Download 5602 - Radboud Repository

PDF hosted at the Radboud Repository of the Radboud University
Nijmegen
The following full text is a publisher's version.
For additional information about this publication click this link.
http://hdl.handle.net/2066/15430
Please be advised that this information was generated on 2017-06-19 and may be subject to
change.
Chapter 11
Semantic and Syntactic Aspects of
Remembering Sentences: a Review
of Some Recent Continental Research
W. J. M. Levelt and G. Kempen
T he study of sentence m em o ry and retrieval, which had becom e a d o m in an t
issue in the psycholinguistic work of the sixties, u n d erw en t a m a jo r change at
the approach of the seventies. It need not be explained here that the original
theorizing was m ore or less directly derived from norm al syntactic notions in
linguistics (for a review see Levelt, 1 9 7 3 -1 9 7 4 ). T h e change was induced by
renew ed attention to sem antic and extralinguistic aspects of internal re p re s e n ­
tations. It becam e doubtful w h e th e r an implicit assumption of the earlier work
could be m aintained, namely the presupposition that what is stored in m em ory
is some sort of linguistic object. A n d one can now rightly question, as Flores
d ’Arcais (1 9 7 4 ) does in the title of a recent paper, w h e th e r there is a m em ory
for sentences at all.
T h e alternative point of view has been expressed most attractively by
B ransford and Franks (1972), and by Barclay (1973). T heir position is that a
person who reads and tries to m em orize a sentence or text does this by building
an internal rep resen tatio n of the object, action or situation which is described
in the text. In case the text contains m ore than a single sentence, the person
integrates the contents of the different sentences into a ‘holistic’ semantic
rep resen tatio n in which there is no trace of the original syntactic boundaries.
T h e linguistic structure of the stimulus material is normally quickly lost. This
can only be p rev en ted by giving the subject the explicit additional task of
verbatim re p ro d u ctio n (Fillenbaum, 1966; Sachs, 1967; Flores d ’Arcais,
1974).
A seemingly obvious consequence of these d ev elopm ents is a devaluation of
the paradigm in which verbatim sentence reproduction is required of the
subject. In this way, it is said, the subject is forced to store the linguistic object,
namely the sentential form, over and above w hat he normally extracts from it.
Since only the latter process is natural and interesting, the paradigm only
diverts attention to an artifact. Such would be the case for J o h n s o n ’s (1965)
202
results which seemingly indicate that surface constituent structure is stored into
m em ory, but from which it cannot at all be concluded that normally such
storage takes place, let alone that this would be the form at in which sentential
material is recalled. A n additional arg u m en t for this interpretation might be
derived from J o h n s o n ’s later (1970) work, in which he can correctly predict
transitional errors in the recall of nonsense letter strings from the way letters
are grouped, just in the same way as he could predict transitional errors in the
recall of sentences from the way words are g rouped into constituents.
Intuitively, however, the m em orization of a meaningful sentence is clearly
different from the m em orization of a nonsense letter string. J o h n s o n ’s original
study might therefore have concerned an artificial grouping effect which had no
relation to sentence m e m o ry per se, but only to the additional task of verbatim
recall.
A similar critique seems applicable to B lu m e n th a l’s (1 9 6 7 ) work. T h o u g h
this a u th o r suggests that sentences are coded in terms of their deep structures
instead of their surface forms, it might again be due to the verbatim recall
task that subjects are forced into creating such codes. T h ey may have little
relation to what a subject normally abstracts from a sentence. (It should, by the
way, be noted that B lu m e n th a l’s results could also be an artifact for other,
purely experim ental reasons. If one repeats his p r o m p te d recall paradigm for
am biguous sentences giving rise to the same case shift as in the John is
eager/easy to please sentence pair, the original effect disappears, as Levelt and
B onarius (1973) have shown for D u tch and mutatis m utandis for Finnish.)
T h o u g h these arg u m en ts unm istakably have som e face value, one has to be
careful not to throw o ut the baby with the bath water. T he verbatim recall
paradigm is not only attractive from the ex perim ental point of view of easy
scoreability, but there is also strong evidence that it does not necessarily create
the artificialities. O n the one hand one can use the paradigm for the analytic
study of the role of sem antic factors in the recall of sentential or text material.
In the next p a ra g ra p h we will discuss a few, mostly continental, studies along
these lines. Some of these studies have up till now only tak e n the form of
unpublished dissertations or reports. O n the o th e r hand, we will show that, just
in the context of the verbatim recall task, it is pro b a b ly not the case that
subjects base their rep ro d u ctio n s on an internal sentential re p re se n ta tio n
(surface or phonetic), this in spite of the fact that they do show clear constituent
b o u n d a ry effects. It will be argued, again on the basis of as yet mostly
unpublished continental work, that such effects are at least partly due to
syntactic retrieval plans, not to syntactic traces in m em o ry . This is d o n e in the
second section of the chapter. In a last, concluding section we will try to relate
these findings to som e diverse n o n -c o n tin en ta l studies; this may be tak e n as an
exercise in bridge building.
Semantic Effects on Verbatim Reproduction of Sentences
O n e of the first to d e m o n s tra te sem antic effects in verbatim re p ro d u c tio n of
sentences was R o s e n b e rg (1968). H e studied the o p e ra tio n of a factor which he
203
called ‘semantic in teg ra tio n ’ and which is d e te rm in e d by interw ord associations
within the sentence to be re m e m b e re d . A case of high semantic integration is
the sentence The old king ruled wisely as o p p osed to The poor king dined
gravely , which has little semantic integration. R o se n b e rg showed that the
p attern of transitional errors only reflected constituent b o u n d a ry effects in the
case of weakly integrated sentences, i.e. m ore transitional errors betw een than
within constituents. F or highly integrated sentences, it seem ed, subjects were
apparently able to construct larger units than (m a jo r) constituents.
In a follow up of this finding, H o r m a n n and E n g e lk a m p started a m ore
analytic ap p roach to this m atter. T h ey tried to isolate one factor in semantic
integration which they called ‘sem antic im plication’ ( H o r m a n n , 1971; E n g e l­
kam p, M erdian and H o r m a n n , 1972). In the latter p a p e r sem antic implication
is defined as a p ro p e rty of a su b ject/p red icate/o b jec t sentence in the following
way: sem antic implication of such a sentence is high if it is hard to change one of
the three constituents given the o th e r two; it is low in the o th e r case. An
exam ple of high sem antic implication is the sentence The river erodes the b a n k ,
since there are only few o th e r things than rivers which ero d e banks, few o th e r
things than banks that are e ro d e d by rivers, and few o th e r things than eroding
which rivers can do to banks. Low sem antic implication, it is re m a rk e d , is the
case for the sentence The pupil finds the book. It will not surprise that the
authors are indeed able to show that learnability increases with degree of
sem antic implication. H o w ev er, to g eth er with this factor E n g e lk a m p et al.
varied a second factor that might contrib ute to sem antic integration, namely
negation. T h e y theorized that sem antic implication would only contribute to
sem antic integration if the implication is not denied. T hey did an e x p e rim e n t in
which the subject had to give v erbatim re p ro d u ctio n of the sentence on the
occasion of a p r o m p t word from the sentence. T h e ir general finding was that
high-implication sentences were re p ro d u c e d b e tte r than low-implication sen ­
tences, b ut only for affirmative sentences. F or negative sentences sem antic
implication show ed no effect. It is necessary, however, to m a k e one qualifica­
tion. T h e ex p erim en tal material contained three types of negation: negation of
subject, of predicate, or of object. Instances of these are:
Not the river has eroded the bank (subject negation)
The river has not eroded the bank (predicate negation)
The river has eroded not the bank (object negation).
(T he latter construction is quite norm al and acceptable in G e r m a n .) It tu rn ed
o u t that the interaction of negation and sem antic implication only occurred in
the latter two cases. Subject negation did n ot reduce the effect of sem antic
implication on sentence rep ro d u ctio n . T h e au th o rs relate this to the relative
in d e p e n d e n c e of the subject constituent, even in sentences with high sem antic
implication (this in d e p e n d e n c e is given credit in linguistics by the classical
su b ject/p red ica te phrase dichotom y). In o u r opinion a m o re obvious e x p la n a ­
tion should be explored first. It might well be that the finding is due to a
204
property of the sentential material used in the experim ent. It may, f u r th e r­
more, be the case that the subject can ignore the negative elem en t m ore easily if
it precede (or follow) the rest of the sentence. In the experim ental sentences
this is only the case for subject negation, but in G e rm a n this could easily be
done for predicate (verb) negation as well:
Der Schüler findet das Buch nicht
(The pupil does not find the book)
In his dissertation E n g e lk a m p (1 9 7 3 ) ex tend ed the notion of semantic
implication to relations betw een constituents o th e r than subject, predicate and
object. In fact he analysed sentences in terms of case relations and expressed
semantic implication in terms of case structure. In one of his ex perim ents he
used sentences such as (a), (b) and (c):
(a) The tradesman with the merchandise hit the boy
(b) The tradesman with the spot hit the boy
(c) The tradesman with the stick hit the boy
A ccording to E n g e lk a m p ’s analysis the type (a) sentence contains two
relation terms, namely trade with argum ents (cases) man and merchandise in
subject and object position, and hit with man and boy as arguments:
(a) trade ( m a n , merchandise)
hit (m an, boy)
H ere, the prepositional phrase with the merchandise participates in the trades
relation. In the type (b) sentences the prepositional phrase expresses an
in d e p e n d e n t relation that we will characterize by the verb possess:
(b) trade (m an)
possess (m an, spot)
hit (m an, boy)
In sentences of type (c) it seems that the prepositional phrase relates to the
main verb as an instrument:
(c) trade (m an)
hit (m an, boy , stick)
E n g e lk a m p found in a verbatim recall ex p erim en t that transitional errors
b etw een tradesman and with , i.e. within the grammatical subject, were highest
for type (b) sentences, in accordance with his analysis. In that sentence type,
w here the prepositional phrase is neither integrated with trade , nor with hit , it
has low sem antic implication.
205
Studies such as these certainly d e m o n stra te the effect of semantic variables
on verbatim sentence recall and give a first theoretical analysis, though a
n u m b e r of new problem s arise. A re the semantic effects on reproduction due to
the m e m o ry code or to some retrieval strategy? Is the degree of semantic
implication m ore than or different from the sum of word association strengths,
i.e. is it possible to construct sentences for which the associations betw een
words are controlled and equal and where there is nevertheless a difference in
sem antic implication? Is semantic implication different from combinability of
phrases? This latter question needs some explanation which can be done by
m eans of a n o th e r continental study (Levelt, 1967; N o o rd m a n and Levelt,
1970). In this one, O s g o o d ’s (1970) ‘word intersection’ m e th o d was used. It
consists of having subjects judge the acceptability of word com binations such as
v erb /ad v erb ( apologize proudly , kill instantly , etc.), adjective/noun (lazy stone,
diligent nurse , etc.) n o u n /v erb (cars eat , children play , etc.). T h e intention is to
infer from such ju d g m en ts som ething a b o u t the feature structure of the words
involved. T h e underlying notion is that acceptability is low in the case of words
o p p o sed on some feature. For lazy stone the critical feature might be
animateness. Levelt (1 9 6 7 ) studied the com bination of interpersonal verbs and
adjectives. N o o r d m a n and Levelt (1970) analysed v e rb /n o u n com binations in
the sentence frame ‘T h ey verbed the noun (e.g., they received the growth). T he
study involved 13 verbs and 480 nouns and all com binations were judged. It
turned out that a specification on only four features was sufficient to predict
nearly all acceptabilities (errors: 3*5 per cent), though these features were not
sufficient to predict all non-acceptibilities; in fact, m any of them were predicted
as acceptable (errors: 24 per cent). T h e four features were concrete/abstract,
living/non-living, h u m a n / n o n - h u m a n , and generic/non-generic. Findings of
this sort may lead to further analysis of the notion of sem antic implication. Can
sem antic implication be partly or fully expressed in terms of (non-) opposition
of certain general sem antic features? D oes verbatim recall of sentences
increase if certain case or modifier relations go with feature similarity betw een
the argum ents?
A n o t h e r ap p ro ach can be found in L o o s e n ’s (1972) dissertation. R etu rn in g
to the point of d e p a rtu re according to which m em orizing sentences is in fact
m em orizing re p re se n tatio n s of subjects, situations, events; it is not a big step to
assume that in the process of decoding a sentence the first things to be
m em o rize d will* be the basic structure of such situations, events, etc. M ore
peripheral details will only be ad d e d if time and m e m o ry load allow. T h e
notion of ‘basic’ or ‘essential’ traits, as o p p o sed to peripheral aspects is not well
defined in its generality, but within restricted dom ains of objects it might be
possible to give a m o re stringent definition. For instance, if the described object
is a visual p a tte rn one could define a hierarchy of traits in terms of coding
systems such as L e e u w e n b e r g ’s (1971). A m ore intuitive version of this
‘essentials first’-hypothesis is quite old in psychology. In 1894 Binet and
H enri p re se n ted d a ta from which it a p p e a re d that the most im p o rta n t parts of a
text were re p ro d u c e d b e tte r than parts of secondary im portance. H ow ever, in
206
this and in all later studies no a tte m p t was m ade to provide an in d e p e n d e n t
estimate of the im portance of various words in a sentence, or of passages in a
text.
Loosen has filled this gap. H e developed an elegant m e th o d to extract the
‘kernel id e a ’ from a sentence. H e presen ted subjects with written (D u tch)
sentences and asked them to underline one to three words that they considered
most essential for the m eaning of the sentence. Next, he d e te rm in e d over
subjects for each pair of words the relative frequency that both words had been
underlined together. T he resulting symmetric data matrix was then analysed by
J o h n s o n ’s hierarchical cluster analysis (see Levelt, 1970). (This p ro c ed u re was
justified because the matrices tu rned out to be highly ultrametric, i.e. hierarchi­
cal.) T h e result of such an analysis can be pictured as a tree diagram, an
exam ple from L o o s e n ’s dissertation is given in Figure 11.1a. It shows the
(a) importance d i a g r a m
round high interested historical with
towers charm castles visitors
(b) reproduction d ia g r a m
1
interested
high
historical
round with
towers
castles charm visitors
Figure 11.1. Importance and reproduction diagrams for H is­
torical castles with high round towers charm interested visitors
(after Loosen, 1972)
im portance hierarchy for the sentence Historical castles with high round towers
charm interested visitors (Historische kastelen m et hoge ro n d e torens b e k o re n
belangstellende bezoekers). T h e lower the node, the higher the relative
frequency that words d o m in a te d by that node were jointly underlined as most
im p o rta n t to the m eaning of the sentence. It is clear from the diagram that the
kernel idea of the sentence is contained in castles with towers charm visitors , or
even m ore strongly: castles charm visitors. O t h e r words such as function words
and modifiers are m o re peripheral to the idea expressed by the sentence.
T hese ‘kernel id e a ’ data were now co m p a red with the results of a verbatim
recall experim ent. In that ex p e rim en t a different group of subjects p erfo rm e d
w hat was essentially a continuous memory task. A long list of sentences was
acoustically p re sen ted ; the s u b je c t’s task was to listen to each sentence, to
judge (by ‘yes’ or ‘n o ’) w h e th e r it was plausible (this in o rd e r to stimulate real
u n d erstan d in g of the sentence), an<^ finally to re p ro d u c e v erbatim the test
207
sentence from m em ory. This was a high-loading task, which resulted in a
substantial a m o u n t of reproduction errors. T he data were analysed in different
ways, but the main p rocedu re was to d eterm in e for each pair of words from a
sentence the relative frequency (over subjects) with which they were jointly
reproduced. For each sentence the obtained symmetric data matrix was
analysed in the same way as the im portance data previously referred to.
T h e main finding of the ex p erim en t was that there was a striking similarity
betw een the im portance diagrams and the verbatim reproduction diagrams.
A n exam ple of the latter is given in Figure 11.1b which summarizes the
rep roduction data for the same sentence as in Figure 11.1a. T h e c o rre sp o n ­
dence b etw een the diagrams is self-evident. F u rth e rm o re , it tu rned out that, at
least for the content words, the chance of being underlined was highly
correlated with the chance of being reprod uced . For the same sentence this
relation is depicted in Figure 11.2. It may be noted in this connection also that
Teigeler (1972) found a positive relation betw een im portance and probability
of reproduction, but his im portance m easure is based on purely linguistic
considerations. His results are, m oreover, ra th er atypical, as E n g e lk a m p
(1973) rem arks.
%
80
• castles
§
• visitors
• charm
60
• historical
• with
CO
<->
•
• towers
interested
CD
^ 40
O
CD
• round
k_
1—
o
o
• high
20
importance ( N =2160)
“i
I
I
20
40
60
I
80 %
Figure 11.2. Relation between importance and rep ro ­
ducibility for the sentence Historical castles with high
round towers charm interested visitors (after Loosen,
1972)
As a conclusion to this section we can state that the different studies clearly
d e m o n s tra te the sem antic character of verbatim sentence recall. E ven if a
subject learns a sentence by h eart he does not treat it as a purely
syntactic-linguistic object, but tries to create an efficient code from which the
sentence may be reconstructed. T h e re is no evidence that the code is
208
isomorphic to the syntactic structure of the sentence, or stronger, that it is to a
substantial degree linguistic in character. This conclusion m akes it all the m ore
interesting to study the origin of the syntactic effects that are usually found in
verbatim recall experim ents. This is the subject of the following section.
The Origin of Syntactic Effects in Verbatim Reproduction of Sentences
Two types of experim ental paradigm have been used in o rd e r to d e m o n stra te
that surface constituent structure is reflected in the pattern of transitional
difficulties in sentence recall. T h e first has been m en tio n ed above: it consists of
registering transitional errors during sentence m em orization. T he o th e r type of
p roced ure consists of m easuring reaction times from presentation of a word
from the learned sentence to su b je c t’s reproduction of the next word in the
sentence (probe latencies, or probe reaction times). In both ways it is possible
to show constituent b o u n d ary effects.
A co m m o n elem en t in the different explanations which have been put
forward for this p h e n o m e n o n ( cf . Johnson 1965, 1970; Wilkes and K ennedy,
1969) is that the cohesion of words within the same constituent results from
learning during the experiment. O ne may, of course, differ in opinion ab o u t the
character of this learning. T h e sentences may have been stored in L TM in the
form of chunks that are more or less related to clauses or constituents. O r
sentences are stored in a different format, but during the learning an additional
retrieval program is constructed which consists of subroutines that are related
to different m a jo r parts of the sentence (cf. Johnson, 1970). In both cases,
however, syntax comes in during learning. In this section we will p resent data
which strongly suggest that this assum ption is wrong, and, m ore particularly,
that syntactic effects in sentence reproduction are caused by retrieval plans
which have a p re-ex p erim en tal existence, i.e. which are not created during the
ex p erim en t but which are part of o u r stock of syntactic skills.
A first indication in this direction was o b tain ed by Loosen (1972). H e
re p e a te d L ev elt’s (1 9 7 0 a) ex p erim en t with one essential change. In the
original ex perim ent subjects had been p re sen ted with sentences e m b e d d e d in
noise. T h eir only task was to write down everything they could reconstruct
from what they had heard. T h e pattern of transitional errors showed not only
strong constituent b o u n d ary effects, but from com puting conditional recall
probabilities for all word pairs in a sentence it was possible, m oreov er, to show
that the chunking pattern was highly hierarchical in nature. Levelt explained
these findings in terms of perceptual partitioning procedures. Loosen argued
that the results could as well be explained by syntactic retrieval procedures, and
to show that the actual stimulus was of relatively m inor im portance and that
therefore a perceptual explanation might be less attractive, he re p e a te d the
ex p erim en t in the following way: instead of presenting whole sentences in
noise he presented the words of each sentence in h a p h a z a rd o rd e r (and noise),
and instructed the subjects to try to re p ro d u c e the list of words as a
sentence.
209
T he results were ab o u t identical to the original results: they showed the same
hierarchical constituent b o u n d ary effects in spite of the absence of prosodic
information or o rd e r information in the stimuli. Also, no particular syntactic
frame was given or suggested during the experim ent. A p p arently, subjects are
able to apply syntactic structures of their own m aking to hap h azard word lists,
enough to show strong syntactic structuring in their reproduction. This is
consonant with the idea that syntax comes in during reproduction and is neither
a prop erty of the m e m o ry code, nor of a retrieval plan that is learned during the
experim ent. A t the same time, however, the ex p erim en t does not prove this. It
may be the case that the subject listens to the word list, constructs and stores a
sentence and finally reproduces the sentence from memory. Stronger results
are required apparently.
M ore definite conclusions can be drawn from a series of experim ents by
K em p en (1974). It is good practice in m em o ry research to unravel storage and
retrieval processes by investigating which aspects of the m em orized material
can be retrieved in the reproduction phase of the ex p erim en t by m eans of
retrieval pro ced ures that are distinct and in d e p e n d e n t from retrieval p ro c e ­
dures which have been learned during the acquisition phase. K em p en applied
this m eth o d to a n u m b e r of variations of the p robe latency technique which was
m en tio n ed above.
In each of his experim ents subjects learned a set of four D utch sentences by
heart. T h e sentences could be qualified as having weak semantic integration;
their verbs were ‘middle verb s’, i.e. they could be used transitively as well as
intransitively. E xam ples are Those two Finns wrote texts; Those three Greeks
learned laboriously. T h e critical m anipulation was the paradigm by which the
different transitions were m easured. If the usual probing paradigm was used,
i.e. presentatio n of p ro b e words in ra n d o m o r d e r — each probe word followed
by the subject quickly m entioning the next word in the sen te n c e— significant
constituent b o u n d a ry effects were obtained. T hese effects were even stronger if
the subject was instructed to react with the preceding word in the sentence
(backw ard reactions) instead of the following word. Since in these two
paradigm s the probe words could be taken from the whole sentence, they are
called the ‘sentencew ise’ paradigm s ( ‘fo rw a rd ’ and ‘b a c k w a r d ’, respectively).
T he results for these two sentencewise paradigm s are sum m arized in the u p p er
pair of d o tte d lines of Figure 11.3. It is clear that the transition from subject
noun to main verb and inversely (Finns, wrote) gives longer p robe latencies
than the transition from main verb to object noun and inversely (wrote, texts).
T he first transition corresponds to a m a jo r constituent break, the second is a
w ithin-constituent transition.
H ow ever, the profile of p ro b e latencies changes drastically if a different
paradigm is used, which K e m p e n called the ‘pairw ise’ paradigm. In this case,
subjects were (after learning) instructed that all p ro b e words would come from
one of the two positions aro u n d a p re d e te rm in e d syntactic transition, for
instance the n o u n /v e rb transition, and that if the one word is presen ted (e.g.,
the n o u n ) the subject has to re p ro d u c e the o th e r (the verb) and conversely.
210
A fter this instruction a particular transition was chosen at ra n d o m and all
backw ard and forward probes were d o n e for all learned sentences. T h e n the
e x p e rim e n te r a n n o u n ced a shift to a n o th e r transition and again all forward and
backw ard probes were done. This w ent on systematically until all transitions
had been m easured. This pairwise paradigm led to a com plete d isap p earan ce of
the constituent b o u n d a ry effect. This can be seen in Figure 11.3, w here the
continuous horizontal lines sum m arize the forward and backw ard latency data
for this paradigm. [A control ex p erim en t could successfully eliminate an
alternative explanation according to which the subjects engaged in silent
rehearsal of the word pairs from which the probes were selected during a given
series of pairwise latency m ea su rem en ts. For the details see K e m p e n (1974).]
Before interpreting this disappearance it is necessary to describe the results
of a third experim ent. In that ex p erim en t K e m p e n was able to g en erate a
constituent b o u n d a ry effect by m eans of the pairwise paradigm . This is
im po rtan t because one might ‘accuse’ the pairwise paradigm of being insensi­
tive, or at least too insensitive to m easure subtle constituent break effects. In
o rd e r to explain this third ex p erim en t a quick course in D u tch is required. T he
ex p erim en t differed from the second one only in term s of its syntactic material.
In D u tch the o rd e r of the main verb and object is different for a main clause and
a subo rd in ate clause. For Those two Finns wrote texts the word o rd e r in D utch is
the same as in English: the verb precedes the object. H o w ever, in the
subordinate clause the o rd e r inverts in the following way: Because those two
Finns ‘texts wrote\ they needed some light. It should be obvious that for D utch
ears these two orders sound equally natural if used in the correct context. T h e
third ex p erim en t differed from the second in that the subjects learned a set of
subordinate clauses w here subject nouns were always followed by object
nouns. So a typical stimulus clause was Because two Finns texts wrote (there was
no main clause added, so the stimuli were incom plete sentences). H e r e the
constituent b re ak is b etw een Finns and texts , the subject and the object which
are juxtap osed in these constructions.
It tu rn ed out that this latter transition led to relatively long p ro b e latencies,
both forward and backw ard, w hereas the w ithin-constituent transition b e ­
tween texts and wrote gave short latencies. T hese results are sum m arized in
Figure 11.3 as the b o tto m pair of d o tted lines.
In o rd e r to interpret these data, K e m p e n reaso n ed as follows. (F o r a detailed
description of the argu m ent, see K e m p e n , 1974.) First, it seems clear that the
constituent b o u n d a ry effect d e p e n d s on the retrieval task which the subjects
have to perform , not on w hat the subjects learned during acquisition. T h e latter
was nam ely identical for E x p e rim e n ts I and II; these ex perim en ts differed only
in retrieval task for which instructions were given after learning had been
com pleted. In the first ‘sen ten cew ise’ paradigm the subject could not know in
advance which transition would be p ro b ed . A t any time he could expect any
p ro b e word. In the second ‘pairw ise’ paradigm , how ever, the subject did know
in advance which transition was going to be tested. T h e second step, then, was
to consider what advantage the subject could have from this knowledge. In a
211
RT (msec)
Figure 11.3. P ro b e reactio n tim es for th re e e x ­
p e r im e n ta l c o n d itio n s (a fte r K e m p e n , 1974)
first ap p roxim ation, K e m p e n hypothesized that in the pairwise paradigm the
subject could so m eh o w limit his attention to a su b p art of the semantic
re p re se n ta tio n of the sentences, w hereas in the first paradigm the whole
internal re p re se n ta tio n had to be kept accessible and retrievable. T h e pairwise
paradigm only required the subject to consider a small (meaningful) part of the
internal re p rese n tatio n , nam ely the inform ation that Finns w rote for the first
transition, and that it is texts that were written, for the second transition. This
explains, to start with, why p ro b e latencies are smaller in the second paradigm .
But how to explain the difference in constituent b re ak effect b etw een E x p e r i­
ments II and III? F or this K e m p e n m a d e the additional assum ption th at the
subject, in retrieving such sem antic units, m akes use of particular syntactic
constructions. A syntactic construction is a string of syntactic categories,
expressing one or m ore case or modifier relationships. (In this p a p e r we will
refer to specific syntactic constructions by m ean s of labels for the expressed
relationships, e.g. S-V, etc., since the intend ed category sequence is always
clear from context.) T h a t is, the subject tries to m ap a maximally specific
syntactic construction on the inform ation to be retrieved. W h a t is the most
specific construction that the subject might use in the first ‘sen ten cew ise’
p aradigm ? Since in that case the subject does not know which transition is
going to be tested there is no o th e r recourse for him than using as a retrieval
frame the syntactic construction of the sentence as a whole. T h e subject first
expresses the sem antic inform ation in phrases co rresp o n d in g to the syntactic
frame and only then he is able to find the particular transition. T h e transition,
212
therefore, is ‘r e a d ’ from such parsed information and thus shows the co n ­
stituent b o u n d ary effect. In the pairwise paradigm, however, the subject knows
in advance what part of the information is to be retrieved and can use a much
m ore efficient syntactic construction. In E x p e rim e n t II the inform ation that
Finns wrote can be efficiently captu red in the syntactic construction S -V ; no
o th e r constructions have to be considered. A n d similarly the information that
texts were written is easily caught by using a V - O construction. T h e re is no a
priori reason to think that one of these is easier to apply than the o th e r so that no
differential effect is to be expected, which is in accordance with the data. In
o th e r words, the subject can perform his task in E x p e rim e n t II by applying a
completely overlearned syntactic construction, be it S - V or V - O . These
constructions can be different from the syntactic structure of the learned
sentences, and still be very effective for retrieval purposes. T hey are preexperim ental in the sense that they are part of the su b je c t’s syntactic skills.
Let us now turn to E x p e rim e n t III and consider why the constituent
b o u n d ary effect re ap p ears in that case. It should be r e m e m b e r e d that the main
difference betw een this and E x p e rim e n t II is in the o rd e r of subject, verb and
object, which is not S - V - O any m ore, but S - O - V ( Because Finns ‘texts wrote ’).
In this case the smallest meaningful unit related to the pair Finns, texts is the
information that Finns wrote texts. T he sequence S - O does not capture this
information, it is m o re o v e r not a syntactic construction such as S - V and V - O .
For retrieving texts, given Finns, it is therefore necessary to first use the larger
construction S - O - V for retrieving Finns texts wrote and then to read off Finns
texts. T h e retrieval of wrote given texts (or inversely) can again be easily don e by
applying the overlearned O - V construction to the inform ation that texts were
written. T h e pattern of latencies is in co rresp o n d en ce with this analysis.
If this analysis is correct, we are able to reconcile the now p o p u lar viewpoint
that sentences are m em orized in semantic or im agery-type form at with the
always recu rren t finding of syntactic effects on verbatim reproduction. This
reconciliation can be m ade, m oreover, without agreeing with the critics who
explain these effects by saying that during a verbatim recall ex p erim en t the
subject not only stores the content of the sentence, but also sets himself the
additional and completely artificial task to m em orize the syntactic frame, either
independently, or as a retrieval plan. K e m p e n ’s experim ents showed that a
syntactic construction used for retrieval is in principle in d e p e n d e n t of the
syntactic structure of the learned sentences. It d e p en d s on the experim ental
task what sort of syntactic construction is going to be used by the subject in
o rd e r to retrieve (parts of) the stored information. In a verbatim recall
ex p erim en t the subject may use the construction which he perceived in the
learned sentence, but at least for simple sentences such constructions are
overlearn ed already and it should not require m uch effort to label a particular
construction for retrieval purposes; this is quite different from learning a new
syntactic construction which would indeed be an artificial task.
As a su m m ary conclusion it can be stated that the data in this section led us to
seek the origin of syntactic effects in verbatim sentence recall in the r e p r o d u c ­
213
tion phase instead of the storage. T h ey are caused by the use of retrieval
program s that correspond to overlearn ed syntactic constructions and that are
applied to read out and verbalize the semantic information in memory.
Discussion
T h e experim ents discussed in the preceding two parag rap hs lead to the
following global description of what a subject does during verbatim sentence
learning. H e creates a semantic re p resen tatio n and in some cases an image of
the subject, event, etc. of which the sentence is a description. If u n d er m em ory
or time pressure, he tries to store the syntactically parsed string of words, but
d e p e n d in g on his expectations with respect to the recall task, he may label a
particular syntactic fram e as a retrieval program . Only in the case where
relatively complex sentences are learned and the subject is anticipating
verbatim recall will the storage of the retrieval program involve some syntactic
learning. D uring the reproduction phase, the subject will, d e p e n d e n t on the
instructions, choose from his stock of overlearned syntactic frames one which is
most specific to retrieving the information required by the task. Only these
syntactic frames can cause syntactic recall effects.
A t this point we want to m a k e three qualifications. Firstly, it seems unlikely
that any freq u en t syntactic construction can be used as a retrieval program. As
we have re m a rk e d earlier, the construction should be ap p ro p riate to be
m a p p e d on some unit of information in m em ory. O u r knowledge of the
structure of such units is still very limited but m any theories of semantic
m e m o ry (for a review see Frijda, 1972) re p rese n t sentential information in the
form of a predication over argum ents or cases. If this is correct we would expect
that certain syntactic constructions would be particularly suited for retrieval,
such as su b je c t-m a in verb, main v e rb -o b je c t, main v erb -p re p o sitio n al phrase,
as well as different types of modifier relations, e.g. v e rb -a d v e rb ,
ad je c tiv e -n o u n . T h e re may in fact be a close co rresp o n d a n ce betw een effective
retrieval program s and the syntactic constructions that figure in B e v e r’s ( 1970)
perceptual strategies. T h e re also the subject tries out syntactic frames which
have a high chance of leading him to the most im p o rta n t semantic relations.
Secondly, by p ro m o tin g syntactic constructions to retrieval program s we do
not intend to deny the existence of o th e r m eans to retrieve sentential
information from m em ory.
Thirdly, we w ant to be careful in drawing conclusions with respect to
sp o n ta n e o u s sentence production. It should be clear from the above that,
contrary to the p resen t trend in text m em o ry work, we do not consider syntactic
effects in sentence recall as peripheral, artificial, or unnatural p h e n o m e n a . On
the contrary we w ant to take them as expressions of L T M -o p e ra tio n s which are
of m uch m ore general use, especially in sp o n tan e o u s speech. Also there the
sp e a k e r tries to frame inform ation from m em o ry into syntactic construction of
his choice. A n d similarly, a syntactic construction may guide his search for
those aspects of the activated information that have to be verbalized at a
214
particular instant in speech. C are is required, however, because we do not
intend to say that in speech the syntactic p ro gram precedes the retrieval of
information from m e m o ry in all cases, nor that it is the only or most im p o rtan t
m eans of retrieval.
T o round up this discussion we finally turn to m entioning some n o n ­
continental studies which in some form or a n o th e r have also pointed to
syntactic constructions as retrieval plans.
E rv in-T rip p (1 9 6 1 ) has suggested that in o rd e r to carry out a free association
task the subject might use syntactic constructions to find a response word. If he
does, the result is a syntagmatic association. P aradigm atic associations can be
explained similarly: the stimulus word activates a syntactic fram e, the response
word can replace the stimulus word in that frame. A possible reason for the
availability of replacer sets is their useful function in speech perception: the
listener can anticipate the s p e a k e r by activating one or m o re words which
would be a likely com pletion given the syntactic fram e u n d e r construction.
Miller (1 9 6 9 ) proposes that certain asym m etries in the occu rren ce of word
associations be explained in a m a n n e r which is quite close to o u r view. O n e
typical asym m etry is the fre q u e n t association from e x e m p lar to category (e.g.
collie-dog ), w hereas the converse is rare. A n o t h e r case is the w h o le -p a rt
association ( hand-finger ) which is m o re fre q u e n t than the converse. A ccording
to Miller the subject tries to find an association word by m aking use of
predicates like 4 . . . is a . . .’, or 4 . . . has a . . . ’.
Polzella and R o h r m a n (1 9 7 0 ) found in an association e x p e rim e n t that
transitive verbs as stimuli were m ore effective in evoking n o u n reactions than
intransitive verbs. T h ey explain this by supposing that in the internal lexicon
transitive verbs have a slot for a nominal ob ject constituent. A p p a re n tly the
subject uses a little V - O construction in o rd e r to g e n e ra te an a p p ro p ria te
response. This construction is not activated in the case of intransitive verbs.
E x te n d in g this line of thought, B acharach , Kellas and M c F a rla n d (1 9 7 2 )
p e rfo rm e d a free recall e x p e rim e n t from which it a p p e a r e d that a pair of
intransitive verb and C V C trigram was easier to learn than a pair of transitive
verb and C V C trigram, but only in the case w here the trigram precedes the verb.
A p p a re n tly , trigrams sub sum e the role of subject p h ra se; in the case of a
transitive verb the ‘s e n t e n c e ’ rem ains incom plete and is th e re fo re h a r d e r to
learn. T h e difference d is a p p e a re d completely, how ever, in conditions w here
trigrams followed the verbs: there trigrams could eith er be in the role of o b jec t
phrase for transitive verbs or of ad verb for intransitive verbs, leaving the
‘s e n t e n c e ’ equally incom plete in the two cases.
W right (1 9 7 2 ) d e te rm in e d e r r o r rates for subjects answ ering questions such
as The doctor helped the nurse. B y whom was the nurse helped? T h e results led
her to suggest that in o rd e r to retrieve the answ er (e.g. the doctor) from L T M it
is necessary for the subject to use a m ediatin g sen ten ce context (e.g., the nurse
was helped by . . . ).
P ro bably closer to o u r view is a recen t p a p e r by Jam es, T h o m p s o n and
Baldwin (1973). T h e y related retrieval of a sen ten ce from m e m o ry to the
215
constructive process in norm al speech production, as we did above. M ore
specifically they select two characteristics of norm al free speech for study in a
free recall task, namely preference for active constructions over passive, and a
tendency to start a sentence with the (semantic) them e. T hey were able to
d e m o n s tra te that errors in free recall show the same biases, which is con so n an t
with o u r view that subjects may use syntactic retrieval plans which are quite
different from what they learned during acquisition. It should be added,
however, that the authors only m en tio n the reconstructive role of syntax, not its
role in m e m o ry search.
It is o u r opinion that both the re p o rte d continental studies and the
h e te ro g e n e o u s collection of results in this latter section d e m o n s tra te the
im p o rtan ce of syntactic factors in getting access to non-syntactic inform ation in
m em ory. O n e would like to see that the presently active study of sem antic
storage is c o m p le m e n te d by an equally active and systematic study of the
(partly syntactic) p ro c e d u re s em p lo y ed in the retrieval and verbal recasting of
inform ation from m em ory.
R eferences
Bacharach, V. R., Kellas, G. and McFarland, C. E. (1972). Structural properties of
transitive and intransitive verbs. Journal o f Verbal Learning and Verbal B ehavior , 11,
4 8 6 -4 9 0 .
Barclay, J. R. (1973). The role of comprehension in rem em bering sentences. Cognitive
Psychology, 4, 2 2 9 -2 5 4 .
Bever, T. G. (1970). The cognitive basis for linguistic structures. In J. R. Hayes (Ed.)
Cognition and the developm ent o f L anguage. New York: Wiley.
Binet, A. and Henri, V. (1894). La memoire des phrases (M emoire des idées). A n n é e
psychologique, 1, 24-59 .
Blumenthal, A. L. (1967). P rom pted recall of sentences. Journal o f Verbal Learning and
Verbal B ehavior, 6, 2 0 3 -2 0 6 .
Bransford, J. D. and Franks, J. J. (1972). The abstraction of linguistic ideas; a review.
Cognition , 1, 2 1 1 -2 4 9 .
Engelkam p, J. (1973). Sem antische Struktur und die Verarbeitung von Sätzen. Bern:
H uber.
Engelkam p, J., M erdian, F. and H ö rm ann , H. (1972). Semantische Faktoren beim
Behalten der V erneinung von Sätzen. Psychologische Forschung, 35, 9 3 -11 6.
Ervin-Tripp, S. M. (1961). Changes with age in the verbal determ inants of wordassociation. A m e r ic a n Journal o f Psychology, 74, 3 6 1 -3 7 2 .
Fillenbaum, S. (1966). M em ory for gist: some relevant variables. L a n g u a g e and Speech ,
9, 2 1 7 -2 2 7 .
Flores d ’Arcais, G. B. (1974). Is there a m em ory for sentences? A c ta Psychologica , 38,
3 3 -5 8 .
Frijda, N. H. (1972). Simulation of hum an long-term memory. Psychological Bulletin,
77, 1-31.
H ö rm an n , H. (1971). Semantic factors in negation. Psychologische Forschung, 3 5 ,1 - 1 6 .
James, C. T., T hom pson, J. G. and Baldwin, J. M. (1973). T he reconstructive process
in sentence memory. Journal o f Verbal Learning and Verbal Behavior, 12, 5 1 -6 3 .
Johnson, N. F. (1965). The psychological reality of phrase-structure rules. Journal o f
Verbal L earning a n d Verbal Behavior, 4, 4 6 9 -4 7 5 .
216
Johnson, N. F. (1970). The role of chunking and organization in the process of recall. In
G. H. Bower (Ed.) The Psychology o f Learning and Motivation, Vol. 4. New York:
Academic Press.
Kempen, G. (1974). Syntactic constructions as retrieval plans. Report 7 4 F U 0 2 ,
D epartm ent of Psychology, University of Nijmegen, Nijmegen.
Leeuwenberg, E. L. J. (1971). A perceptual coding language for visual and auditory
patterns. A m erican Journal o f Psychology, 84, 307-350.
Levelt, W. J. M. (1967). Sem antic features: a psychological model and its m athem atical
analysis. Report, Center for Comparative Psycholinguistics, University of Illinois,
Urbana.
Levelt, W. J. M. (1970). Hierarchical clustering algorithm in the psychology of
grammar. In G. B. Flores d ’Arcais and W. J. M. Levelt (Eds) A d v a n c e s in
Psycholinguistics. Amsterdam: North-Holland.
Levelt, W. J. M. (1970a). Hierarchical chunking in sentence processing. Perception and
Psychophysics , 8, 99-102.
Levelt, W. J. M. (1973-1974). Formal G ram m ars in Linguistics and Psycho-Linguistics.
Vol. I: Introduction to the theory of grammars and autom ata, 1973. Vol. II:
Applications in linguistic theory, 1974. Vol. Ill: Applications in psycholinguistics,
1974. The Hague: Mouton.
Levelt, W. J. M. and Bonarius, M. (1973). Suffixes as deep structure clues. Methodology
and Science, 6, 7-37.
Loosen, F. (1972). Cognitieve organisatie van zinnen in het geheugen. Dissertation,
Leuven.
Miller, G. A. (1969). The organization of lexical memory: are word associations
sufficient? In G. A. Talland and N. C. Waugh (Eds) The Pathology o f Memory. New
York: Academic Press.
Noordman, L. G. M. and Levelt, W. J. M. (1970). N oun categorization by noun-verb
intersection for the Dutch language. Heymans Bulletins HB 70-59, D ep artm ent of
Psychology, Groningen University, Groningen.
Osgood, C. E. (1970). Interpersonal verbs and interpersonal behavior. In J. L. Cowan
(Ed.) Studies in Thought and Language. Tucson: The University of Arizona Press.
Polzella, D. J. and Rohrman, N. L. (1970). Psychological aspects of transitive verbs.
Journal o f Verbal Learning and Verbal Behavior, 9, 537-540.
Rosenberg, S. (1968). Association and phrase structure in sentence recall. Journal o f
Verbal Learning a n d Verbal Behavior, 7, 1077-1081.
Sachs, J. S. (1967). Recognition memory for syntactic and semantic aspects of
connected discourse. Perception and Psychophysics , 2, 4 3 7 ^ 4 2 .
Teigeler, P. (1972). Satzstruktur und Lernverhalten. Bern: Huber.
Wilkes, A. L. and Kennedy, R. A. (1969). Relationships between pausing and retrieval
latency in sentences of varying grammatical form. Journal o f Experim ental Psych­
ology, 7 9 ,2 4 1 - 2 4 5 .
Wright, P. (1972). Some observations on how people answer questions about sentences.
Journal o f Verbal Learning and Verbal Behavior, 11, 188-195.