sta-

'The centralnotion of syntactic diffusionalong the axis of saliency, upon which the structural
portion of this paper is based, was sketched in Naro & Lemle 1976. In that work, only three
speakers,with a total of less than ten hours of recordedspeech, were studied. Later, much more
extensive tests, involving twenty speakers and 140 hours of speech, were made as part of the
'CompetenciasBasicas do Portugues'researchproject(sponsoredjointly by the FundacaoMo-
vimento Brasileirode Alfabetizafco and the Ford Foundationoffice in Brazil, administeredby
MiriamLemle of the Museu Nacional, UniversidadeFederaldo Rio de Janeiro).The results of
these tests appearin the final report of the project (Lemle & Naro 1977:17-50).In the present
paper, I use the same data base as in earlierversions, with only minorcorrectionsand additions;
but I reportsomewhatdifferentresults because of the inclusionhere of social variablesand some
changes in the morphologicalcategorization.I am gratefulto Lemle for allowingme to use the
data gatheredduringher project, as well as for many insightfuldiscussions duringthe course of
the work; without her keen intuitionsinto the workingsof Portuguese,this researchwould not
have been possible. I am likewise gratefulto GregoryGuy, to whom I owe my initiationinto the
use of statisticaltechniquesin linguistics,as well as many hours of fruitfulcollaboration;and to
Shana Poplack and David Sankoff for comments and suggestions that have profoundly affected
the final form of this paper. Michael Stanton of the Departamento de Informatica of PUC/RJ has
patiently helped me learn how to communicatewith a computerin a mannerto which it will
respond more or less appropriately, even if not always willingly and quickly. All calculations
reportedhere were carriedout on the IBM/370-165installedat the Rio Datacentro,to which PUC/
RJ generously provided me free and unlimitedaccess. Differences between this and previous
versions reflect what I consider to be improvements or corrections.
63

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
64 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

bilized standardlanguages, a deceptive appearanceof categoricalregularityis


produced.
The goal of this article is to present a detailed quantitativestudy of one
particularon-going syntactic change-the loss of 3rd person subject/verbcon-
cord in modern BrazilianPortuguese-and to suggest (on the basis of this and
other evidence) a general hypothesis about syntactic change.
2. SUBJECT/VERBCONCORD.
In standard Portuguese, a finite verb must agree
with its subject, whether this latter element be present or deleted, preposed
or postposed.2 Some prescriptivemanuals do countenance a few exceptions
to categorical agreement, such as optional singularor pluralwith subjects of
the type um grupo de manifestantes 'a group of demonstrators';but in the
standardwritten language and in the normal speech of the educated classes,
the rule of subject/verbconcord is nearly categorical.
Withinthe noun phrase, numberagreementalso operates obligatorilyin the
standardlanguage.As we shall see in more detail in ?9, below, this is not true
of the spoken language-where, usually, only one element of the NP shows
an explicit plural inflection. But in contrast to some varieties of Spanish, the
presence of at least one explicit plural mark in the NP is nearly categorical,
even in the most relaxed styles of the lowest socio-economic levels. This
circumstance facilitates the identification of plural subjects. Generic NP's,
which are more common in Brazilian Portuguesethan in English, are gram-
maticallysingulareven in the standardlanguage;e.g. Feijao preto ta em falta
'Black bean(s) is out of stock.' For this reason, they will not be considered
here.
In the verb, pluralizationis indicatedby meansof the person-numbersuffix-
which, in the 3rd person, takes the basic forms -u or zero for the singular,and
-un or -n for the plural(Camara1972:126-7).On the surface, this suffix always
occurs fused with the tense-mood suffix (and sometimes even with the theme
vowel or root). Thus the actually occurringinflection can be complex; and it
may show additionalchanges in form, resultingfrom the fact that the tense-
mood suffix may also be dependentupon number(Camara1972:125).To give
a more complete view of the facts, I have listed in Table 1 the majorpossible
surface realizationsof the complete inflections for the 3rd person, along with
examples and glosses. Both the standardorthographicrepresentationsand the
phoneticforms of the inflectionsare indicated;the latterare furthersubdivided
into popularand standardvariants. The popularvariantstend to occur more
frequentlyin the speech of the lower socio-economiclevels, while the standard
variantsare found more frequentlyin the speech of the highersocio-economic
levels, on radio and television, etc.
A few irregularverbs for which singularand pluralforms are homophonous
(or nearly so) in the present have been excluded from the data base in that
tense (e.g. ter 'he has', tem 'they have', where the accent is merely ortho-
2
In fact, even an infinitivemust agree with its subjectundercertaincircumstances.However,
since inflectedinfinitivesconstitutean areaof debatein bothprescriptiveanddescriptivegrammar,
and occur very infrequentlyin spontaneousspeech, they have been left out of consideration.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 65

CLASS ORTHOGRAPHIC FORM POPULAR STANDARD

I. PRESENT
1. 1st conjugation sg.: falA 'speaks' ['a] ['a]
(regular) pl.: falAM'speak' ['fi, 'u, ,]
2. 2nd & 3rd conj. sg.: partE'leaves' ['i] [li]
(regular) pl.: partEM'leave' [ I] ['ey]
3. 2nd & 3rd conj. sg.: faz 'does' 0 0
(-r or -z) pl.: fazEM 'do' [-i,:i]
4. Monosyllables sg.: dA 'gives' [a] [a]
p.: dAo 'give' [aw] [aw]
II. PRETERIT IMPERFECT
All verbs sg.: falavA 'spoke' ['a] ['a]
pl.: falavAM 'spoke' ['u, 'u, 'a] [Iaw]
III. FUTURE
All verbs sg.: falarA 'will speak' [-a] [-a]
pl.: falarAo 'will speak' [-aw] [-aw]
IV. PRETERIT PERFECT
1. 1st conjugation sg.: falou 'spoke' [-6] [-6, -6w]
(regular) pl.: falARAM'spoke' [-rfi, -aru] [-drAW]
[-araw]
2. 2nd conjugation sg.: aprendEu'learned' [-ew] [-dw]
(regular) pl.: aprendERAM'learned' [-eru, -dru] [-draw]
3. 3rd conjugation sg.: partiu 'left' [-iw] [-fw]
(regular) pi.: partIRAM 'left' [-frfi,-fru] [-fraw]
4. stress-shifting sg.: trouxE'brought' [li] [-ri]
(irregular) pl.: trOUXERAM 'brought' [-rfi, -dru]
5. radicalchanging sg.: fez 'did' 0 0
(irregular) pi.: fiZERAM 'did' [-erufi,-eru] [-eraw]
TABLE I

graphic).Also excluded from considerationare several constructionsin which


it is not clear that a surface subject is present. For example, accordingto the
prescriptivegrammarsused in Brazilianschools to teach the most prestigious
forms of the standardlanguage, the copula in time expressions should agree
with the NP that denotes the hour; thus Era (sg.) uma hora quando Giselle
chegou 'It was one o'clock when Giselle arrived';Eram(pl.) duas horas quando
Ant6nio chegou 'It was two o'clock when Anthony arrived.' However, in this
construction,the time mightbe considereda predicateelement, and the subject
as empty. Similarlyin clefts: Eram (pl.) tres mofas que vimos 'It was three
girls that we saw.' Since these types of sentence show great variationeven in
the standard,because of varyingsyntacticconstruals,it would be inappropriate
to include them in a study of agreementas such.
Finally, this study is limited to verbal agreementwith true plural subjects,
disregardingimpersonalsubjects of the type in Dizem que nao ter mais vaga
'They (people) say that there is no room left.' Earlierwork (e.g. Naro & Lemle
1976) showed that these impersonal subjects impose a higher rate of plural
markingthan do true plurals. However, they are excluded here because of
several peculiarities in behavior that are not directly relevant to the central
theme. The data thus excluded amount to about 14.6% of the total corpus
originallycollected.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
66 LANGUAGE, VOLUME57, NUMBER 1 (1981)

UNDERSTUDY.As mentioned at the outset, subject/verb


3. THEPOPULATION
concord is nearly categoricalin the educated classes. The only circumstances
under which large-scale variationoccurs in this group is when the subject is
postposed to the verb. Thus it was decided to limitthe study to people of lower
socio-economic levels, where variation is massive. Furthermore,since the
principalsponsor of the research was MOBRAL(the Braziliangovernment's
literacy agency), all but one of the seventeen people whose speech was to be
studied were chosen from among studentsin that agency's classes. At the time
of the interviews, they had completedabouthalf of the six-monthcourse. Some
were repeats, havingfailed or droppedout of the previous cycle-but all were,
for practical purposes, still essentially illiterate.3 In order to eliminate the
geographicvariable, all the speakers chosen were natives of the state of Rio
de Janeiro, or had moved there as infants. Eight of them had spent their
formative years in the surroundingareas outside the city proper, while nine
were natives of the city itself. The speakers included eight females and nine
males. They were concentratedin two disjoint age groups (roughly, 25 years
or younger and 35 years or older) in order to provide a clear test of the dis-
tributionwith respect to this variable;but there are only six speakers in the
older group. The principalfacts regardingeach person are listed in Table 2.

NAME SEX AGE AREA PROFESSION


VAN F 15 S maid
JOM M 16 C unemployed
LID M 16 C errandboy
SID M 16 C oddjobs
HEN M 17 S helper(laundry)
PAU M 19 S errandboy
NAT M 20 C (unknown)
SON F 22 C maid
FAU F 23 S custodian(school)
IZE F 23 C maid
MAR F 26 C maid
GCL M 37 C janitor
UBI M 42 C janitor
ILD F 43 S maid
ELV F 45 S maid
MAD F 45 S maid
CAN M 54 S porter(club)
TABLE 2.
(M = male; F = female; C = city; S = state [environs])

Since syntactic data occur with a relativelylow frequencyin naturalspeech,


it was decided to record about seven hours of speech for each person, in
roughly hour-long interviews. However, NAT was interviewed for only one
hour. The interview site, the interviewer,and topics underdiscussion were all
variedas much as possible in an attemptto elicit the full rangeof each person's
3
GCL, the one non-MOBRALstudent, is also illiterate.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 67

potential behavior. Most of the interviews were carriedout by university stu-


dents, using a questionnairedesigned primarilyto elicit informationon the
speaker's background.Several topics relatedto popularculturewere also sug-
gested, but the interviewershad instructionsto switch to any other topic which
the interviewee seemed interested in discussing. In general, interviewer and
interviewee were of the same sex. In some cases, one of the people in the
sample interviewed another, and there were even attempts to obtain secret
recordings. The total quantity of relevant data obtained varied greatly, how-
ever, rangingfrom 42 to 818 tokens4
In a given sentence with plural subject, a
4. MODELSANDMETHODOLOGY.
variety of forces (sometimes called 'variableconstraints')operate simultane-
ously to make the application of the agreement rule more or less probable,
resultingin the variableappearanceof a pluralinflectionon the corresponding
verb. It is precisely the identificationof such forces and the evaluationof their
relative importancefor rule applicationthat constitute the central topic under
investigationhere.
For purposes of analysis, variable constraints may be classified into two
main types: linguisticand social. The first is dependentupon relationsthat the
item under study bears to others within the grammaticalsystem; the second
depends upon the relations of the speaker to the surroundingcommunity. I
shall postpone the study of social dimensions until ?6.
The basic determinantof the linguistic constraints can be deduced from
comparison,in their spoken forms, of pairsof sentences like Eles come banana
cor mel 'They eats (sg.) bananaswith honey' and Eles e do Nordeste 'They
is (sg.) from the Northeast.' The first, despite its lack of agreement,does not
sound like an unusualutterance,while the second attractsattentionmuch more
strongly, and even sounds somewhat odd to many native speakers. Note, in
this regard, that if the verb of the first example were pluralized,the resulting
form comem 'eat' would not be very differentphoneticallyfrom come 'eats':
in the popularvariantsoverwhelminglypreferredin the corpus under study in
this article, the difference is merely nasalizationof an unstressed final [i]. In
the second sentence, however, the singulare 'is' is totally differentfrom the
plural sao 'are'. In other words, the second opposition is more SALIENT or
noticeable than the first because the phonetic segments that realize the op-
position are more different. On the basis of these roughobservations, a multi-
leveled scale of phonetic saliency of the morphological realization of the
singular/pluralopposition was set up, and was postulatedas one of the struc-
tural constraintgroups.
Along these same lines, lack of agreementis more salient when the plural
subject occurs immediatelypreposed to the correspondingverb than when it
is separatedfrom the verb by interveningelements (adverbs, relative clauses

4 A 'relevant' token is defined as


any occurrence of a verb with a plural subject, whether or not
the verb actually has a plural inflection. Furthermore, it is not required that the plural subject be
explicit in the sentence.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
68 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

etc.) which physically and temporallyseparatethe dependentpluralverb from


the determiningpluralsubject. Even less salientis the case in which the subject
follows the verb.
The effects of both the phonetic and the positional scales of saliency will be
treated in greater detail below. I refer to these scales as 'constraintGROUPS'
because they are set up in such a way that at'most one (in fact, exactly one)
category of each group is applicableto a given verb.
4.1. UNIFORMITY
OF EFFECT.As we have already seen, any variable phe-
nomenon is under the influence of a number of simultaneousvariable con-
straints. Even if the constraintsrelevant to the phenomenonunderstudy were
successfully identified, a seemingly unresolvabledilemma would remain:we
wish to know the individualcontributionof each constraintto the total situation,
but we can never empiricallyobserve only ONEconstraintin action. In the real
world, a certain numberof constraints are always operatingin any given en-
vironment. Thus, in his classic study of the deletion of the copula in Black
English, Labov 1969establishedthat this rule was constrainedby the following
factor groups: (1) the natureof the subject (pronounor full NP); (2) the nature
of the predicate(NP, locative, predicateadjective, gerund, or gonna); and (3)
the phonetic type of the preceding segment (sibilant, other consonant, or
vowel). But the only available empirical data are of this type: With an NP
subject ending in a vowel and with a gerund in the predicate, deletion occurs
with frequencyf. How, then, can we evaluate the contributionof the fact that,
say, a gerund is present, if there is always some effect other than that of the
gerund in any environmentcontainingthis element?
Note first that, merely by asking this sort of question, we are makingthe
implicit assumption that it makes sense to talk about THE effect of a given
element; i.e., we are assuming that whatever the effect of a following gerund
may be, it is the same in all environments,regardlessof whatever other ele-
ments may be present. This is, in general, not a self-evident truth. A simple
example to the contrary can be given from NP agreement(Scherre 1978;dis-
cussed in greater detail in ?9, below). Here there is a strong tendency for the
plural morphemeto be absent in all elements of the phrase following a non-
agreeingelement (i.e., for 'my first friends', as minhasprimeiraamiga is more
common than as minhaprimeiraamigas, where -s is the pluralmorphemeand
as is the definite article). However, the effect of lack of -s is just the opposite
of this if it occurs in the first element of the phrase (i.e., a minhas amiga is
more common than a minha amigas, althoughboth are rare). Thus it makes
no sense to speak simply of 'the effect of a preceding uninflected element'.
This problembecomes even more serious when social factors are introduced,
since it is quite possible for the effect of a certainfeatureto vary among social
groups (Kay 1978, Kay & McDaniel 1979).
When situations of this type are encounteredin the course of research, two
options are available: (1) to reformulatethe variable constraint that violates
the uniformityof effect assumption, or (2) to carry out separate analyses for
each domain in which this assumptionis satisfied.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 69

Given uniformityof effect, at least within acceptable limits, we still need a


model of the way the effects of each constraintcombine to produce the total
effect of the environmentin which they are jointly present. The necessity for
such models has, of course, arisen in other empirically-basedsciences; the
statistical solutions already found can, fortunately,be adaptedfor linguistics.
MODELS.Basically, there are three approaches. The
4.2. MATHEMATICAL
first ignores the whole problemby assumingthat, if a sufficientnumberof data
involve a given constraint c(i) in a wide variety of combinationswith other
constraints, the over-all result will be for the individualeffects of the latter to
cancel each other out. Under this assumption,it would suffice simply to note
the frequency of rule applicationover all tokens in which c(i) is present (re-
gardless of whatever other constraintsmightalso be present), and then to take
this raw frequency as a measure of c(i)'s effect on rule application.This sim-
plistic approachdoes, in fact, workin manycases; but since its success depends
more upon a felicitous distributionof the data than upon any factor that can
be controlled, it would not be prudentto base any importantconclusions on
the results of this model alone. Furthermore,skewed distributionsare usually
unavoidablein naturallinguisticdata, since we have no way of controllingwith
precision the distributionof environmentsin the flow of speech.
The second approachpostulates that the effects of each factor present in a
given case cumulate additively, i.e. that the total effect of an environmentis
the SUMof the individualeffects of the parts. Since we are here dealing with
relativefrequency of rule application,expressed as a percentage,this amounts
to assumingthat each variableconstraintis associated with a certainquantity-
also stated as a percentage,either positive or negative, that measuresthe effect
of this particularconstraint. According to this model, the experimentallyde-
terminedpercentageof rule applicationin any given environmentshould tend,
within the limits of statisticalfluctuation,to be the sum of these percentages.
To provide a non-arbitrarystandardagainst which to measure these effects,
the grand mean, or over-all frequency of rule application, is added into the
sum; the effects are then viewed as deviations from the grandmean resulting
from the presence of the factors. In symbols, this model is:
(1) f(t) = f(O) + f(l) + ... + f(n)
Here f(t) is the total frequency of applicationin a given context, f(O) is the
grand mean, and f(l) ... f(n) are the effects of the n factors present in the
context, stated as deviations from the grandmean.
A statistical technique known as multipleclassificationanalysis (MCA) has
been used to estimate the factorialdeviationsfromthe rawrelativefrequencies,
by adjustingthe raw deviationof each factor for the effects of the other factors
in such a way as to satisfy formula I as closely as possible. The result is called
an 'adjusteddeviation'. MCA is usually used in conjunctionwith analysis of
variance, the purpose of which is to determinewhether the postulatedfactors
have a statisticallysignificanteffect. Until recently, neitherof these techniques
was usable in areas of study where experimentscould not be predesignedand
controlled, because their mathematicsrequiredthat the total numberof tokens

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
70 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

in each environmentbe the same.5 Fortunately,improvementsnow allow this


restrictionto be relaxed under certain conditions. One advantageof MCA is
that it provides an explicit measure of the relative importanceof each GROUP
of factors. In non-technicalterms, this statistic, called a 'partialbeta', indicates
(but does not measure exactly), when squared, the proportionof the total
variation attributableto the group to which it refers. Thus the betas vary
between zero and one, and a higher value indicates a more importantgroup.
In linguistics, the use of an additive model was first suggested by Labov
1969;but this suggestion was never implementedin any meaningfulway, and
Labov himself abandoned the additive model in his 1972 revision. The pre-
mature demise of this model was apparentlycaused by a combinationof the
practicaldifficultiesmentionedabove (althoughthe first version of the OSIRIS
package, available in 1969, contained an appropriatecomputer programthat
did not require equal cell counts) with a misinterpretationof the underlying
assumptionsof the model. In the first place, Labov was forced to make a series
of unrealistic assumptions (labeled the 'postulate of geometric ordering') in
order to rule out the possibility of the sum in formula 1 becominggreaterthan
100%or less than 0%-or, in terms of relative frequencies (ratherthan per-
centages), greaterthan 1or less than0. However, thereis no reasonto constrain
the model in this way: if, for a particularenvironment, the factorial effects
estimated on the basis of all the data happen to add up to 110%,when the
empiricallydeterminedrate is 95%, this should simply be counted as a failure
to model the data correctly, in a way parallelto a predictionof 90%when the
correct figure is 75%. In both cases, the proportionof variationaccountedfor
will diminishand the errorwill increase. Unfortunately,later researchers(Kay
& McDaniel, 160, 166-8) imposed a host of even more arbitraryand unnec-
essary assumptions, since Labov's original 'postulate' turned out not to be
sufficient for the purpose for which it was intended. All these unfortunate
assumptionshad the result of makingthe additive model quite unattractive.
One otheraspect of the additivemodel shouldbe mentioned.We have already
noted above that the effect of each feature must be assumed to be constant,
as a prerequisiteto any discussion that makes sense. In the additive model,
this has the consequence of requiringthat each constraintadd (or subtract)an
equal amountto the frequencyof rule applicationin all environmentsin which
it is present. Consider an example from Labov's data (as reportedby Kay &
McDaniel, 165):if contractionof the copula in the environmentof a preceding
vowel occurs 90%of the time before a verb, but only 77%before a non-verb,
the effect of the following non-verbis - 13%.If, in the environmentof a pre-
ceding non-vowel, the frequency of contractionwhen a verb follows is 75%,
the model thus requires that this frequency be 75% - 13% = 62% when d non-
verb follows. To the extent that real data do not conformto such requirements,
there is said to be INTERACTION
among the constraints; i.e., at least one of the

I.e., if a speakerhappenedto uttern tokens of 1st conjugationverbs in the presenttense with


immediatelypreposed plural subjects, he/she would have to utter exactly the same numberof
tokens underidenticalcircumstancesin the preteritperfect, and so on.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 71

constraintshas an effect that varies accordingto the environmentin which it


is found. In Labov's data, the empirically determinedfrequency was 47%,
ratherthan 62%;so apparentlysome interactionwas involved.
If interactioneffects are strong, the model in formula 1 will not be realistic,
and the MCA based on it will be meaningless.Analysis of variancecan be used
to test for statisticalsignificanceof interaction;but statisticalsignificancedoes
not necessarily imply linguisticsignificance,since even minuteinteractionwill
be statistically significantwith large amounts of data. The only practicalway
out of this difficulty is for interactioneffects, if present, to be lumpedin with
the general error term. If the proportionof variation accounted for is none-
theless sufficiently high, the relegationof interactionto erroris justified. This
expedient has been used in the additive results presented below-which rep-
resent, as far as I know, the first successful use of the additive model in a
linguisticstudy. This model will, however, always be used in conjunctionwith
other models that have already proved their usefulness in previous linguistic
investigations.
The third approach utilizes a multiplicativemodel for the combined effect
of the component features of an environment.In this case, it is customaryto
speak of relative frequencies in the strict sense ratherthan percentages (i.e.
withoutmultiplyingby 100),andto speakof the factorialeffects as probabilities.
This is because the first multiplicativemodels proposedin linguisticsfollowed
the usual law of independentjoint effect of probabilities(Cedergren& Sankoff
1974:338-9):
(2) p(t) = p(O) x p(l) x *. x p(n)
Here the p(i) (for i from 0 to n) are in the interval between zero and one;
p(t) representsthe probabilityof rule application,correspondingto the empiri-
cally determined relative frequency up to statistical fluctuation, for an envi-
ronment that consists of factors 1 to n; and p(O) is an 'input' probability that
plays the same role as the grandmean in formula 1. The factorialprobabilities
are written as p(l), p(2) etc.
Note that this model has the unfortunateconsequence that the combined
effect of highprobabilities(say, 0.8 and0.9) is lower thantheirindividualeffects
(0.8 x 0.9 = 0.72). Furthermore,this model also requiresa series of unmo-
tivated and unattractiveassumptions, such as setting equal the probabilities
associated with all factors that happen to occur in the environmentin which
the rule is most frequently applied. And, although it is true that 2 formally
defines independence of applicationprobabilities,we mightjust as well have
postulated independence of non-applicationprobabilities, replacing p(i) by
(1 - p(i)). This introducesa furtherelement of arbitrariness,since it gives us
two equally well-motivatedmultiplicativemodels from which to choose. Fur-
thermore,the basic reason for adoptionof 2 disappears.For these reasons, the
simple multiplicativemodel has proved inadequateand has been replaced by
the so-called logistic model. Developed by statisticiansfor use in other areas,
this model was introducedinto linguistics by D. Sankoff 1975, and has been
used with great success in a numberof research projects. The basic formula

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
72 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

is:
p(t) (O) p(l) p(n)
(1 - p(t)) (1 - p(O)) (1 - p(l)) (I - p(n))
The symbols here have the same interpretationas in 2, and continue to be
called probabilities-although the appropriatenessof this terminologyis ques-
tionable, since the logistic model is motivatedmerely by the fact that it works,
ratherthan by any abstractconsiderationsof probabilitytheory. Representing
the entire right-handside of 3 by p(R) and solving for p(t), we find:
(4) p(t) = p(R) / (1 + p(R))
From this it is easy to see that, as p(R) becomes large, p(t) tends toward the
numberone; and as p(R) becomes small, p(t) tends toward zero. Since p(R),
in turn, is a product of terms of the form
(5) p(i) I (1 - p(i)),
p(R) will increase for each such term largerthan 1, and decrease for terms less
than 1. These terms themselves will be greaterthan 1 if p(i) is greaterthan 0.5,
and less than 1 if p(i) is less than 0.5. If p(i) is exactly 0.5, the corresponding
termbecomes 1, and makes no contributionto the product.Thus, in the logistic
model, the 'probabilities'can be classified as favoring, inhibiting,or indiffer-
ent-depending on whether they are greater than, less than, or equal to 0.5.
Furthermore,accordingto this model, the combinedeffect of favoringfactors
(with values of 0.9 and 0.8, say, as in the example given above) is more highly
favoring than that of each of the individualfactors (the result is 0.97 in our
example).
In the mathematicalsense, the data and the model in 3 are not sufficient to
determine uniquely the values of the p(i)'s; only a relationshipamong these
probabilitiesis determined.In order to calculate a unique numericalvalue for
each p(i), we must add one mathematicalconstraintper factor group. Although
this constraint is, from the mathematicalviewpoint, strictly arbitrary,it can
be chosen in such a way as to facilitateinterpretationof the resultantnumbers.
In any case, since all logistic results obey the same constraint,it will ordinarily
not be necessary to consider its consequences. The constraintcurrentlyused
is to requirethat the average of the terms in 5 for each factor group be zero.
For groups containingonly two factors, this has the consequence of requiring
that the logistic probabilitiessum to 1.
In order to estimate the values of the p(i)'s satisfying 3 (and the constraint
just mentioned)on the basis of the availabledata, a statisticaltechniqueknown
as maximumlikelihoodestimationis employed. For any given model, the like-
lihood statistic itself is a measure of how likely it is that a particularset of
data has been generated by the model. The estimationprocedure maximizes
this likelihood for the set of data under study, in accord with the model being
used.
Within the limits of a particularmodel and data set, the significanceof the
results providedfor any factor or factor groupby the estimationprocedurecan
be measured by eliminatingthe p(i)'s associated with such factors or groups,
and verifying whether this has a significant adverse effect on the resultant

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 73

maximized likelihood. For example, let us returnto Labov's study of copula


deletion: we might suspect that, in the factor group referringto the nature of
the predicate, the effect of a gerund is really not significantlydifferentfrom
that of gonna-or we might suspect that this whole groupis irrelevant,i.e. that
all its component factors really have the same effect, and that any apparent
differences are statistically invalid. To test for statistical significance along
these lines, we need only find the difference of the logs of the likelihoods for
the two analyses, since twice this difference should approximatea chi-square
distributionif no significant effect results from the additionalfactors of the
more complex model. The numberof degrees of freedom involved in the dis-
tribution is the number of factors eliminated from groups-or, if an entire
group is eliminated, one less than the number of factors in the group. (For
more details, see Rousseau & Sankoff 1978and the sources cited there.)
Only highly sophisticatedstatisticaltechniquesinvolving rigorousmeasures
of the closeness of fit of the models to a wide range of empiricaldata could
determinewhich (if any) of them is to be preferredin general. Fortunately,this
problem does not exist for the phenomenon studied here-since it turns out
that, with certainexplainableexceptions, the three models give parallelresults.
Furthermore,since the models are intended only as approximationsto joint
effect, and are neither theoreticallynor empiricallyjustified in detail, we will
be concerned more with relative strength of constraints than with exact nu-
merical values.
4.3. CODIFICATIONANDCOMPUTATION. The tape-recordedinterviews were
transcribedby the staff of 'CompetenciasBasicas do Portugues',who located
and markedverb tokens with plural subjects. Each such token was classified
in accord with the structuralconstraintspresent in its environment(i.e. mor-
phologicalcategory of the verb and position of the subject).For computational
purposes, a single alpha-numericsymbol was assigned to each variable con-
straint(N for the morphologicalclass offala/falam, A for immediatelypreposed
subjects), and a tally was made of the numberof times agreementapplied in
each environment.This resultedin a series of observationsof the type 'Speaker
VAN appliedagreement37 times out of 54 in the environmentsymbolizedNA.'
After a third alpha-numericsymbol was added to identify the speaker, these
data were punched on cards in a standardformat.
The cards were then submittedto SWAMINC4(Naro 1978),an editing pro-
gram that checks for errors of numerous sorts, removes data containingcat-
egorical (or 'knock-out') features, and provides a frequency count in accord
with the first approach outlined above. Results based on the additive model
were obtainedby using the ANOVA subprogramof the SPSS (version7) pack-
age (Nie et al. 1975:411-29);correspondingresults according to the logistic
multiplicativemodel were provided by a modifiedversion of Sankoffs VAR-
BRUL2 programthat also calculates the log likelihood.
5. LINGUISTIC DIMENSIONS: RESULTS FOR MORPHOLOGICAL
CATEGORY.AS
stated in ?4, the basic observationthatguidedthe orderingof the morphological
categories was that increasingphonic salience of the singular/pluralopposition

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
74 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

seemed to correlate with increasingchances of agreement.In order to reflect


this centralidea, a hierarchyof categorieswas set up so as to reflect increasing
oppositionalsaliency. This morphologicalhierarchywas divided into two lev-
els. The first level contains those pairs in which the phonetic segments that
realize the opposition are UNSTRESSEDin both members; the second level con-
tains those pairs in which these segments are STRESSED in at least one member
of the opposition. Given this fundamentaldistinction in terms of stress, it is
clear that all oppositionson the first level are less salientthanany on the second
level.
The first level of the hierarchywas further divided into three classes, ar-
rangedin order of increasingmaterialdifferentiationof the singular/pluralop-
position. In Class la, which consists of the presentof the 2nd-3rdconjugations,
this oppositionis normallyrealizedin our sampleas [i]/[i];i.e., the pluralshows
only nasalization, with no change in quality. Class lb, containingthe present
of the 1st conjugation,usually shows a more salient oppositionof the type [a]/
[fi], where the nasalizationin the plural is variable, but the differentiationin
qualityof the vowels serves to distinguishthe forms. However, a few speakers
have [a]/[a]for this opposition, and others show considerablevariationamong
[a u] when the plural is utilized. In Class Ic, which contains the present of
2nd-3rdconjugationverbs ending in -r or -z, the entire unstressed segment is
missing in the singular,resultingin a zero/[i] or zero/[i] opposition.
On the first level of the hierarchy,the oppositionof Class la consists of the
presence or absence of the suprasegmentalfeature of nasalization;Class lb
shows an opposition markedby a greaternumberof distinctive segmentalfea-
tures; and Class Ic shows addition of an entire segment. Thus oppositional
saliency increases throughthe three classes.
At the second level, we enter the range of the hierarchyin which the desin-
ences are stressed. On this level, too, we distinguishseveral classes, and ar-
rangethem in orderof increasingmaterialdifferentiationof the opposition.The
least salientpluralon this level shows nasalizationof the stressed vowel, which
does not change, and additionof an offglide. This occurs in Class 2a, containing
the present of several common monosyllables (e.g. ddldao 'gives/give'),6 as
well as the future of all verbs (e.g. falardlfalarao 'will speak [sg./pl.]') Class
2b, the next most salient, also shows no change in the stressed vowel, but an
entire syllable, consisting of two segments, is added in the plural, contrasting
with an offglide in the singular.All pairs in Class 2b are in the preteritperfect,
but there are several distinct varieties. In the 2nd conjugation,the constant
vowel is [e], and the opposition is [-ew/-eru](comeu/comeram'ate [sg./pl.]'),
while the 3rd conjugationhas [i] and [-iw/-iru](partiulpartiram'left [sg./pl.]')
One unique case with constant vowel [o] and singularoffglide [y] (foilforam
'was/were' or 'went [sg./pl.]') is also includedin Class 2b.7Class 2c, in addition
to the syllabic contrast of 2b, shows a change in the stressed vowel from [o]
to [a] in the opposition [-6/-ari] (falou/falaram'spoke [sg./pl.]') On the second
6
Although slightly different, the opposition vailvao 'goes/go' is also included here.
7
Thus Class 2b combines the classes labeled J and Z in Lemle & Naro.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 75

level of the hierarchy,then, the oppositionof Class 2a consists of the presence


or absence of nasalization (accompaniedby an offglide); Class 2b shows an
opposition marked by a greater number of segments, although the stressed
segment remainsconstant; and in Class 2c even the stressed segment changes.
Thus, oppositionalsaliency increases throughthe three classes of this level as
in the first level.
As stated, Class 2c shows no sharedsegmentbetween the stresseddesinences
that realize the singular/pluralopposition, althoughthe verb root is constant.
The present of the verb ser 'to be' (el/so 'is/are')is unique:it is a monosyllable
in which no clear separationof root from desinence is possible-and in which,
as in the stressed desinences of Class 2c, singularand pluralhave no shared
segment. To test the behavior of this unique case against that of Class 2c, it
was assigned to a special Class 2d.
Finally, stress-shiftingand radical-changingpreterits(types 4 and 5 of Table
1) are includedin Class 2e.8 In these verbs, the changingposition of the accent
introduces a high level of saliency; but the fact that there is no stressed des-
inence in the singularreduces the over-all effect.
A summaryof the hierarchyof oppositionalsaliency, with the morphological
classes arrangedin accord with the levels of saliency discussed above, is as
follows:9
(6) LEVEL1 (unstressed):
Class a. [-i/i]
b. [-a/iu]
c. [-0/-]
LEVEL2 (stressed):
Class a. [-a/-aw]
b. [-ew/-eru,-fw/-fru,O-6y/-6ru]
c. [-6/-aru]
d. unique case: elsao
e. ['0/-eru,-i/-eru]

8
Since these forms occur quite infrequentlyin real speech, it was not possibleto maintainthem
as separateclasses. The verb 'to come', which in the prescriptivestandardhas the infinitivevir
and the preteritperfect veiolvieram'came (sg./pl.), is currentlyundergoinga complex series of
changes involvingthe introductionof nasalizationthroughoutthe paradigm,possibly in order to
avoid confusionwith ver 'to see' (see Camara[1965]1975:89),which has similaror identicalforms
in its paradigm(e.g. vimos 'we saw' or 'we come'; veem 'they see' and vem 'they come', where
the differenceis merelyorthographic).This process has resultedin the creationof a new infinitive
vim [vi] for 'to come' (all other infinitivesend in -r, but the most frequentspoken realizationof
this segmentis zero), which happensto coincide with lsg. preteritperfectvim 'I came'. In the 3rd
person preteritperfect, the new form is vinheram,replacingstandardvieram'they came'; but 3
sg. veio 'he came' is apparentlyremainingstable. Because of this situationof changein the plural,
youngerspeakers of both groups overwhelminglypreferto use the stable veio, tendingto avoid
both vieramand vinheram.This producesa decrease in the frequencyof agreementfor 'to come'
in the preteritperfect, independentlyof the generalrule under study here; hence data from this
verb have been removed from Class 2e for the younger speakers, by means of data-exclusion
commandsof SWAMINC4and SPSS.
9 The final vowels of Classes Ib, Ic, 2b, 2c, and 2e may be oral.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
76 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

The results obtained for the morphologicalvariable according to all three


models are shown in Table 3. These results, as well as all others given in this
section, were calculatedfor the groupof 17speakersas a whole; i.e., individuals
were aggregated into a group (for justification, see ?7 below). The column
labeled FREQUENCYreports the raw rate of agreementin accord with the first
approachoutlined above. For the logistic model, the factorialprobabilitiesof
agreementcalculated by means of VARBRUL2 are listed under the heading
PROB;the correspondinginput parameteris 0.41. The results accordingto the
additivemodel are cast in the form of deviationsfrom the grandmean, or over-
all frequency of agreement, resultingfrom the presence of the factors. These
deviations, calculatedby the MCA statisticsoptionof the ANOVA subprogram
of SPSS, are listed underthe label ADJDEV(adjusteddeviation);the grandmean
is 3002/6310, or 47.6%. In all instances, the ANOVA analysis itself showed
that all main effects, as well as most interactions,were statisticallysignificant
at the .001 level; but interactionwas included in the error term, as outlined
above.

CLASS EXAMPLE FREQUENCY PROB ADJ DEV


la comelcomem 110/755 = 14.6% 0.07 -33.7%
b falalfalam 763/2540= 30.0% 0.19 -18.2%
c fazlfazem 99/273 = 36.3% 0.26 -11.3%
2a ddldao 604/927 = 65.2% 0.58 +17.9%
b comeulcomeram 266/365 = 72.9% 0.69 +25.5%
c faloulfalaram 524/672 = 78.0% 0.78 +33.3%
d elsao 539/662 = 81.4% 0.79 +33.6%
e disse/disseram 97/116 = 83.6% 0.80 +35.4%
(Beta = 0.74.)
TABLE3.

Inspection of the figures reported in this table shows that increasing rate
operationof the agreementrule indeed correlateswith increasingoppositional
saliency. In all three models, the largest jump in rate of agreement occurs
between levels 1 (stressed) and 2 (unstressed) of the oppositional hierarchy.
Thisjump amountsto about 30%in termsof eitherraw or adjustedfrequencies,
and about 0.3 in terms of probabilities.WithinLevel 1, the least agreementis
shown by Class la (nasalization,i.e. minordifferentiation),followed by Class
lb (change in vowel quality, i.e. greaterdifferentiation)and Class lc (addition
of a segment, i.e. complete differentiation).Level 2, similarly,shows its lowest
rate of agreementin Class 2a (nasalization,i.e. minordifferentiation),followed
by Class 2b (addition of a syllable, i.e. greater differentiation)and Class 2c
(desinences with no shared segment, i.e. complete differentiation).
Recall that, in additionto Class 2c, which shows total differentiationof the
stressed desinences, we set up for purposes of verificationa special Class 2d,
consisting of the single opposition elsao. As regardsoppositionalsaliency, this
item also shows total differentiationof forms, and these forms are always
stressed. The opposition of Class 2c does, however, possess the constant un-
stressed verbal root in both items (fal- in our example), while a constant root

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 77

of this sort is absent in Class 2d. Table 3 indicates that this circumstancehas
little or no effect, since all three models show nearlyidenticalresults for these
two classes. Furthermore,the log likelihood test in the logistic model shows
that the small difference found in that model is not significanteven at the .5
level. It can thus be concluded that Class 2d is really not differentin behavior
from Class 2c, and that these two classes should be amalgamated.
Class 2e contains the stress-shiftingpreterits-which, parallelto Class 2c,
show no shared segment in the desinences that mark the singular/pluralop-
position. Here, however, the singularshows root stress; and the singularde-
sinence is either zero, as infez 'did (sg.)', or unstressed [i], as in disse 'said
(sg.)' The pluraldesinence is stressed -eram (disseram'said [pl.]',fizeram 'did
[pl.]'), identical to Class 2b. Thus the opposition in this class is [i, -0/-eru];
and in some verbs the root vowel also changes. Despite these nuances, the
fundamentalsimilarityto Class 2c remains clear: there is complete differen-
tiation of the desinences, and they are stressed in the plural.
Table 3 shows slightlymore agreementfor Class 2e than Class 2c in all three
models; but in the logistic model, the log likelihoodtest reveals that the minute
distinction found there (0.78 vs. 0.80) is not valid even at the .7 level. These
negative results are, of course, partiallycaused by the fact that Class 2e has
the fewest tokens of all the classes. Until more data are forthcomingon this
class than those found in the approximately115 hours of speech considered
here, it must be concluded temporarilythat Class 2e is very similarin behavior
to Class 2c, and that the nuances distinguishingthem are not important.From
this point on, Class 2c will be understoodto includeboth 2d and 2e. The results
of this amalgamationare shown in Table 4.
CLASS EXAMPLE FREQUENCY PROB ADJ DEV
la comelcomem 110/755 = 14.6% 0.11 -33.7%
b falalfalam 763/2540 = 30.0% 0.26 -18.2%
c fazlfazem 99/273 = 36.3% 0.35 -11.3%
2a ddldao 604/927 = 65.2% 0.68 + 17.9%
b comeulcomeram 266/365 = 72.9% 0.78 +25.5%
c faloulfalaram 1160/1450 = 80.0% 0.85 +33.6%
(Beta = 0.74.)
TABLE4.

Comparisonof Tables 3-4, which refer to the same data under slightly dif-
fering analyses, reveals several points of interest. The raw frequenciesfor all
classes up to 2c do not, of course, change at all; but the new result for 2c
represents an average of the former 2c, 2d, and 2e weighted by number of
tokens. In the logistic model, however, all the new probabilitiesare numerically
differentfrom those of Table 3, althoughthe relationshipsamong various cat-
egories have not changed. The reason for this is the mathematicalconstraint
imposed upon the average of the terms in Formula5 for each group. Since the
numberof factors in the group has changed from eight to six, the averaging
gives different results. We need not, however, concern ourselves with this
matter,since all thatis of interestis the constantrelationshipamongcategories.
In the adjusted deviations, the new results bear the same sort of relationship

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
78 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

to the old as in the case of the raw frequencies. It is important to note, however,
that the partial beta, which reflects the proportion of total variation accounted
for by the factor group, remains constant at 0.74. Thus, confirming the log
likelihood test in the logistic model, the more complex analysis gives no im-
provement in the results of the additive model.
The final morphological categorization of Table 4 was subjected to two fur-
ther types of test. First, using the log likelihood technique, it was confirmed
that the results for these six categories are indeed statistically different at the
.005 level; thus no additional amalgamations would be valid. Then several at-
tempts to improve the analysis were made by splitting some of the six categories
into subcategories. For example, in Class 2b the opposition foilforam, the ir-
regular preterit of both ser 'to be' and ir 'to go', was separated from regular
2nd-3rd conjugation preterits. Forfoilforam a frequency of 111/156, or 71.2%,
was found; the regular preterits' rate of agreement was 155/209, or 74.2%. In
the logistic model, this gave results of 0.72 vs. 0.76, a distinction which the log
likelihood test shows to be insignificant even at the .25 level. Based on this
sort of statistical testing, I conclude that the six-leveled hierarchy realistically
portrays the true situation.
5.1. A FEATURE ANALYSIS OF SALIENCY.In the informal discussion of the
concept of oppositional saliency, it was argued that two main forces were si-
multaneously operating. The first of these was stress; and it was assumed, quite
naturally, that stressed oppositions are more salient than unstressed opposi-
tions. This distinction was used to set up two levels, on each of which the
secondary criterion of material differentiation was used. Thus one might argue
that the hierarchy of morphological categories is a derivative notion, and that
saliency would be more insightfully analysed in terms of its component features.
To carry out such an analysis, two new factor groups were postulated-
stress and differentiation:
(7) STRESS:
Class a. unstressed (Morphological Classes la-c)
b. stressed (Morphological Classes 2a-c)
MATERIAL DIFFERENTIATION:
Class a. minor differentiation (Morphological Classes la, 2a)
b. greater differentiation (Morphological Classes lb, 2b)
c. complete differentiation (Morphological Classes Ic, 2c)
Through appropriate recoding commands of SWAMINC4 and SPSS, the factor
group representing the morphological hierarchy on the original data cards, was
replaced by the appropriate values of these two factor groups. For example,
the symbol representing Morphological Class lb was rewritten as symbols rep-
resenting Stress Class a and Material Differentiation Class b. The results are
shown in Table 5.
For the Stress feature group, all three models agree and show extremely
sharp polarization, indicating that this is a very important determinant of agree-
ment. In the Material Differentiation constraint group, however, we see for the
first time a case where the raw frequencies do not agree with the probabilities

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 79

TYPE CLASS FREQUENCY PROB ADJ DEV


STRESS
a unstressed 972/3568 = 27.2% 0.22 -20.6%
b stressed 2030/2742 = 74.0% 0.78 +26.8%
beta = 0.70
MATERIAL DIFFERENTIATION
a minor 714/1682 = 42.4% 0.35 -10.8%
b greater 1029/2905 = 35.4% 0.54 +2.0%
c complete 1259/1723 = 73.1% 0.62 +7.2%
beta = 0.20
TABLE 5.

of the logistic model, or with the adjusted deviations of the additive model.
Both the latter are in accord with our expectations:increasingmaterialdiffer-
entiation correlates with increasingrate of agreement.Furthermore,although
some of the numericaldifferences involved are relatively small, the two ap-
propriatelog likelihood tests show that the results are statisticallydistinct at
the .005 level.
The reasons for the deceptive resultsin the rawfrequencycolumnare varied.
In the first place, DifferentiationClass a is a more or less evenly-weighted
average of MorphologicalClasses la and 2a; but DifferentiationClass b is
heavily weighted toward MorphologicalClass lb, and away from 2b, because
the former happens to have far more tokens. Just the opposite occurs with
DifferentiationClass c, which is heavily weighted in favor of Morphological
Class 2c, and away from Ic. Thus DifferentiationClass b is weighted in favor
of the lower rate of agreementclass, while DifferentiationClass c is weighted
toward the higher rate of agreementclass. If all three classes had been evenly
weighted, the results for DifferentiationClasses a, b, and c would have been
39.9%, 51.5%, and 58.2%, respectively. It is just this sort of infelicitous dis-
tributionof data that was referredto in the first paragraphof ?4.2, above, as
invalidatingraw frequencies as a generally valid measure of effect. What is
importantis the joint effect of the stress and differentiationfeatures:for each
class of stress, each class of differentiationhas a paralleleffect, namely that
of increasing(or decreasing)rate of agreement.The additiveandlogistic models
are capable of reflectingthis fact, despite the skewed distributionof data.
5.2. RESULTSFORPOSITIONAL
CATEGORY.
The ordering of the positional cat-
egory was guided by the same central notion of saliency used in the morpho-
logical category. Here, however, the relevant type of saliency is not opposi-
tional, but depends instead upon the relative position of the subject-which
is the element that determinesthe appropriateform of the verb, and the verb
itself.
The positional relationshipis most salient when the determiningsubject im-
mediately precedes the determinedverb. This situation, in which the deter-
miner/determinedrelationshipis totally transparent,is labeled as Positional
Class la, defining 'immediately'as allowingup to five syllables of intervening
material.This cut-off point, while essentially arbitrary,has the result of allow-
ing a series of two or three short adverbs (such as jd 'already', ndo 'not', or

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
80 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

nunca 'never'), while excluding most relative clauses and adverbialphrases.


Less salient is Positional Class lb, in which more than five syllables separate
subjectfrom verb within the same surface sentence. The situationin which the
subject/verbrelationshipis least salient is Positional Class Ic-where the de-
terminingelement, the subject, follows the determinedelement, the verb.
The so-called 'hidden' or deleted subject presents a special problem. It is
clearly distinct from any of the classes discussed above, because the subject
is not even present in the surface sentence-the unit usually taken as basic in
syntactic integration.?0Thus the verbal desinence is not redundantwithin the
sentence; and if the verb and its extra-sententialsubjectare to have any struc-
turallink at all, it can only be accomplishedthroughthe agency of agreement.
Otherwise, there could be no grammaticalalgorithmto establish the cross-
sententialconnection from verb to subject, since there wouldbe no markwhat-
ever of pluralityin the sentence containingthe verb with deleted subject. On
this functional basis, 'hidden' subjects are set up as a separate case outside
the positionalhierarchy,which is valid only withinthe limits of a single surface
sentence." The positional classes are summarizedbelow, with the positional
hierarchylisted in descending order:
(8) CASE1 (within a single surface sentence):
Class a. immediately preposed subject (eles falam)
b. distant preposed subject (eles ... falam)
c. postposed subject (falam eles)
CASE2 (cross-sentential):
Class a. deleted subject (... falam)
Results for the positional variableare shown in Table 6.

CLASS EXAMPLE FREQUENCY PROB ADJ DEV


la eles falam 1802/3680 = 49.0% 0.71 +3.3%
b eles ... falam 84/254 = 33.1% 0.41 - 16.1%
c falam eles 59/252 = 23.4% 0.24 -28.2%
2a ... falam 1057/2124 = 49.8% 0.65 -0.4%
(Beta = 0.21.)
TABLE6.

Within sentential limits (i.e. Case 1), all three models show parallelresults
that accord perfectly with the principle of saliency: Class la, consisting of
immediatelypreposedsubjects, shows the highestrateof agreement-followed
by Class lb, where the subject is still realized and preposed, but separated
from the verb by some distance. The lowest rate of agreementof the hierarchy
is induced by the postverbal subject of Class Ic.
Case 2 uniformly shows results of approximatelythe same magnitudeas
Class la. I see no way in which this fact can be explained in terms of the

'0 For an explicit model along these lines, see Kimball 1973.
" It would, in fact, be
tempting simply to exclude sentences without realized surface subjects
from the data base, if this did not involve the loss of over 40% of the data. Nonetheless, the results
obtained in this way do not differ markedly from those reported in the text.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 81

principleof saliency; rather, it seems necessary, as noted above, to postulate


that other principles of a functional nature interact with saliency, inhibiting
what would amount to complete loss of any syntactic realizationof plurality
in a sentence with deleted subject and unmarkedverb.
All runs reported above contained, in addition to the
6. SOCIALDIMENSIONS.
structuralconstraints (morphologicaland positional factor groups, or the two
groups constituting the feature analysis of the morphologicalclasses as well
as the positional group), a factor group consisting of identificationsof each
individualspeaker. This was done in order to include in the calculations the
social variables that had been left out of explicit consideration. In fact, the
logistic model probabilitiesfor these factors range from 0.20 to 0.89, and the
adjusted deviations from -24.9% to +30.9% (partialbeta = 0.43), showing
that social variablesmust be quite importantfor a complete account of verbal
agreement.
Table 7 reports the over-all frequency of applicationof the agreementrule
for each speaker. The TOTALcolumn gives the number of verb tokens with
column gives the subtotal of those verb
plural subjects, while the APPLICATIONS
tokens that actually showed a plural inflection. The table is in order of de-
creasing relative frequency, given as a percentage.
NAME APPLICATIONS TOTAL FREQUENCY
SON 421 523 80.5%
VAN 180 258 69.8%
CAN* 96 141 68.1%
SID 312 505 61.8%
UBI* 315 622 50.6%
MAD* 132 263 50.2%
ELV* 344 701 49.1%
LID 72 184 39.1%
HEN 127 328 38.7%
ILD* 191 494 38.7%
JOM 139 368 37.8%
GCL* 77 209 36.8%
FAU 299 818 36.5%
IZE 145 417 34.8%
MAR 114 346 33.0%
PAU 28 91 30.8%
NAT 10 42 23.8%
Total 3002 6310 47.6%
TABLE7. (Asterisk marks older speakers.)

On the average, this sample of illiteratespeakersapplies the agreementrule


in about half the appropriatecases; but there is an extremely wide range of
variation,from NAT's 24%,on one extreme, to SON's 81%,on the other. As
it happens, these two speakersare approximatelythe same age, and both grew
up in the city of Rio. They are of opposite sexes; but MAR, near the bottom
of the list, shows the same age group, area of origin, and sex as the speaker
with the highest incidence of agreement. Similarly, SID has the same social
characteristicsas NAT, but is very near the top of the list.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
82 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

The older speakers show considerablymore uniformityof behaviorthan the


younger, being concentrated in the upper range with an over-all variation of
only about 25%(approximately40%-65%).Nonetheless, youngerspeakersare
found above, below, and among the older ones.
The general impressionof indifferenceof effect of the social variableslisted
in Table 2 is reinforcedby inspection of the generaldistributionof frequencies
with respect to these categories, shown in Table 8.
WEIGHTING TOKENS SPEAKERS

city 1605/3216 = 49.9% 45.0%


environs 1397/3094 = 45.2% 47.6%
younger 1847/3880 = 47.6% 44.2%
older 1155/2430 = 47.5% 48.9%
male 1176/2490 = 47.2% 43.1%
female 1826/3820 = 47.8% 49.1%
TABLE8.

The second column of Table 8 was calculatedby simply aggregatingall data


across all speakers. This procedure has the effect of weighting more heavily
the behaviorof people who happento have producedmore data. For example,
in the aggregated sample, the behavior of FAU, with 818 data tokens, is
weighted about 20 times more than that of NAT, who has only 42 data tokens.
In part, this sort of weightingis justified because the results for speakers with
more data are probably more reliable; but good statistical practice would re-
quire a more sophisticated weighting algorithm.The third column of Table 8
goes to the opposite extreme, weightingeach speaker equally, without taking
into account quantityof data (or any other property)in any way.
To judge by the results of Table 8, the social factors reportedthere would
not be thoughtvery important.The largestdifferencesare on the orderof only
about5%('city' aheadof 'environs'when weightedby numberof data; 'female'
ahead of 'male' when weighted by speaker).Furthermore,whateverweighting
procedure one might choose would have to provide results intermediarybe-
tween those of the second and thirdcolumns, reducingthese small differences
even further.
However, serious difficulties of two sorts invalidatethe results of Table 8.
In the first place, the data of 'CompetenciasBasicas do Portugues'were not
controlledfor a balanceddistributionof social variables:the principalsponsor
was, at the time, interestedonly in grammaticalinformationthat could be used
in the planningof teaching materialsand techniques. For example, of the nine
speakers who spent their formativeyears in the city, six are males; but of the
eight speakers broughtup in the outlying areas, only three are males. In nu-
mericalterms, of the 3216 tokens producedby the city group, 1930(or 61.0%)
were from males; but of the 3094 tokens spoken by the group from the sur-
roundingareas, only 560 (or 18.1%)were producedby males. This is exactly
the sort of skewed distributionof data that can lead to deceptive raw frequen-
cies, as noted above in ?5.1. Furthermore,there are even some combinations
of social categories, such as olderfemalesfromthe city, thatare not represented

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 83

at all in the data, creating empty cells. However, since both the logistic and
the additivemodels are capableof dealingwith skewed distributionsand empty
cells, these statistical difficulties can be overcome.
A second, and more fundamental,problemis that the proposed social var-
iables includeno measureof the speaker'spossible relationshipsto other socio-
economic levels in the surroundingcommunity. As already mentioned, the
speakers under study are all illiterateand from the same (low) socio-economic
level. They are, further,all orientedtowardimprovingtheir lot, a fact attested
by their presence in adult literacy courses. Without exception, they hope to
'better' their lives throughthe course. One even dreamsof becominga doctor,
while anotherhopes eventuallyto own a stable of race horses and an apartment
in the most expensive area of the city. In general, older speakers have more
modest ambitions.Thus ELV (age 45) hopes to 'have a familyand get married'
(she is a maid, living by herself in a room in her employer's apartment);UBI
(age 42) wants only to 'stop being a janitor' (his currentprofession);ILD (age
43) hopes to become a seamstress or a manicurist,etc. By way of contrast,
SID (age 16), HEN (age 17), and MAR (age 26) want to travel outside Brazil
as a preliminaryto their future careers, a privilege usually reserved for the
childrenof the elite. VAN (age 15) is the would-be doctor, while IZE (age 23)
hopes to get a bankingjob. Althougha study of such aspirationswould indeed
be interesting, the high correlation with age rules out their use as a social
indicator.
What we really wish to measure is the degree to which a given speaker has
managed to penetrate the dominant culture of the middle and upper-middle
socio-economic levels.12 I have been careful to limit the desired measure to
middle-level CULTURE because, in point of fact, only one of the 17 speakers
consideredhere has anythinglike middle-classstatus in real life.'3 Despite this
lack of actual status, however, some speakers show that they are able to
participatein middle-class concerns and anxieties, and seem able to relate to
their middle-class interviewers perfectly well on this level. These speakers
have, as it were, a clear view from the outside, althoughthey are not actually
able to enter the higher socio-economic levels; the social view of the others
is strictly limited to their own level.
Since the interviews were conductedwith no specific concern for measuring
social factors, we must be satisfiedwith an indirectindicator.For this purpose,
I propose to use the speaker's reaction to television. In this connection, it is
importantto bear in mind that the target of Braziliantelevision programming
seems very narrowlyconfined to the urbanmiddle (or upper-middle)class-
probablybecause this group, althoughnumericallysmall in comparisonto the
total population of the country, controls a very large proportionof the total
purchasingpower. Prime evening time is taken up mainly by 'novelas', i.e.
12
My occasional use of the term 'class' is strictly non-technical. In particular, I do NOT have
in mind the Marxist concept of class or any of the associated theoretical apparatus.
13
This is CAN, who lives in his own apartment in one of the less expensive middle-class areas
of the city. All the other speakers either live with their employers or in areas of
markedly lower
socio-economic status.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
84 LANGUAGE, VOLUME57, NUMBER 1 (1981)

short-run soap operas that are heavily based on middle-class morality and
anxieties. These soap operas are followed by their viewers with great dedi-
cation, and are successfully used by businessmenas a basis for a whole series
of secondary products, rangingfrom record albums and magazines through
jewelry and other fashion items. Most of the comedy shows, too, presuppose
a culturaland political backgroundnot attainedby the people in our sample.
Nonetheless, a relativelysmallerproportionof broadcasttime (mainlySaturday
and Sunday afternoons)is given over to variety shows featuringpopularsing-
ers, sambas, amateurcompetitions of various sorts, and so on.
During investigations preliminaryto the 'Competencias Basicas do Por-
tugues' research project, it was quickly discovered (contraryto expectation)
that most of the speakers interviewed watched little if any television, even
when they had free access to a set. Those who did watch television over-
whelmingly preferred the weekend variety shows or broadcasts of soccer
matches. Whenrequested,most of the illiteratepeople interviewedwere unable
to recount the plots (or even, for that matter, the titles) of the most popular
novelas. The reason for these people's lack of interest in novelas seems to be
their inability to follow the story lines: they simply cannot understandthe
motivationbehind the characters'actions and reactions.
In the sample of 17 speakers under study here, only five are followers of
novelas. Of these, four are from the youngergroup-two females (SON, VAN)
and two males (SID, JOM)-while one (CAN, a male) is from the older group.
The remaining twelve speakers either watch no television at all, or watch
mainly the variety shows. Table 9 shows the usage of the verbal agreement
rule by these two groups. In accord with the ideas developed above, novela-
watching is taken as indicative of a wider culturalview-or what I shall call,
at Gillian Sankoffs suggestion (p.c.), a 'vicarious' orientation,in contrast to
the 'experiential'orientationof the remainingspeakers.
WEIGHTING TOKENS SPEAKERS
vicarious 1148/1795= 64.0% 64.9%
experiential 1854/4515= 41.1% 42.1%
TABLE 9.

Althoughbased only on potentially-deceptiveraw frequencies, Table 9 con-


firms the idea that culturalorientationis an importantdeterminantof rate of
use of the agreement rule. However, given the highly skewed and partially
defective distributionof data withinthe varioussocial and linguisticcategories,
use of a statistical model of joint effect becomes imperative.
In the logistic model, D. Sankoffs VARBRUL2 programimposes no limi-
tations on the quantity of data or factor groups in a particularanalysis, aside
from the storage capacity of the computer utilized. However, SPSS (version
7) limits the capacityof its MCAanalysisto only five factorgroupsper analysis.
This limitation constitutes a serious problem, since we have four candidates
for social groups (orientation,sex, age, origin)and three for structuralgroups
(stress, differentiation,position), makinga total of seven. For this reason, the
data, analysed according to all seven factor groups, were first run in VAR-

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 85

BRUL2 (see Table 10 for results), and appropriatelog likelihood tests were
performedin orderto evaluate the statisticalsignificanceof each group. These
tests show that each of the postulated factor groups, with the exception of
origin(city of Rio or surroundingareas), is significantat the .005 level or better,
but that this latter group's significanceis worse than .75. The tests rankedthe
factor groups in the following order: stress, cultural orientation, position of
subject, materialdifferentiation,age, sex. The first five of these groups were
then run in SPSS's MCA program.In order to test the remaininggroups (sex
and origin), the stress and materialdifferentiationfactor groups were replaced
by the morphologicalcategories of Table 4, and two separateruns were made,
each with the four surviving factor groups (morphologicalcategory, cultural
orientation,position of subject, age) as well as one of the two remainingvar-
iables. For this reason, the additivemodel results in Table 10 are, in fact, from
three separate runs.
TYPE CLASS FREQUENCY PROB ADJ DEV
STRESS
a unstressed 972/3568 = 27.2% 0.24 -20.5%
b stressed 2030/2742 = 74.0% 0.77 +26.7%
beta = 0.69
CULTURAL ORIENTATION
a vicarious 1148/1795 = 64.0% 0.69 + 19.7%
b experiential 1854/4515 = 41.1% 0.31 -7.8%
beta = 0.37
POSITION OF SUBJECT
la preposed 1802/3680 = 49.0% 0.71 +3.6%
b distant 84/254 = 33.1% 0.41 -17.6%
c postposed 59/252 = 23.4% 0.24 -29.9%
2a deleted 1057/2124 = 49.8% 0.65 +0.5%
beta = 0.22
MATERIAL DIFFERENTIATION
a minor 714/1682 = 42.4% 0.34 -11.0%
b greater 1029/2905 = 35.4% 0.54 +2.3%
c complete 1259/1723 = 73.1% 0.61 +6.9%
beta = 0.20
AGE
a older 1155/2430 = 47.5% 0.58 +6.9%
b younger 1847/3880 = 47.6% 0.42 -4.3%
beta = 0.16
SEX
a female 1826/3820 = 47.8% 0.54 +2.1%
b male 1176/2490 = 47.2% 0.46 -3.3%
beta = 0.08
ORIGIN
a city 1605/3216 = 49.9% 0.50 +1.6%
b environs 1397/3094 = 45.2% 0.50 -2.9%
beta = 0.06
(r = 0.82.)
TABLE 10.

The first data column of Table 10 repeats informationon raw frequencies


that has already been presented for individualfactor groups. The second (lo-

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
86 LANGUAGE, VOLUME57, NUMBER 1 (1981)

gistic model) and third (additive model) columns, however, representfor the
first time a complete overview of the joint effect of all variable constraints,
both structuraland social. The results for the former differ very little from
those already presented, since the social factors had been included implicitly
in the earliercalculationsby means of variablesrepresentingindividualspeak-
ers. The various factor groups are listed in Table 10 in order of statistical
significancewithin the logistic model as determinedby the log likelihoodtest.
This correspondsexactly with the orderingaccordingto partialbetas.
The statistic r is, roughly speaking, to the whole what the betas are to the
parts-it indicates (exactly), when squared, the total proportionof variation
in frequency of agreementattributableto the additiveeffects of all the factors.
The level of better than 0.8 achieved here is generally considered extremely
successful in empiricalstatistical investigations.
The results for the social variables are quite revealing. By far the most
importantsocial conditioningis attributableto the speaker's degree of pene-
tration into the culture of the surroundinghigher socio-economic levels, as
reflected in his or her habits of television viewing. It is true, of course, that
television provides a massive dosage of the standardlanguagewith its cate-
gorical agreementrule; but the same is true of radio, a mediumdirected more
towardthe lower classes and enjoyed by a good numberof the speakersin our
sample-even if its horoscopes, advice to the lovelorn, outragedcommentary
on local issues etc. are interspersedwith a good deal of popular music. The
devotion to novelas on the part of these people would seem to indicate that
theircomprehensionof middle-classlife style has reachedthe point where they
can follow the plots and develop an interest in them. It is the fact that these
speakersparticipatevicariouslyin a socio-culturalcontext foreignto theirown
milieu that is reflected by the culturalorientationvariable.'4
The age variable, which seems indifferentin terms of raw frequencies (but
see the last column of Table 8, where the data from each speakerare weighted
equally), shows a significantly higher rate of agreementfor the older group
when the simultaneouseffects of the other variablesare taken into account on
either the logistic or the additive model. This age-gradingin apparenttime is
the 'synchronic'reflex of the 'diachronic'trajectoryof the agreementrulefrom
the grammar.As we have come to expect in this sort of linguistic change,
women are closer to the prestige form than men-although, in this particular
case, the difference is very small and accounts for only a minute proportion
of the over-all variation,a fact reflected in an extremely low partialbeta. The
area of origin of the speakers shows no effect at all in the logistic model, and
a negligibly small effect in the additive model.
14For the
record, I know of one person for whom this variable does not provide the desired
results. This speaker, diagnosed as mentally retarded by the state authorities, has an over-all rate
of application of the agreement rule of only about 30%; yet she spends virtually the whole day
planted in front of a television set and thus watches the full range of programming, from the variety
shows through the novelas and even news broadcasts. Unfortunately, a thorough investigation of
the assumed general correlation between television viewing habits and cultural orientation would
require sophisticated sociological and anthropological techniques far beyond the goals of this paper.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGF 87

In the social variables, as with the materialdifferentiationvariable, there is


a certain lack of correspondence between raw frequencies and the results
accordingto the models of joint effect. As already noted in the discussion of
Table 8, this results in partfrom the skewed distributionof data among speak-
ers. In fact, both models provide results more similar to the last column of
Table 8 than to the 'Tokens' column. There is, furthermore,a significantpe-
culiarityin the behaviorof the culturalorientationand age variables.Consider
the frequencies reportedin Table 11.
OLDER YOUNGER
All 47.5% 47.6%
Vicarious 68.1% 63.6%
Experiential 46.3% 35.7%
TABLE11. Raw percentagesof verbalagreement
accordingto age and culturalorientation.

The first row of this table repeats the over-all frequency for all older and
younger speakers, considered as an aggregatedgroup; the second and third
rows break down this group into vicarious and experiential orientation
subgroups. Within both age groups, the speakers with a vicarious cultural
orientationshow more agreementthan those with an experientialorientation;
within each orientationgroup, older speakers agree more than younger speak-
ers. Unfortunately,this latterfact does not come throughto the over-allfigures
for age groups, because the skewed distributionof data causes the weighted
averages in the first line to turn out nearly equal. Nonetheless, the models of
joint effect are capable of overcoming this difficulty; they present a correct
evaluation of the effects of each variable.
Note that the orientation variable has a much greater effect than the age
variable. This can be seen in several ways. Formally, the partialbeta of the
additive model is more than twice as large for orientationas for age. Further-
more, the net difference associated with the vicarious/experientialdichotomy
is 27.5%, while that of older/youngeris only 11.2%.In the logistic model, the
probabilities for the orientation categories are more distant from 0.50 (the
neutral point) than those of the age variable. Finally, informalinspection of
Table 11 quickly makes one realize that the differencesbetween the rows are
far greaterthan those between the columns.
In sociolinguisticterms, the relativestrengthof these two variableconstraints
reflects a real-worldconflict between opposing trends. On the one hand, we
see in the age variablethe continuationof a slow process of rule death, related
to similar changes in the NP and elsewhere. This variable is, so to speak,
exterior to the speaker, impingingupon him or her without any internalmo-
tivation. The orientationvariable, on the other hand, reflects a resurgencein
the use of this same rule that comes from within the speaker, dependingupon
his or her own culturalorientation.In contrast to the relatively small effect of
being older, possessing a vicarious culturalview implies a large differentialin
incidence of the agreement rule, when comparedwith other speakers having
otherwise similarsocial characteristics.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
88 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

It is temptingto speculate on what the long-termlinguisticoutcome of these


opposing social forces might be. Purely on the basis of informalobservation,
I have the impression that the vicarious orientationis much more prevalent
amongthe young than among the old; and the distributionin the sample under
study here confirmsthis. If the vicariousorientationcontinuesto spreadamong
youngerspeakers, given the relativelysmallage-differential,the situationcould
soon reach a point where that groupmightproducea higherrate of agreement
than the older speakers. Since there are no social barriersbetween the two
orientationsubgroups, at a certain point in time the resurgencein use of the
agreementrule mighteven spreadthroughoutthe youngergroup,independently
of the orientation variable, and produce a reversal in the trend toward the
eliminationof the rule. Future studies will determineif there is any validity to
this speculation.
7. THEINDIVIDUAL
WITHIN
THEGROUP.All results presented up to now have
referred to the behavior of a group of speakers defined in terms of social
parameters.Given current psycholinguistictheories, according to which lin-
guistic activity is controlledby the individualbrain, such collective data would
seem to be meaningfulonly insofar as all membersof the group behave simi-
larly; otherwise, group data would amount to no more than a meaningless
average of disparatebehaviors. If this occurred, the group figures would not
describe correctly the performanceof any individual,and could have no psy-
chological reality. Furthermore,given that the forms used by a speaker are
similaror identical in all relevant aspects to those used by other speakers, the
scale of saliency is invariantacross speakers, and there is thereforeno reason
to expect conflicting orderingsof constraintsin the hierarchizeddimensions,
either among speakers, or between individualsand the group.
To evaluate the behavior of the individualwith respect to the group, I shall
use techniques developed by Guy 1975,modifiedsomewhatfor the case under
study here and incorporatingrecent statistical advances. In any comparison
of statistics for a group of speakers with the correspondingresults for its
component members, it is very importantto bear in mind the fact that the
group always has the advantageof possessing a greaterquantityof data; thus
accuracy is increased, and the chances of noticeable fluctuationsin less fre-
quently occurringcategories, resultingfrom probabilisticallydeterminedout-
comes, are reduced. In this sense, the idea that the groupis more regularthan
the individualis a triviality.
The evaluation of group vs. individualbehaviorwill be carriedout first for
the morphologicalhierarchy, since this is the most complex constraintgroup
and thus most likely to show distinctions. There are six morphologicalcate-
gories in the hierarchy. In principle, it would be possible for an individual
speaker's results for a given category to fall anywhere with respect to any
other category; i.e., Class la might turn out to show a higher incidence of
agreement than Class 2c. In practice, however, we rarely find deviations of
more than one rank with respect to the group pattern. Thus we are interested
in determiningthe individual'sbehaviorin five pairs of neighboringcategories
of the hierarchy. Earlier, in dealing with the group data (?5), the .005 level

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 89

was used as a criterion for accepting numerically distinct results for such
neighboringcategories as statisticallyvalid. However, because of the relative
scarcity of data on individuals, it becomes necessary to relax this criterion
here to the 0.1 level.
For purposes of comparisonwith the grouped results, six individuals,rep-
resentinga balanced distributionof the three most importantsocial variables,
were chosen. It was requiredthat four of these speakers be from the younger
group, and two from the older. Withineach age sample, half were male, and
half female. Furthermore,in the case of the younger speakers, one individual
of each sex showed the vicarious culturalorientation,and one the experiential
orientation.15To amelioratethe problemof scarcity of data on individuals,the
speakers actually selected in accord with the above criteria were those who
had the greatest numberof data. Because of low cell counts, however, some
speakers showed categorical behavior (i.e. 0% or 100%rule application)in
certain cells (e.g., IZE had 0% for MorphologicalClasses la and lc). The
logistic model is unableto deal with such factors because the terms in Formula
3, of the form in 5, become zero or infinity under these conditions. For this
reason, speakers with categoricalcells were eliminatedfrom the comparison.
As it happens, the five individualswith the greatestnumberof dataare included
in the sample chosen in accord with the above requirements.
Table 12 reports the results for the morphologicalhierarchyin individuals.

CLASS FREQUENCY PROB FREQUENCY PROB


Speaker:ELV Speaker:UBI
la 9/92 = 9.8% 0.08 12/115 = 10.4% 0.08
b 102/297 = 34.3% 0.29 89/219 = 40.6% 0.35 *
c 20/32 = 62.5% 0.53 = 10/41 = 24.4% 0.18
2a 56/91 = 61.5% 0.57 = 86/107 = 80.4% 0.78 =
b 25/37 = 67.6% 0.73 25/30 = 83.3% 0.82 =
c 132/152 = 86.8% 0.87 93/110 = 84.6% 0.86
Speaker:FAU Speaker:HEN
la 1/62 = 1.6% 0.03 5/36 = 13.9% 0.16 =
b 50/380 = 13.2% 0.22 31/144 = 21.5% 0.24 =
c 21/48 = 43.8% 0.60 = 3/17 = 17.7% 0.19
2a 46/89 = 51.7% 0.68 25/54 = 46.3% 0.51
b 29/42 = 69.1% 0.83 = 14/18 = 77.8% 0.90 =
c 152/197 = 77.2% 0.89 49/59 = 83.1% 0.88
Speaker:SON Speaker:SID
la 26/53 = 49.1% 0.23 20/59 = 33.9% 0.18
b 175/223 = 78.5% 0.54 * 107/210 = 51.0% 0.31 =
c 1/5 = 20.0% 0.06 10/15 = 66.7% 0.44 =
2a 55/62 = 88.7% 0.70 = 66/86 = 76.7% 0.62
b 34/36 = 94.4% 0.85 = 35/39 = 89.7% 0.82 *
c 130/144 = 90.3% 0.76 74/96 = 77.1% 0.63
TABLE12.

15
It was not possible to impose this requirementon the older sample, since
only one older
speakerhad vicariousorientation(CAN), and he providedtoo few data tokens to'makethe com-
parisonworthwhile.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
90 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

Both rawfrequenciesand the logistic modelprobabilitiesare shown. The equals


sign following a probabilityindicates that the difference between it and the
next probabilityin the hierarchy(listed immediatelybelow) is not statistically
valid at the 0.1 level, a situationthat I shall referto as 'neutralized'.An asterisk
indicatesthat the differencebetween categories is valid at 0.1 or a better level,
but is the opposite of whatwouldbe expected fromgroupbehavior.No notation
is made when the difference between probabilitiesassociated with a class and
the next class in the hierarchyis statisticallyvalid and agrees with the group
order.
In general, this table reveals that the behavior of individualsis in accord
with the group pattern, but there are exceptions. Thus UBI's value for Class
Ic is lower than for lb, and the same is true of SON; while SID's Classes 2b
and 2c are in the wrong order. Of the thirteen neutralizedpairs, ten are in
accord with the group order, while only three are in the opposite order. Thus,
of the thirty neighboringpairs in Table 10, fourteen are correct, thirteen are
neutralized(but still about 80%in the correct order), and only three are in the
wrong order. In the entire table, only one result is more than one rank away
from the group order: this is SON's value of 0.06 for Class Ic, which is not
only lower than her value of 0.54 for Class lb, but also lower than her value
of 0.23 for Class la. Significantly,it is precisely SON's Class Ic that has the
fewest data tokens in the table (only five).
It is instructiveto consider furtherthe question of numberof data tokens-
althoughthis is not, of course, the only determinantof statisticalsignificance.
Since a comparisonof values for membersof a pair is only as accurate as the
LEASTaccurate of the values involved, the number of tokens in the SMALLEST
cell of each pair is the numberof empiricaldata tokens upon which the com-
parison rests. For example, the pair la-lb in ELV is taken as based upon 92
tokens (the total for la) rather than 297 tokens (the total for lb). Scoring in
this manner, the fourteen correct pairs average 47.1 tokens, the thirteen neu-
tralized pairs, 30.2 tokens; and the three incorrect pairs, 28.3 tokens. Fur-
thermore, the results for the positional hierarchyare exactly parallel:of the
twelve neighboringpairs of Case 1, eight are correct and show an average of
20.5 tokens, while four are neutralized(but in the correct order) and average
only 14.0 tokens. As these figures show, it is not possible to establish a fixed
requirementfor statisticalsignificancein terms of numberof tokens. The prin-
cipal reason is that the degree of differenceof behaviorbetween categories is
also a strong determinantof statistical significance. We can conclude that
apparentindividualdeviations from the group patterntend to occur when the
numberof tokens is small, relative to the numberof tokens in other cells-
and that, for the most part, individualsdo conform to the group pattern.-
8. THE ORIGIN
OF LOSSOFVERBPHRASE
AGREEMENT.
Loss of agreement in
the VP involves changes of very disparate sorts on the surface level. In the
classes on the upper end of the morphologicalhierarchy,the process appears
to be strictly one of generalizationof privilege of occurrence of the singular
morphemesat the expense of their plural equivalents (thus, in Class 2c, [-6]

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 91

seems to be taking over the territoryof [-arf]). But on the lower end of the
morphologicalscale, this same process appears to be the result of a simple
phonologicalrule that denasalizes final vowels (e.g., in Class la, [i] becomes
[i]). In fact, such a rule exists in BrazilianPortugueseand applies variablyto
classes other than verbs (i.e. to nouns such as vagem 'stringbeans', adverbs
such as ontem 'yesterday', etc.) Furthermore,this rule is attested even in the
oldest texts, and thus predates the evolution of a non-agreeingsystem by at
least several centuries. In some areas of Portugal,principallyin the region of
Entre-Douro-e-Minho,the rule survives in popular speech, where its effects
have been reported in both nouns and verbs of Class la (virgem 'virgin' as
birge; vertem 'they spill' as berte-Vasconcellos [1901] 1970:86-7), although
no general variability in agreement is mentioned for the areas in question.
However, the view of loss of agreementas a consequence of a phonological
change is fully applicableonly in Class la; from Class lb upward, more than
the simple loss of nasalizationis involved, and the process can no longer be
viewed as a phonologicalrule.
The developments in real time mentionedabove lead one to investigate the
correspondingphenomenain apparenttime (i.e. in age groups). For this pur-
pose, Table 13 lists, for all three models, the results for Class la in the older
and younger groups, and compares these results with Class lb, the next cat-
egory. The results of the logistic model for these two categories are shown to
be statistically significantat the .005 level by the log likelihood test.

GROUP: OLDER GROUP: YOUNGER


MODEL CLASS VALUE DIFFERENCE VALUE DIFFERENCE
Frequency la 11.4% 17.0%
b 31.4% 29.3%
20.0% 12.3%
Logistic la 0.10 0.11
b 0.31 0.23
0.21 0.12
Additive la -36.8% - 31.0%
b - 16.6% - 19.2%
20.2% 11.8%
TABLE 13.

In all models, the older speakers evidence a wider gap between these two
classes; this indicates that, at earlierstages of the history of loss of agreement,
Class la was even further in the forefront of non-agreementthan it is now.
Since it is precisely Class la that represents the uniquely phonologicalcom-
ponent of the change, this in turn suggests that the loss of final nasalization
was once responsible for a largerproportionof the over-all process, and may
have been an actuatingforce behind it.
If we look at the same problemfrom a differentperspective, it is instructive
to compare loss of nasalizationin verbs with that in other grammaticalcate-
gories. In verbs, several distinct subcategories of nasalizationloss must be
considered. The results of denasalizationare summarizedin Table 14.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
92 LANGUAGE, VOLUME57, NUMBER 1 (1981)

CLASS FORM [ + NASAL] [- NASAL]


1 comem [k6mi] [k6mi] (= sg.)
falam [fala] [fala] (= sg.)
2 falam [falu] [falu] (sg. fala)
fazem [fazi] [fazi] (sg. faz)
3 comeram [kom6rfi] [kom6ru] (sg. comeu)
falaram [falarfi] [falaru](sg. falou)
TABLE14.

In Class 1, the forms of the [ + nasal] column are plural, while those of the
[- nasal] column are singular;here the loss of nasalizationtotally destroys the
opposition. In Class 2, the forms of both columns must be classified as plurals,
since they are both distinct from the singular,and the loss of nasalizationdoes
not destroy the opposition. Finally, in Class 3, as in Class 2, both columns are
plural.
Votre 1978, 1979has studied the general process of denasalization-in par-
ticular, the relationshipof the rates of retention of nasalizationin the three
verbalclasses of Table 14 vs. that in nouns. In these considerations,the nouns
serve as a sort of control against which the behavior of the verbal classes is
measured.
Table 15 shows Votre's results (1979)for retentionof nasalizationin two age

MODEL CLASS OLDER YOUNGER

Logistic Verbs 1 0.24 0.17


2 0.83 0.88
3 0.67 0.43
Nouns 0.24 0.46
Frequency Verbs 1 108/668 = 16.2% 31/202 = 15.3%
2 155/258 = 60.1% 24/44 = 54.5%
3 110/301 = 36.5% 10/99 = 10.1%
Nouns 454/683 = 66.5% 244/359 = 68.0%
TABLE15.

groups, correspondingto subsets of the same speakersstudiedin the mainpart


of this paper. Consider first the older speakers. In terms of logistic model
probabilities,the four classes of Table 15 show a clear split into two levels-
a high rate of retentionof the nasal in verbs of Classes 2 and 3, and a low rate
in Class 1 verbs and nouns. As far as the verbal classes are concerned, the
relatively frequent absence of the nasal in Class 1, where it constitutes the
pluraldesinence, is in accord with the principlethat 'deletion of a morphemic
segment is more likely ... if the morphemeis redundant'(Naro 1980:166).In
this case, the plural desinence is most often determinedby a previously oc-
curringplural subject within the context of the discourse. The nasals in the
other two verbal classes are not redundant,and are thus more resistant. The
relatively low rate of nasal retention in nouns, however, is specific to the
history of Portuguese, which has a general long-termdrift toward variability
in this class on both sides of the Atlantic. The most importantpoint to note,

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 93

however, is the fact that, in the older speakers, loss of nasalizationin Class
1 verbs, which is equivalentto loss of subject/verbagreement,is no morelikely
than loss of nasalizationin nouns.
There are some strikinginconsistencies between the logistic model proba-
bilities and the correspondingraw frequencies in Table 15. In particular,al-
though Class 1 verbs and nouns show equal probabilitiesof nasal retention,
their frequencies for this same phenomenon are at opposite extremes, with
Class 1 verbs at 16.2%and nouns at 66.5%. The reason behindthis seemingly
strange situation is the concurrenteffect of stress. In verb Class 1, the nasal
is always in an unstressed syllable, while in nouns it is often in a stressed
syllable (alguem 'someone', mafa 'apple', etc.) This is a very important distri-
butionalproperty, since stressed nasals are nearly always retained(frequency
434/444 = 97.1%, probability0.95), while unstressed nasals often disappear
(frequency 393/1466 = 26.8%, probability0.05). Thus the difference in fre-
quencies between these two columnsin the older speakersis caused by a strong
correlationwith stress, ratherthan by a loss of nasalizationpeculiarto one or
the other class.
In the younger speakers, the situationis quite different.For this group, the
frequencies and probabilitiesagree in attributinga considerablylower rate of
nasal retentionto verb Class 1 thanto nouns. Furthermore,verb Class 3, which
had a higher rate of retention than nouns for the older speakers, is equal to
Class 1 for the younger speakers. This means that in the younger group, ac-
cording to the logistic model, verb Class 1 does have its own unique loss of
nasalization.In other words, in this age group, the loss of nasalizationin Class
1 verbs, and consequently of subject/verbagreement,can no longer be attrib-
uted to a correlationwith stress, even though stress remainspowerful (in the
logistic model, 0.93 for stressed and 0.07 for unstressed).Similarly,the peculiar
loss of nasalizationin verbs has spread,althoughnot as strongly,to verb Class
3, where nasalization is not redundant.Verb Class 2, however, has not yet
been affected. From these facts we may conclude that, in the youngergroup,
the loss of nasalizationin Class 1 verbs has at least two components-a general
denasalizationthat can be seen in the nouns, and a specific morphologicalshift
in favor of singulardesinences. The over-all view that emerges is this: within
the class of verbs, the phonologicalprocess of denasalization(and consequent
loss of agreement) made its first inroads at the weakest point (i.e. Class 1,
where the nasalizationis syntacticallydetermined),later generalizingto other
environments.
In general terms, the evidence indicates that, in the earlier stages of the
evolution of verbal agreement, environmentsinvolving simple loss of nasali-
zation were furtherin the forefrontof agreementreductionthanin laterstages-
and that, in the earlier stages, nasalization loss in Class 1 verbs was more
similar(in fact, nearly equal) to nasalizationloss in non-verbs. Extrapolating
backward,I conclude that one of the actuatingforces that set off the evolution
toward a non-agreeingmorphologicalsystem was the phonologicalrule of de-
nasalization.This rule was, and continues to be, variable.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
94 LANGUAGE, VOLUME 57, NUMBER I (1981)

9. SALIENCYHIERARCHIES
IN NOUN PHRASEAGREEMENT.As in the VP, the
agreement rule in the NP is disappearingin modern Brazilian Portuguese.
However, this change has a wider socio-economic distribution,and seems to
be furtherevolved. Within the NP, variationin the realizationof plural mor-
phemes is extensive, even in highly educated speakers who show nearly cat-
egorical markingof the verb.
The behaviorof nominalpluralshas been studiedin the area of MinasGerais
by Braga 1977, and in the area of Rio de Janeiroby Scherre 1978. These two
investigations produced essentially parallel results; but since the latter work
is partiallybased on a subgroupof the speakers in our sample, it will be cited
here. Scherre's illiterate speakers were all selected from among the younger
group. The over-all frequencies are given in Table 16 (Scherre. p. 98).
NAME APPLICATIONS TOTAL FREQUENCY
VAN 1084 1588 68.3%
SID 923 1365 67.6%
SON 619 947 65.4%
HEN 518 992 52.2%
IZE 587 1144 51.3%
JOM 738 1559 47.3%
TABLE16.

In addition, Scherre presents data from four speakersof the same age level,
but of higher socio-economic and educationalstatus. These speakers average
a frequency of 80.0% in applicationof agreement.
A hierarchyof saliency of the singular/pluralopposition, similarto that of
Table 3, was set up for the nominalsand postulatedas a constraintgroup. Pairs
in which the plural mark is in an unstressed syllable were assigned to Level
1, while Level 2 was reserved for the more salient case in which a pluralmark
occurs in the stressed syllable. WithinLevel 1, two classes were distinguished.
In Class la, the plural differs from the singularmerely by the addition of a
final -s (ano/anos 'year/years'), but in Class lb the plural form acquires an
extra syllable throughthe additionof-es (lugar/lugares'place/places').16Level
2 also has two subclasses. Class 2a consists of pairs in which the plural -s
attaches to the stressed final syllable, usuallywith furthermorphophonemically
determinedchanges: ancestral/ancestrais[-aw/-ays] 'ancestor/ancestors',dis-
cussdo/discuss6es [-iaw/-6ys]'discussion/discussions'.At the top of the hier-
archy is Class 2b, containingthe so-called metaphonicplurals, in which there
is a very salientchangein the qualityof the stressed root vowel with concurrent
additionof -s. (ovolovos [dvu]/[5vus]'egg/eggs'). A summaryof the hierarchy
of oppositional saliency is as follows:'7
16
Items of the type vez/vezes 'time/times',in which the -es follows anothersyllable endingin
a sibilant, were not includedin Class Ib: they undergoa haplologicalprocess of reduction,in-
dependentof the generalagreementrule.
17 The final -s of Classes
lb, 2a, and 2b may be deleted. In Scherre'sdata, the correspondence
to stress is not perfect, since she includespluralsof the type cafes 'coffees' in Class la and of the
typefdceis 'easy' (pl.) in Class 2a. Nonetheless, these items occur infrequently.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONS OF A SYNTACTIC CHANGE 95

(9) LEVEL1 (unstressed):


Class a. [-k/-s]
b. [-0/-es]
LEVEL2 (stressed):
Class a. [-aw/-ais, -ew/-eis, -iw/-is, -aw/-6os, -aw/-as, -aw-aws]
b. [6/3-s]
The results obtainedby Scherre(p. 79) for the morphologicalvariablein the
two socio-economic groups she studied are representedin Table 17.

CLASS EXAMPLE FREQUENCY PROB


Educated speakers
la ano/anos 2705/3365 = 80.4% 0.39
b lugar/lugares 76/86 = 88.4% 0.57
2a discussaoldiscussoes 116/122 = 95.1% 0.85
b ovo/ovos 21/21 = 100.0% +
Illiterate speakers
la anolanos 4291/7240 = 59.3% 0.29
b lugar/lugares 37/80 = 46.3% 0.25
2a discussaoldiscussoes 39/101 = 38.6% 0.49
b ovolovos 30/34 = 88.2% 0.96
TABLE 17.

Consideringfirst the educated group, we see a perfect correlationbetween


increasing oppositional saliency, on the one hand, and increasingfrequency
or probabilityof actualizationof a pluralmark, on the other. For the illiterate
group, the logistic model probabilitiesreveal a neutralizationof effect on Level
1 (unstressed), but this level can be clearly distinguished from Level 2
(stressed), which is not neutralized.
As regards the morphological constraint group of the illiterate speakers,
there is a discrepancy between the relative orderingsof the probabilitiesand
the frequencies. This is caused by the distributionof these constraints with
respect to the positional group. In NP's, the first element, labeled as position
0, nearlyalways carriesan actualizedpluralmark,while in succeedingpositions
the chances of pluralmarkingdecrease, as shown in Table 18.
POSITION FREQUENCY PROB
0 3212/3289 =
97.7% 0.98
1 1172/3847 =
30.5% 0.45
2 84/416 =
20.2% 0.31
3 1/39 =
2.6% 0.06
4 4 =
0.0%
TABLE18.

These results for the nominalposition groupare importantfor my purposes,


since they confirm the operation of the principle of saliency along the same
lines as in the verbal positional group. The interactionwith the morphological
group results from the fact that position 0, in which markingis nearly cate-
gorical, is almost always occupied by an article(os, as 'the') or a demonstrative
(estes 'these', aqueles 'those'), both of which belong to MorphologicalClass

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
96 LANGUAGE, VOLUME 57, NUMBER 1 (1981)

la. Thus the raw rate of agreementfor Class la is inflatedbecause a relatively


higher proportionof la tokens occurs in position 0 than tokens of the other
morphologicalcategories.
The results of Table 17 suggest that there is no relationshipbetween fre-
quency of occurrence of a paradigmaticclass and the correspondingrate of
agreement. In the educated group, for example, the most frequent class (la,
with 3365 tokens) has the lowest rate of agreement, while the least frequent
class (2b, with only 21 tokens) shows categoricalagreementin the sample. The
situation is parallel in the case of VP agreement-referring back to Table 3,
the highest rate of agreementis found in the least frequentclass (2e, with 116
tokens), while the most frequentclass (Ib, with 2540tokens) shows the second
lowest rate of agreement. It might seem, however, that in considerationsin-
volving frequency, the appropriateunit would be the individualmorphological
item rather than paradigmaticclasses. Indeed, the item ser 'to be' (elsdo 'is/
are', Class 2d of Table 3) has a high frequency of occurrence and a high rate
of agreement, while the individual verbs that compose the less frequently
agreeingclasses would each also have a low frequency. Note, however, that
Class 2e has a low frequency but a rate of agreementthat exceeds even that
of Class 2d. Furthermore,the subclass of 2e consistingexclusively of the items
disseldisseram 'said (sg./pl.)' and trouxeltrouxeram 'brought (sg./pl.)' (com-
bined with Class 2e in this paper precisely because of low frequency; see fn.
8) shows agreementin 28 of the 31 (= 90.3%)instances in the data. Thus, even
though these two morphologicalunits occur very infrequently, each has a
higherrate of agreementthan the very frequentelsdo. AlthoughI do not have
data for individualitems on the lower extreme of the hierarchy,the results so
far available make it seem unlikely that further investigationwould reveal a
correlationbetween frequency of occurrence and rate of agreementon either
the paradigmaticor individualitem levels.
10. CONCLUSION: THE ROLE OF SALIENCY IN SYNTACTIC CHANGE. Two aspects
of the over-all process of linguisticchange may be distinguishedfor analytical
purposes: (a) ACTUATION, i.e. the beginning point, or first context, of a change;
and (b) DIFFUSION,i.e. the subsequent spread of the change to other environments.
In the loss of verbalconcord, I have arguedabove thatthe linguisticactuating
force was a low-level phonetic rule of denasalizationof final vowels. For the
verbs of MorphologicalClass la, this rule, operatingupon a pluralform such
as comem [komi], produces a form that coincides exactly with the singular.18
Thus, for speakers having this rule, the singular/pluraloppositioncollapses on
the surface in a certain part of the language.The diffusionof the non-agreeing
system throughoutthe languagethen proceeds, in accord with the principleof
saliency, along several distinct dimensions-extending itself most strongly in
those contexts where it involves the smallest (i.e. least noticeable) change in
surfaceform. The change thus sets in at the zero point of surfacedifferentiation
between the old and the new systems, and spreads to other points along the
'8 This is also partially true of Level lb, where some speakers may show a [fala] / [fala]
opposition
rather than the more usual [fala] / [falu].

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
DIMENSIONSOF A SYNTACTICCHANGE 97

path of least surface differentiation.This view of the process reinforces and


confirms the general idea of syntactic change as a surface phenomenon(Naro
1976), according to which change originatesin contexts that admit more than
one syntactic analysis with no change in surface form.
Since salience is a general propertyof the linguistic system, one would not
normallyexpect groups of speakers at different stages of evolution to exhibit
conflicting orderings of variable constraints in the hierarchizeddimensions,
althoughsuch constraintscan be neutralizedundercertaincircumstances.The
principaldifference between a highly advanced and a less advanced group of
speakers should lie in the general propensity to use the rule. In other words,
as noted in ?7, the relative impedance of the hierarchizedconstraints should
be the same across speakers, because the forms involved-and thus their
relative salience-are the same. However, the general level of use of the rule
might vary from group to group without alteringthe relative orderingof the
constraints. In fact, this is exactly what we have found to be the case in the
study of agreementin the NP and VP.
Thus we conclude with a general view of a majortype of syntactic change
as originatingin an environment,or set of environments,where the structure
of the rest of the languageis such that more than one construal(i.e. syntactic
derivation)of the same surface structureis possible. However, even though
this particularoriginatingenvironment may involve no (or only very incon-
spicuous) surfacedifferencesbetween the old and new analyses, consequences
in other parts of the languagewill lead to the creation of forms (or whole new
constructiontypes) not previouslypresent. These new items do not, in general,
enter the language simultaneously,or even at the same rate-despite the fact
that, from the linguist's point of view, they form a 'natural' set or logical
consequence of the new analysis. Rather, the principle of saliency, possibly
along with other universal principles or language-particular restrictions, gov-
erns the introduction of new elements. In graphic terms, syntactic change
begins at a corner, thence spreadinginward, i.e. from a region of zero differ-
entiationbetween the old and the new constructionsto regions of successively
greatersurfacedifferentiation.In this way, the new structureand its associated
propertiescan be partiallyblocked for a certainperiod of time. It is only in the
long run, not in its initiationor progress, that linguistic change produces rel-
atively uniformand 'natural'results.
REFERENCES
BRAGA, MARIALUIZA. 1977.A concordanciade nimero no sintagmanominalno Triangulo
Mineiro. PUC/RJmaster's dissertation.
CAMARA, JOAQUIM MATTOSO. 1965. A propos d'un vulgarismedu portugaisdu Bresil.
Omagiului AlexandruRosetti, 534-35. Bucharest:EdituraAcademiei Republicii
Socialiste Romania. [Portuguesetranslationin Dispersos de J. Mattoso Camara
Jr., ed. by Carlos Eduardo Falcao Uch6a, 89-93. Rio de Janeiro: Editora da
FundagaoGetfilio Vargas, 1975.]
--. 1972. The Portugueselanguage.Translatedby Anthony J. Naro. Chicago: Uni-
versity of Chicago Press.
CEDERGREN, HENRIETTA J., and DAVID SANKOFF. 1974. Variablerules: Performanceas
a statisticalreflection of competence. Lg. 50.333-55.

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions
98 LANGUAGE, VOLUME57, NUMBER 1 (1981)

GuY, GREGORY R. 1975. Variationin the group and the individual:The case of final
stop deletion. (Pennsylvaniaworking papers on linguistic change and variation,
1:4.) Philadelphia:US Regional Survey.
KAY, PAUL. 1978. Variable rules, community grammar, and linguistic change. Linguistic
variation:Modelsand methods,ed. by DavidSankoff,71-83. New York:Academic
Press.
--, and CHADK. MCDANIEL.1979. On the logic of variable rules. Language and
Society 8.151-88.
KIMBALL,JOHN. 1973. Seven principles of surface structure parsing in natural language.
Cognition 2.15-48.
LABOV, WILLIAM. 1969. Contraction, deletion, and inherent variability of the English
copula. Lg. 45.715-62. [Revised version in his Languagein the innercity: Studies
in the black English vernacular,65-129. Philadelphia:University of Pennsylvania
Press, 1972.]
LEMLE, MIRIAM,and ANTHONYJ. NARO. 1977. Competencias basicas do portugues. Rio
de Janeiro: Fundacao Movimento Brasileiro de Alfabetizacao. [Available from
MOBRAL-CETEP/CEDOC,Ladeira do Ascurra 115, Cosme Velho, Rio de Ja-
neiro, Brasil.]
NARO, ANTHONYJ. 1976. The genesis of the reflexive impersonal in Portuguese: A study
in syntactic change as a surface phenomenon.Lg. 52.779-810.
--. 1978. The SWAMINC4program.Unpublished,Departamentode Letras, PUC/
RJ.
--. 1980. Review article on Linguisticvariation:Models and methods, ed. by David
Sankoff. Lg. 56.158-70.
-- , and MIRIAM
LEMLE.1976. Syntactic diffusion. Papers from the Parasession on
Diachronic Syntax, 221-39. Chicago: CLS. [Reprinted in Ciencia e Cultura
29.259-68.]
NIE, NORMAN H., et al. 1975. SPSS: Statistical package for the social sciences. 2nd ed.
New York: McGraw-Hill.
ROUSSEAU, PASCALE, and DAVIDSANKOFF. 1978. Advances in variable rule methodology.
Linguistic variation: Models and methods, ed. by David Sankoff, 57-69. New
York: Academic Press.
SANKOFF, DAVID. 1975. VARBRUL2. Mimeographed, Universite de Montreal.
SCHERRE, MARIA MARTAPEREIRA.1978. A regra de concordancia de nfimero no sintagma
nominalem portugues.PUC/RJmaster's dissertation.
VASCONCELLOS, J. LEITE DE. 1901. Esquisse d'une dialectologie portugaise. Paris: Ail-
laud. [2nd ed., ed. by Maria Adelaide Valle Cintra. Lisboa: Centro de Estudos
Filologicos, 1970.]
VOTRE, SEBASTIAOJOSUE. 1978. Variacao fonologica na fala dos mobralenses da area
do Rio de Janeiro. PUC/RJdoctoraldissertation.
-- . 1979. Phonologicaland syntactic aspects of denasalizationin spoken Brazilian
Portuguese.Paper presented at NWAVE8, Universite du Quebec a Montreal.
[Received 20 March 1980.]

This content downloaded from 128.226.37.5 on Mon, 21 Dec 2015 14:10:20 UTC
All use subject to JSTOR Terms and Conditions

You might also like