156 the more complicated processing of the question is

156
TRANSFORMATIONS AND THE UNDERSTANDING
OF SENTENCES
PATRICIA WRIGHT
Medical Research Council
Applied Psychology
Research Unit,
Cambridge
Five seconds after hearing a simple affirmative sentence that was either active
passive, subjects were asked a question about the sentence. The question was sometimes in the active and sometimes the passive voice. It was found that when both
sentence and question had the same voice fewer errors were made than when there was
a syntactic mismatch. Furthermore in the matched conditions there was no difference
between the active and the passive voice. This was interpreted as evidence that
subjects did not transform the sentences after hearing them, for had they done so
the passive questions might have been consistently more difficult, and perhaps also the
passive sentences. The greater difficulty observed when the voice of the question
differed from that of the sentence was attributed to subjects having to make transformations in this situation. If forcing people to make transformations results in more
errors, this also suggests that people do not normally carry out such processing as an
integral part of understanding sentences.
Analysis was carried out of the errors made on different parts of the sentence, but no
clear interpretation of this data was possible. The verb was seen to be closely related
to the grammatical subject of the sentence, which is to be expected if the sentences
were not being transformed. But fewer errors were made when the correct answer was
the agent of the verb. This might have been the effect of a specific question form.
or
INTRODUCTION
There is evidence (Mehler, 1963, Savin and Perchonock, 1965) that in order to
remember a sentence verbatim a subject transforms it into a kernel string, a form most
akin to the active voice, and separately stores the information about the syntax of the
original sentence. The present experiment is addressed to the problem of whether such
transformations are a general part of understanding sentences when the memory load
is
slight.
Consider the two voices active and passive. Theories of linguistic performance which
are based closely on transformation theory (Chomsky, 1957) would predict that the
active voice requires less processing and therefore should be easier than the passive.
That is to say active sentences and active questions will result in fewer errors than
passive sentences and passive questions. When a question in one voice is paired with
a sentence in another it can be predicted that as long as the subject has had time to
carry out the necessary transformations on the sentence, active questions about passive
sentences will be answered more easily than passive questions about active sentences.
The reason is that when the question is passive, the sentence must be remembered while
the more complicated processing of the question is undertaken, whereas when the
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
157
passive this more
question being presented.
sentence
is
difficult
processing
has been carried
out
prior
to
the
The rationale behind the present experiment is based on the idea that if subjects
have to transform sentences, and questions concerned with the sentences, in order to
understand them (as distinct from having to commit them to memory) then strong
predictions can be made about the order of difficulty to be expected when sentences and
questions of different syntactic forms are combined.
For the four combinations of active and passive sentences and questions the following
rank order of difficulty can be predicted from transformation theory:
(1) Active sentence + Active question is easier than
(2) Passive sentence + Active question, which is easier than
(3) Active sentence + Passive question, which is easier than
(4) Passive sentence + Passive question.
But, if the processing of the sentence is finished before the question is presented
then the voice of the question will be the only treatment effect observed. i.e. (1 ~- 2)
both easier than (3 = 4). These predictions are tested in the following experiment.
METHOD
Procedure
Subjects heard a series of tape recorded simple affirmative sentences numbered
consecutively 1-60. A 5 sec. unfilled time interval followed each sentence and then a
relative question was asked about the sentence, e.g. &dquo; The cat watched the bird,
What was watched by the cat ? &dquo; Subjects were required to write down their answer
as briefly as possible using just one or two words. 5 sec. later the number of the next
sentence was given and 2 sec. after this the next sentence was heard.
Sentences were either in the active (A) or passive (P) voice. Similarly half the
questions were A and half were P. Thus there were four distinct syntactic combinations
of sentence and question AA, AP, PA, PP (where the first letter of each pair indicates
the voice of the sentence, and the second gives the voice of the question). For
48 sentences the order of these four pairings was counter-balanced so that
each followed every other pairing an equal number of times and each followed itself
once. There were also twelve sentences which had &dquo; trick &dquo; questions.
These were
inserted randomly in the series of genuine items. Subjects were instructed to write
not told &dquo; as the answer to a trick question. An example of a trick question following
...
&dquo;
the sentence &dquo; The cat watched the bird &dquo; would be &dquo; What did the bird watch ? &dquo;
Each of the four syntactic pairings of sentence and question occurred three times in the
trick items.
Two groups of subjects heard the series of sentences in the reverse order to the
other two groups to ensure that there were no artifactual order effects.
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
158
Sub jects
.
.
Eighty paid
adult volunteers took part in this experiment. Their ages
they were tested in groups numbering 18-24.
from
ranged
20-60 years, and
Materials
The 48
,
.
-
genuine
items
were
made up from the 6
sentences:
the doctor
helped
the
nurse; the cat watched the
child; the fox
saw
bird; the police warned the men; the teacher thanked the
the rabbit; the boy followed the girl. Each sentence occurred once
followed by an active question and once followed by a passive question. Similarly the
passive of each of the above sentences occurred once followed by an active and once
followed by a passive question. The remaining 24 sentences were obtained by reversing
the order of the nouns in the original sentences, e.g. &dquo; The nurse helped the doctor &dquo;,
instead of &dquo; The doctor helped the nurse &dquo;. Again there were 12 active sentences, half
with active and half with passive questions, and likewise 12 passive sentences.
The purpose in reversing the nouns was two-fold. Firstly it meant that subjects
would have to listen carefully to each sentence and prevented anyone learning that, say,
the sentence about the fox was &dquo; The fox saw the rabbit &dquo;. Secondly it counterbalanced
for any order effects of the nouns within the sentence, the order of the nouns being
reversed when a sentence is changed from active to passive, and so enabled an analysis of
the separate parts of the sentence, such as the grammatical subject of the sentence in
active and passive sentences, without confounding content factors.
The 12 trick items were the original 6 active sentences and their passive transforms.
Because it was not possible to fully counterbalance the syntax and content of just 12
sentences across the 4 experimental conditions, the trick items were excluded from the
analysis. The main point of their inclusion was to prevent anyone adopting the
strategy: the correct answer is the noun in the sentence that was not mentioned in the
question.
The questions required for an answer either the subject, the verb, or the object of the
sentence. Arbitrarily it was decided that all questions about the verb should be phrased
with reference to the agent of the verb, e.g. the ‘ verb ’ question in the active voice
following both the sentence &dquo; The fox saw the rabbit &dquo; and &dquo; The rabbit was seen
by the fox &dquo; was &dquo; What did the fox do ?
&dquo;
.
’
RESULTS
Although the mean error rate
was
low, only 13.46% of all possible errors
.._
were
actually
made, nevertheless the errors were not distributed randomly across the items. Moreover,
in a task that is essentially as simple as the one being used here, perhaps it is surprising
that any errors were made at all. But it was found that people would sometimes write
down &dquo; not told &dquo; when in fact they had been given sufficient information to answer
the
question.
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
159
TABLE 1
Errors in each
position
as
percentage of total
errors
in all conditions.
If the predictions from transformation theory are correct and subjects derived from
the input sentence a base structure specifying the relationships AGENT + VERB
+ OBJECT, then the most meaningful comparison between sentence forms must be in
terms of these components of the base structure.
Table 1 shows the errors for each of these three parts of the sentence in each of the
four experimental conditions.
From Table 1 it can be seen that the distribution of errors across the experimental
conditions that obtained when quizzing the verb (passive sentences always resulted in
more errors) differed from that obtaining when the correct answer was one of the nouns
in the sentence. Since it seemed likely that this might be due to the arbitrary decision
made earlier to phrase all verb questions with reference to the agent of the verb,
verb questions were reconsidered in the light of further experimental evidence (see experiment II). Therefore the present analysis will be confined to instances where the
correct answer was either the grammatical subject or object of the sentence.
Table 2 shows the error distribution for the four experimental conditions when
combining the error scores obtained from quizzing the two nouns.
The statistical reliability of the results was determined by noting the error score
for each subject in each of the four conditions (maximum error score per subject per
condition was 8). A sign test between the appropriate pairs of scores gave the confidence
levels shown in Table 2. There it can be clearly seen that fewer errors occur when
the sentence and the question have the same syntax than when the question differs in
voice from the sentence.
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
160
TABLE 2
The
error
distribution for the four experimental conditions when the
from quizzing the two nouns are combined
error
scores
Because of this interaction between the voice of the sentence and question it seemed
preferable that further analysis should be carried out separately for those items where the
sentence and question had the same voice(matched) and where they did not (mismatched). In addition, further analysis was made with reference to the surface structure
rather than the deep structure. Accordingly Table 3 shows the total number of errors
made by subjects in each of the four conditions when the correct answer was the subject
or
the
object of the
sentence.
As above the statistical reliability of the results was assessed by determining the
error score for each subject on the 8 types of item (maximum error score per subject
per cell was 4). A sign test between the appropriate pairs of scores gave the confidence
levels shown in Table 3. Although the matched and mismatched items have been
analysed separately, clearly there is agreement regarding the overall treatment effects.
When people were asked to respond with the surface subject fewer errors were made
with active sentences, when asked about the surface object fewer errors were made with
passive sentences. Correspondingly within active sentences fewer errors are made when
quizzing the subject and within passive sentences fewer errors are made when quizzing
the object.
Two other points should also be noted about the data in Table 3.
Firstly, although the mismatch items tended to result in more errors being made, the
differences between matched and mismatched were in fact statistically reliable only when
4.61, p < 0.0001, twoquizzing the object of the sentence (for active sentences Z
3.59, p < 0.0001, two-tailed).
tailed ; for passive sentences, Z
Secondly, when the question voice matched the sentence voice, fewer errors (Z =
2.43, p < 0.015, two-tailed) were made when quizzing the object of passive sentences
=
=
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
161
~
z
5
fez
O
E
M
~<M)
a
~
z
Q
41
M
pp
V
gg
C
WZ
’3
.S
g
cr ~
Vg
M
u
o
S
g
S~
M
z
W
U
0
§~
cs
u
M
#
z
W o o W
M
u
~~
~
0
~
r
~
<
.
j
.C
41
0
.- o
>
M
m
§
f
a
~ S
CIS
~ u 4,
v
<s
u
0
g
X
a
as
§
g
H
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
162
than when quizzing the subject of active sentences, whereas when there was a mismatch between the voice of sentence and question fewer errors (Z
3.88, p < 0.001,
two-tailed) were made when quizzing the subject of passive sentences than when
quizzing the object of active sentences. No other differences within each contingency
table were reliable.
=
DISCUSSION
The above results are at variance with the simple predictions about item difficulty
that were made directly from transformation theory in the introduction. The data given
in Table 2, particularly the low error rate in the PP condition which was expected to
be the most difficult, clearly show that the match-mismatch variation has a greater
effect on error than the active-passive variable. Therefore it seems that people may have
retained the original sentence in the form that they heard it. Similarities in word ordering
may then account for the two match conditions being easier than the two mismatch
conditions. Wason (1965) has given a similar explanation of certain data on negation.
It was pointed out in the introduction that where there is a mismatch between the
voice of the sentence and question it might be expected from transformation theory that
the passive question would be a greater source of difficulty than the passive sentence.
The data of Table 2 show that PA was easier than AP, and this could be taken as
support for the transformation theory position. But the data of Table 3 raise further
problems. Consider the PA condition. If on hearing the sentence, people derived a
kernel string or base structure form that most closely approximated the active voice, then
at the point in time when the question was heard the PA situation should be comparable
to the AA condition. If Table 3 is redrawn to show the errors for the agent of the verb
and the object of the verb, then a comparison of the matched and the mismatched
conditions does suggest that the PA condition resembles the AA condition (the &dquo; mismatch &dquo; increasing the error rate by approximately half percent for both agent and
object). But this also forces the conclusion that the AP condition resembles the PP
(the &dquo; mismatch &dquo; increasing the error rate by roughly three percent for both agent and
object). This would seem nonsensical in terms of transformation theory although it may
be quite plausible on pragmatic grounds. For example, it would be interesting to know
if the full answer to the question, &dquo; What was seen by the rabbit ? &dquo; was &dquo; The fox
&dquo;
was seen by the rabbit &dquo; or
The rabbit saw the fox &dquo;. It might be thought that subjects
would spontaneously generate answers that agreed in voice with the question. Therefore
in the mismatch conditions of the present experiments perhaps on hearing the question
subjects then had to transform the sentence. The greater difhculty observed in transforming from A to P than from P to A is consistent with the results of Hepler (1967).
It is not clear how the difference in errors that obtained when quizzing the subject and
the object of the sentence can be accounted for, although the suggestion that the sentences
were transformed in the mismatched condition would overcome the difhculty of the
subject of A sentences being harder than the object of P sentences in the matched
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
163
condition but easier in the mismatched condition. With sentence transformations occurring only in the mismatched condition it is clear that fewer errors are made when the
correct answer is the object of a P sentence than when it is the subject of an A sentence,
both in the matched and mismatched conditions.
Nevertheless, the present data are not entirely consistent in their indices of the relative
difficulty of subject and object. Comparing the matched and mismatched data given in
Table 3 it is seen that the increased difficulty in the mismatched condition is apparently
confined to the surface object; i.e. when the answer is the surface subject there is no
difference in the error scores for active sentences in the matched and mismatched conditions, similarly the differences are not statistically reliable for passive sentences in
these two conditions. Whereas the differences were large and reliable (p < 0.001) when
quizzing the surface object in both sentence forms.
Further information on the relative ease of handling the agent of the verb and the
subject of the sentence could be obtained by considering those items where the correct
answer was the verb in the sentence. Because all the verb questions in the present
series were phrased with reference to the agent of the verb, Experiment II was carried
out in which the verb questions were phrased with reference to the noun acted upon.
EXPERIMENT II
procedure of the previous experiment was repeated with the single modification
concerning the phrasing of the verb questions, which were now worded with reference to
the recipient of the verbal action. The same 60 sentences as before were presented
to two groups of subjects, one hearing the list of sentences in the reversed order
The
from the other group.
’
Subjects
Another 28 subjects, again paid adult volunteers, took part in this experiment. In
order to equate the numbers in the two groups one subject, who made no mistakes on
genuine items, was excluded and another subject was dropped at random.
Materials
.
The question forms used in this and the
the sentence &dquo; The fox saw the rabbit &dquo;
previous experiment
are set out
below for
Clearly there is some difficulty in obtaining a question in the active voice that is
phrased with reference to the thing acted upon. This difficulty must be borne in mind
when interpreting the results.
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
164
TABLE 4
Errors made when
*
Data from
_
quizzing the verb, expressed
total number of errors made.
as
a
percentage
Experiment I.
’
Results
Table 4 shows the
of the total number of
made when quizzing the verb, expressed as a percentage
made in each experiment.
These percentages can be compared directly with those given in Table 1 which are
therefore included in the column headed &dquo; agent of verb &dquo;. It can be seen that rephrasing the verb questions had a marked effect on the error distribution within verb
questions. When the question was phrased with reference to the agent, passive sentences
were much more error-prone than active sentences. The reverse was true when the
question made reference to the recipient of the verbal action.
It should be noted that re-phrasing the verb questions has not affected the distribution
of errors made when quizzing the two nouns. If the errors made on the two nouns are
expressed as a percentage of all errors made in Experiment II, then the error scores for
AA and AP were 8.20 and 23.08 respectively; for PP and PA the scores were 7.69
and 18.46 respectively. These scores can be compared with the total errors on the
nouns shown in Table 1. Analysis of the errors made when quizzing the nouns in
Experiment II was carried out by combining the two matched conditions, i.e. AA and
PP, and comparing them with the two mismatched conditions, i.e. AP and PA, by
means of a sign test. This showed that the match-mismatch effect was again significant
(Z 3.49, p G 0.01, two-tailed). This confirms the finding of Experiment I that there
is an interaction between the voice of the sentence and that of the question.
errors
errors
=
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
165
DISCUSSION
The results of Experiment II suggest that the verb is perceived as being related to
the subject of the surface structure of the sentence. This close relationship between
subject and verb has also been noted by Martin, Segal and Talley (1967) using a
rating procedure. Findings such as these support the conclusion from Experiment I, that
passive sentences were not being transformed into their equivalent actives. Therefore
it seems reasonable to conclude from the present data that people do not necessarily have
to transform sentences in order to understand them. Of course these data do not bear
on the processing involved when sentences are committed to memory and should not
be taken as conflicting with the results of studies such as Mehler (1963) or Savin and
Perchonock (1965) which indicate that sentences are transformed when stored in
memory.
The notion that the passive is in fact handled as a passive, fits in well with the
suggestion that pragmatically the passive differs in meaning from the active (JohnsonLaird, 1968). Similarly Johnson (1967) showed that the surface subject of a sentence
had greater &dquo; dominance &dquo; (i.e. rated meaning) than the surface object. This should
not be taken as implying that the deep structure of a sentence is of little consequence.
There is experimental evidence to the contrary (e.g. Blumenthal and Boakes, 1967),
but Mehler and Carey (1967) compared the relative effects of surface and deep structures
and found that surface structure was the more influential. Again the point being made
is not that people cannot, nor even that they do not, transform sentences as a normal
part of processing them, but simply that they need not unless the demands of the
situation warrant it. The requirement of memorization may be one such demand, but
&dquo; understanding &dquo; as studied in the present experiments appears to make no such demand.
Nevertheless there is still no obvious explanation of why fewer errors were made when
quizzing the agent of the verbs in both active and passive sentences, in both the matched
and mismatched conditions. It is possible to interpret this finding as suggesting that for
both syntactic forms, subjects derived the base structure relationships AGENT +
VERB + OBJECT. But if one chooses to reject the null hypothesis only if differences
are significant beyond the 1’% level, then the agent is only easier with passive sentences
in the matched condition (PP) and active sentences in the mismatched condition (AP).
The most obvious thing that PP and AP have in common is the passive question. Therefore this particular aspect of the data might be due to some special effect of the
&dquo;
question form, By
whom ... ?
&dquo;
CONCLUSION
A number of other studies which have examined the
understanding of various
syntactic forms (e.g. McMahon, 1963, Mehler, 1963, Slobin, 1966, Gough, 1966) have
reported finding the passive voice more difficult than the active. But the major
differences observed in the present study were between the matched and mismatched
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016
166
conditions. This raises the question of whether studies finding the passive more difficult
have inadvertently compared active sentences in matched conditions with passive
sentences in
mismatched &dquo; conditions. The experimental technique of requiring
subjects to assess the truth or falsity of a sentence may well be prejudicial to passive
sentences for just this reason. There is no indication in the present data that passive
sentences are more difficult to understand than actives when transformations are
not required. (Combining the data from Experiments I and II for all parts of the
sentence quizzed shows that 21% of all errors were made in AA, 32% in AP,
28% in PA and 19% in PP.) It is therefore concluded that the passive voice can be
understood directly and is not inevitably more difficult than the equivalent active.
&dquo;
&dquo;
&dquo;
REFERENCES
A. L. and BOAKES, R. (1967). Prompted recall of sentences. J
. verb. Learn.
verb. Behav., 6, 674.
CHOMSKY, N. (1957). Syntactic Structures. (The Hague)
GOUGH, P. B. (1966). The verification of sentences: The effects of delay of evidence and
sentence length. J
. verb. Learn. verb. Behav., 5, 492.
HEPLER, N. (1967). Strategies for performing grammatical transformataions. Paper read to
the Eastern Psychological Association.
. verb. Learn. verb. Behav., 6,
JOHNSON, N. F. (1967). Syntactic position and rated meaning. J
240.
JOHNSON-LAIRD, P. N. (1968). The interpretation of the passive voice. Quart. J. exp. Psychol.,
20, 69.
MARTIN, R. R., SEGAL, E. M. and TALLEY, M. G. (1967). The effect of transformations, on
rated word importance. Paper read to the South Western Psychological Association.
McMAHON, L. (1963). Grammatical analysis as part of understanding a sentence. Unpublished
doctoral dissertation. Harvard University.
MEHLER, J. (1963). Some effects of grammatical transformations on the recall of English
sentences. J
. verb. Learn, verb. Behav., 2, 250.
MEHLER, J. and CAREY, P. (1967). Role of surface and base structure in the perception of
sentences. J
. verb. Learn. verb. Behav., 6, 335.
SAVIN, H. B. and PERCHONOCK, E. (1965). Grammatical structure and the immediate recall of
English sentences. J
. verb. Learn. verb. Behav., 6., 348.
SLOBIN, D. I. (1966). Grammatical transformations and sentence comprehension in childhood
and adulthood. 7. verb. Learn. verb. Behav., 5, 219.
. verb. Learn. verb. Behav., 4, 7.
WASON, P. C. (1965). The contexts of plausible denial. J
BLUMENTHAL,
Downloaded from las.sagepub.com at PENNSYLVANIA STATE UNIV on September 12, 2016