./" RC'J.riUlrtl ['81m JUl".:"OAL VUKAL LEA.~J~C A~P \"t:M"AL Ih:u,\\'IIJH. ,0..)lIl1h' 2. Xlllnl..., 1. Nu,'.'mlu" 19(..1 Cu",',i:"t @ 196.1 hy ."C";ul,onai.. rn"1" JU('o P,illlr,l in U. S. .f. JOUR:-lAL OF \'ERB,U LEAR~I~G A~D VERBAL BEHA\"lOR 2, 346-351 (1963) Some Effects of Grammatical Transformations on the Recall of English Sentences1 JACQUES )IEHLER Harvard U"iversity, Cambridge, Massachllsetts It is an all too common observation that we can understand the general significance of a verbal message even though we may be unable to repeat the exact words in which it was originally expressed. To give a precise aOCQuntof. alL.that.;gQe&. on. undet these. cir~. cumstances would, of course, be difficult if not impossible; yet it has been possible to isolate some aspects of the process and to study them by experiments on the memoriza- tion of connected discourse. The present paper represents an attempt to extend this general line of investigation by exploiting certain con- cepts that have proved useful in descriptive linguistics. The fact that a person can often rephrase "in his own words" the general sense of a message that he has not yet memorized in precise detail would seem to indicate that semantic components of a meaningful message are generally easier to recall than are its specific grammatical details. In the present experiment, therefore, we have attempted to explore this possible difference in a systematic manner. In general terms, the strategy adopted was the following: a set of eight 1 This research was supported by funds from the Public Health Service research grant No. M-05120-02 from the l'ational Institutes of Health, by the Na- tional Science Foundation grant No. NSF-G-16486, and by the Carnegie Corporation of New York, to Harvard University, Center for Cognitive Studies. I want to express my gratitude to Professor G. A. Miller for his advice on the form and content of this paper. Also I want to thank Dr. H. B. Savin for his criticism of earlier drafts and for the encourage- ment he gave me. short sentences was presented to the S, who was instructed that he would be tested for recall. The sentences differed systemati. cally in their grammatical forms; they might be either active or passive, affirmative or negative, declarative or interrogative. The S's responses were then scored both for semantic accuracy (did he recall one of the eight sentences, regardless of grammatical form?) and for syntactic accuracy (given that a sen- tence was reca1led, was its grammatical form correct?). The results of a preliminary study using this general approach have already been reported by l\Iiller (1962). The syntactic description used here is es- sentially Chomsky's (1957). According to Chomsky's grammar, most sentences are de- rh.ed from more fundamental ones by certain special rules, ca1led transformations. The fun- damental, or kernel (K), sentences are, in the vocabulary of traditional grammar, simple, acth-e, affirmative, declarative sentences, such as Tlte boy Itas hit the ball, The girl has worn tlte jewel, etc. Only three grammatical trans- formatIons are considered in this experiment: the negative (N), the passive (P), and the interrogative (Q). When the N transforma- tion is applied to The boy has hit the ball it produces a new sentence, The boy hasn't hit the ball. When P is applied to the same K sentence, it produces The ball has been hit by the boy. The Q transformation applied to the same K sentence produces Has the ball been Izit by the boy? These transformations may also be applied in combination to pro- duce, for example, a passive-negative-question . 346 - -- - -- -- ,.
Transcript
./"
RC'J.riUlrtl ['81m JUl".:"OAL VUKAL LEA.~J~C A~P \"t:M"AL Ih:u,\\'IIJH. ,0..)lIl1h' 2. Xlllnl..., 1. Nu,'.'mlu" 19(..1Cu",',i:"t @ 196.1 hy ."C";ul,onai.. rn"1" JU('o P,illlr,l in U. S. .f.
JOUR:-lAL OF \'ERB,U LEAR~I~G A~D VERBAL BEHA\"lOR 2, 346-351 (1963)
Some Effects of Grammatical Transformations on
the Recall of English Sentences1
JACQUES )IEHLER
Harvard U"iversity, Cambridge, Massachllsetts
It is an all too common observation thatwe can understand the general significance ofa verbal message even though we may beunable to repeat the exact words in which itwas originally expressed. To give a preciseaOCQuntof. alL.that.;gQe&.on. undet these. cir~.cumstances would, of course, be difficult ifnot impossible; yet it has been possible toisolate some aspects of the process and tostudy them by experiments on the memoriza-tion of connected discourse. The present paperrepresents an attempt to extend this generalline of investigation by exploiting certain con-cepts that have proved useful in descriptivelinguistics.
The fact that a person can often rephrase"in his own words" the general sense of amessage that he has not yet memorized inprecise detail would seem to indicate thatsemantic components of a meaningful messageare generally easier to recall than are itsspecific grammatical details. In the presentexperiment, therefore, we have attempted toexplore this possible difference in a systematicmanner. In general terms, the strategyadopted was the following: a set of eight
1 This research was supported by funds from thePublic Health Service research grant No. M-05120-02from the l'ational Institutes of Health, by the Na-tional Science Foundation grant No. NSF-G-16486,and by the Carnegie Corporation of New York, toHarvard University, Center for Cognitive Studies.
I want to express my gratitude to Professor G. A.Miller for his advice on the form and content ofthis paper. Also I want to thank Dr. H. B. Savin forhis criticism of earlier drafts and for the encourage-ment he gave me.
short sentences was presented to the S,who was instructed that he would be testedfor recall. The sentences differed systemati.cally in their grammatical forms; they mightbe either active or passive, affirmative ornegative, declarative or interrogative. The S'sresponses were then scored both for semanticaccuracy (did he recall one of the eightsentences, regardless of grammatical form?)and for syntactic accuracy (given that a sen-tence was reca1led, was its grammatical formcorrect?). The results of a preliminary studyusing this general approach have alreadybeen reported by l\Iiller (1962).
The syntactic description used here is es-sentially Chomsky's (1957). According toChomsky's grammar, most sentences are de-rh.ed from more fundamental ones by certainspecial rules, ca1led transformations. The fun-damental, or kernel (K), sentences are, in thevocabulary of traditional grammar, simple,acth-e, affirmative, declarative sentences, suchas Tlte boy Itas hit the ball, The girl has worntlte jewel, etc. Only three grammatical trans-formatIons are considered in this experiment:the negative (N), the passive (P), and theinterrogative (Q). When the N transforma-tion is applied to The boy has hit the ball itproduces a new sentence, The boy hasn't hitthe ball. When P is applied to the same Ksentence, it produces The ball has been hitby the boy. The Q transformation applied tothe same K sentence produces Has the ballbeen Izit by the boy? These transformationsmay also be applied in combination to pro-duce, for example, a passive-negative-question .
346
- -- - -- --
,.
347GRAMMAR AND RECALL OF SE~TE~CES
(PNQ), Hasn't tlte ball been Itit by tlteboy?
Applicationof these transformationsin allpossible combinations wiII generate seventransformed sentences for each K. The rela-tions between these eight sentences can besimply represented by a cube, as in Fig. 1,
FIG. 1. Syntactic relations of theeight sentences. Each \"ertex corre-sponds to a type of sentence, and ad-jacent vertices correspond to pairs ofsentences that differ only by one trans-formation.
where each vertex corresponds to a type ofsentence, and adjacent vertices correspond topairs of sentences that differ by only a singu-lar transformation.
METHODMaterials
The following eight kernel sentences were used:The man has bought the house; The boy has takenthe photograPh; The biologist has made the dis-covery; The girl has worn the jewel; The studenthas written the essay; The car has hit the tree; Theairplane has carried the passenger; and The secre-tary has typed the paper. The eight kernel sentences,together with their seven transformations, give 64sentences, which were dhided into eight groups ofeight sentences in such a way that in each groupno sentence and no transformation were representedmore than once. Each of the eight groups of sen-tences was the experimental material for one groupof 5s.
Procedure
Each set oi eight sentences was presented forfive successh"e trials, each time in a different order.Ten 5s at a time (chosen randomly) listened to thesentences, which were recorded on magnetic tapewith a separation of 2 sec. between the end of onesentence and the beginning of the next.
After each presentation, 5s attempted to recallthe sentences; they wrote their responses in a recallbooklet that had five pages, one for each of thefive trials. On each page were listed eight promptingwords, one for each sentence; four of these prompt-ing words were subjects and four were predicates.For example, a page in the booklet might containthe words mall, boy, biologist, girl, tree, essay, paper,airplalle, listed vertically on the page. The use ofprompting words to improve the recall of the under-lying kernel sentences is the principal methodologicaldifference between the present experiment and thepreliminary study reported by Miller (1962).
The instructions asked the 5s to complete theirrecall as quickly as possible, although no definitetime limit was specified.
There were 80 5s, all native speakers of Englishcurrently enrolled in an American unh'ersity.
RESULTS
Figure 2 shows the acquisition curves foreach of the various transformations, whereper cent recalled correctly is plotted as afunction of the trial number. The moststriking fact in Fig. 2 is the greater facilitywith which Ss. learned the kernel sentences.At each trial the difference between K and allthe other types of sentences is significant atleast at p < .005 for the first and last trialsand at p < .001 for the other three trials. Inthis experiment the number of words in asentence was not a good predictor of theease of learning; although the P sentencescontained more words, they were somewhatbetter learned than the Q sentences.
Figure 3 presents the percentages of thevarious kinds of errors that occurred on suc-cessive trials. A sentence was scored by meansof the following rules:
(a) A sentence is scored as correct if it isa verbatim reproduction of the stimulus sen-tence or if it differs from the stimulus sen-tence only by (i) the replacement of a word
------- - - -- - - -- - - -
348
100
MEHLER
LEARNING CURVES
PER SYNTACTIC FORM
o
TRIAL NUMBER
FIG. 2. Acquisition curves for each of the transformed sen-tences. The kernel is learned with greater facility than any of theother sentences.
Flc. 3. Errors in free recall of sentences. The curve labeled
Syntactic 1 gives the proportion of all sentences that are syntacticerrors. Syntactic 2 shows the same proportion of actual responsesthat are syntactic errors. Omitted gives the proportion of all sen-tences that were completely omitled. All others gives the proportionof miscellaneous errors (e.g., inventions).
by a synonym; (ii) a change from the definiteto the indefinite article; and/or (iii) a changein tense.
(b) A sentence is scored as a syntacticerror if and only if it can be derived from acorrect sentence [as defined in (a) abov~]by applying one or more of the singular trans-formations P, Q, or N or their inverses. If arecalled sentence is correct except for a syn-tactic error, then by definition its sentence-content is the same as the one in the originalsentence.
(c) A sentence is said to be omitted whenno recognizable part of it occurs in recall.
(d) All other errors are errors that do notfit any of the above categories.
Among the various types of other errorswere pure inventions, partial inventions, con-jugation of two sentences, etc. In Fig. 3 it canbe seen that omissions drop steeply betweenTrials 1 and 2 and remain at a low levelthereafter. Syntactic errors, however, increase
to a maximum around Trial 2 and then de-crease; this inflection is due, of course, to themany omissions on the first trial, as can beseen from the curve labeled SYN 2, since Sscould not make syntactic errors on the sen-tences they did not remember at all.
A matrix showing the total number of syn-tactic confusions of each type on all five trialsis given in Table 1; the rows indicate thesyntactic form of the sentence that was pre-sented, the columns indicate the syntacticform of the sentence that was recalled, andthe cell entries give the raw frequencies ofoccurrences of each stimulus-response com-bination. Entries on the main diagonal, ofcourse, indicate the number of sentences ofeach type that were correctly recalled. Thebottom row of the matrix gives the proportionof the total number of errors of each syntacticcategory.
The probability of coding correctly thesentence-content of the sentences is .85. The
---- - -- ---
I-Z
30j 1\ SYNTACTICIwu
\ OMISSIONa::
20w
\a.
0- \
101-
........\:- --0....JALL OTHER
01 I I --2 3 4
TRIAL NUMBER
350
probabilities of recalling correctly each of thetransformations are respectively P~ = .81;Pi' = .83; PQ = .86.
DISCUSSION
These results suggest a partial answer tothe question of how S can remember thegeneral sense of a sentence even when hecannot repeat it verbatim. Very roughly, thesuggested answer is that Ss do not recall thesentence verbatim, but rather that they an-alyze it syntactically and encode it as a kernelsentence plus appropriate transformation. Forexample, if the sentence is The ball Itas beenhit by tlte boy, then S presumably codes it asan underlying kernel plus some "mental tag"that indicates that the passive transformationmust be applied for recall. Exactly how thekernel is encoded, of course, is not estab-lished; it might be an image, an abstract setof symbols, or anything else capable of re-generating the kernel sentence on demand.
In support of this schema-plus-correctionhypothesis, one notes a strong tendency forSs to simplify the syntactic structure. :Kotonly were kernel sentences easier to recall, butfrom Table 1 we can count 468 syntacticerrors that (in terms of Fig. 1) involve re-sponses nearer to the kernel than was thecorrect response, but only 380 errors in theopposite direction. (If we ignore the more orless natural confusions between Q and NQand between QP and NPQ, this difference be-comes even more obvious: 400 to 248.) :More-over, the nature of the errors in K is alsosuggestive; the great majority of them weresimple omissions.
In his report of our preliminary data,Miller (1962) suggested that there are threesyntactic "footnotes" that are rememberedmore or less independently. This hypothesisof independence, however, is not critical forthe general hypothesis that S stores the in-formation after analyzing it into its separatesyntactic and semantic components. In par-ticular there is one notable exception to the
:>IEHLER
proposal that the three transformations mightbe encoded independently: Q and NQ arefrequently confused, as are PQ and N PQ.Questions, both active and passive, are morelikely than nonquestions to be recalled in thenegative. Thus more often than we wouldexpect on the assumption of independence,Q is recalled as NQ, and QP as NQP. Thisinteraction between Nand Q transformationsis undoubtedly related to the fact that affirm-ative and negative questions mean the samething (Has the boy hit the ball? vs. Ham'ttlte boy hit tlte ball?), whereas affirmative andnegative declarative sentences mean very dif-ferent things (The boy has hit the ball vs.TIte boy hasn't hit the ball).
Fortunatefy, nowever, independence of toerecall of the various transformations is notessential for a schema-plus-correction type ofhypothesis. The critical point is that the greatmajority of the errors people made consistedof sentences that could be derived from thecorrect sentence by omitting or applying syn-tactic transformations; it would not be pos-sible to account for this fact by any theoryof recall that neglected these syntactic opera-tions.
:\loreover, if one accepts the kernel sen-tence as psychologically simpler and, in somesense, closer to the way we understand themeanings of sentences, then we may also havea partial explanation for the fact, reportedin numerous studies of thinking and conceptlearning, that negative information is harderto use than affirmative information. The dif-ficulty may be due, at least in part, to thefact that negative sentences involve an addi-tional syntactic transformation. The validityof this explanation might be tested by inves-tigating the use of information expressed inpassive sentences or in questions.
SUMMARY
The recall of English sentences varyingsystematically in syntactic structure wasstudied by the method of prompted recall
GRA:'1MAR AND RECALL OF SE:STENCES
with 80 Ss. Analysis of the errors indicatedthat most of them were due to syntacticalconfusions. The hypothesis is advanced thatSs analyze the sentences into a semantic com-ponent plus syntactic correciions when theylearn them, and that this separation of se-mantic content from syntactic form is onereason that the general meaning of a message
is generally so much easier to recall than itsexact wording.
