- Email: [email protected]
Distinctive feature codes in the short-term memory of children
JOURNAL
OF
Distinctive
EXPERIMENTAL
Feature
CHILD
PSYCHOLOGY
Codes
19, 241-251 (1975)
in the
Short-Term
Memory
of
Children PETER
D. EIMAS
Brown University The present studies investigated the use of distinctive feature codes in shortterm memory tasks by young, school-aged children. Each child was presented with 30 lists of consonant-vowel (CV) syllables for recall. In each list, the initial consonants were different, being drawn from the set of six English stop consonants. The data of interest were the errors in recall, which were not random as would be the case had the encoding unit been the entire consonantal phone. Rather the errors had distinctive features in common with the presented sound. Thus, children encode the consonantal phones into sets of unique features which are recalled independently at times.
Recent experimental and theoretical work in the area of memory has been directed toward an understanding of the structures and control processes associated with the reception, encoding, storage, and retrieval of information in the human adult (e.g., Atkinson & Shiffrin, 1968; Norman, 1970; Shiffrin & Atkinson, 1969). One aspect of memory which appears to be receiving increasing attention is the manner in which information is encoded for storage. Indeed, a very recent book on memory is devoted almost exclusively to the problems of coding (Melton & Martin, 1972). The encoding process also has special importance for developmental psychologists concerned with memorial processes, in that there is evidence indicating that at least some of the age-related deficiencies in memorial tasks are associated with the inability of younger children to select and use an efficient code (e.g., Haith, 1971; Morin, Hoving & Konick, 1970). The present studies were undertaken to investigate one aspect of the encoding process in children, namely, the extent to which they use a distinctive feature code in the storage of phonetic information. Such studies are relevant not only to the understanding of memorial processes in children but also to the development of the phonological component of This research was supported by Grant HD 05331 from the National Institute of Child Health and Human Development. The author wishes to thank Ms. Jane Carlstrom for her capable assistance in the collection and analysis of the data, Ms. Mary O’Brien, Principal of the Martin Luther King, Jr., School, and Mr. William Peacock, Director of the Summer Elementary School Program in Seekonk, Massachusetts, for their cooperation and support. 241 Copyright All rights
0 1975 by Academic Press, Inc. of reproduction in any form reserved.
242
PETER
D.
EIMAS
language. Children were presented with lists of consonant-vowel (CV) nonsense syllables differing only in the initial stop consonant: [b, d, g, p. t, or k]. The data of interest are the errors in recall. Wickelgren (1966). in a similar study with adults, found that the errors in recall were not random as would be expected had the most elementary unit of storage been the consonant phones, themselves. Rather he found that “Some of the features of a consonant can be recalled when others cannot, producing a systematic tendency for the errors in short-term recall to have distinctive features in common with the correct consonant” (Wickelgren, 1966, p. 397). Additional evidence for this conclusion is found in the research of Hintzman (1967), Cole, Haber, and Sales (1968), and Cole, Sales, and Haber (1969). Table 1 shows two exemplars of distinctive feature systems that are capable of describing each of the stop consonants uniquely and that are capable of making at least some differential predictions regarding the pattern of errors in short-term recall. Consider first the system based on a conventional phonetic analysis (CPA), found in most elementary linguistics textbooks. The six stop consonants can be described by two features or dimensions: the manner of production, voiced vs. voiceless, and the point of articulation or in this case the major point of closure. Thus, the stop consonant [d] is a voiced, alveolar stop. It differs from [t] only with respect to voicing, from [b] and [g] only in place of articulation, and from [p] and [k] in both voicing and place of articulation. On TABLE DISTINCTIVE
FEATURE
DESCRIPTIONS
1 OF THE SIX STOP CONSONANTS
Conventional phonetic analysis Place of articulation Bilabial
Alveolar
Velar
Voiced
bl
[dl
kl
Voiceless
LPI
[tl
[kl
Halle’s phonetic analysis Stop consonants Feature
bl
[dl
kl
[PI
it1
kl
Grave Diffuse Voicing
f + +
+ +
+ +
+ + -
+ -
+ -
DISTINCTIVE
FEATURE
CODES
243
the basis of a shared feature hypothesis, given [d] as the presented stimulus and an erroneous response, then [t] , [b] and [g] should intrude more frequently than either [p] or [k], The second system of analysis by distinctive features is one developed by Halle (1964). In this analysis each of the six stop consonants may be uniquely described by a set of three features: grave, diffuse, and voiced. As Wickelgren (1966) has noted, the values of the features are described in articulatory terms, but there is no attempt to validate the system on these grounds. Hence, it is most likely better to consider these features as abstract in that there may be complex relationships between the features and their acoustic and articulatory representations (in addition, see Wickelgren, 1969; Liberman, Cooper, Shankweiler, & Studdert-Kennedy, 1967). Using this system and the shared feature hypothesis, the predictions are that [b] and [t] will intrude more frequently than [g] or [p] which in turn will intrude more frequently than [k], given [d] as the presented sound and an error in recall. From these analyses it is possible to decide whether distinctive features provide an adequate description of the short-term encoding process of the stop consonants and, if so, which of the two feature systems has the greater predictive power. METHOD
Subjects In Expt I, 46 boys and girls served as subjects. The children attended a prefirst-grade summer school in Seekonk, MA, and were on the average 6 years of age at the time of testing. They were randomly assigned to one of two experimental groups of 23 subjects each. The subjects in Expt II were 60 boys and girls attending second grade in the Providence, RI, school system. They were approximately 7.5 years of age at the time of testing. The children were randomly assigned to one of three treatment conditions, each of which had 20 subjects. Stimulus
Materials
and Apparatus
The stimulus elements for all groups in both experiments were the syllables [ba, da, ga, pa, ta, and ka]. Each child was tested on a set of 30 lists composed of these syllables, the length of the lists in a particular set being dependent upon the age of the child and the particular treatment group to which the child had been assigned. In Expt I, the list length was either three or four syllables. One of the two groups received lists of length four, while the second group received lists of length three. In Expt II one group received lists of length five and the second group, lists of length four. In addition, a third group, which received the syllables in the visual mode, were tested on lists containing five syllables.
244
PETER
D.
EIMAS
List length was manipulated at each age level in order to determine whether the extent to which a feature code was used varied with increased demands on the short-term memory system. Although there were small differences in degree to which the distinctive feature code was used as a function of list length, the differences never reached statistically acceptable levels. As a consequence, list length and its effects upon the storage of phonetic information will not be considered further. In order to construct the 30 lists of varying length, the following procedure was employed. Thirty lists of five items each were randomly selected, with the restrictions that each of the five items be a different syllable and that each of the six syllables appear equally often (five times) in each of the five serial positions. From this master set of lists of length five, the two remaining sets of lists, one with four items per list and the other with three items per list, were constructed by merely eliminating the final item or the final two items of each of the 30 master lists. In this manner, each set of 30 lists had each of the six syllables appearing equally often: 25 times per syllable in the 5-item list, 20 times per syllable in the 4-item lists, and 15 times each in the 3-item lists. With the exception of the single group in Expt II that received the lists visually, the lists were recorded on magnetic tape in a monotone voice and played to the children on a Sony portable tape recorder. Each list was preceded by the word “ready” and the individual syllables were recorded at the rate of one syllable every 1.5 sec. In order to present the stimuli visually, the six syllables were printed on separate 3 x 5 cards in large capital letters. The rate of stimulus presentation was approximately equal to that for the auditory groups. Each stimulus was exposed for approximately .75 set, after which each card was turned over and left in front of the child, forming a row of five cards. Procedure
Each child was tested individually for two days either in a small room provided by the school or in a mobile laboratory. After instructions that informed the child that he was to play a game in which he has to try and remember in the same order what the machine had told him (or what he had seen on the cards), a short series of practice trials was given. The trials were intended to verify the child’s comprehension of the instructions and to permit the child some practice in recalling items in the order specified. The child’s spoken answers were recorded immediately by the experimenter on prearranged answer sheets. The child was encouraged to guess when he was not sure of an item, although omissions were permitted. A response was scored as correct only if it was recalled in the correct order. The instructions regarding the recall of the items were simply to begin
DISTINCTIVE
FEATURE
245
CODES
recall as soon as the machine stopped talking or as soon as all of the cards had been presented. Fifteen of the 30 lists were presented on the first day of testing and the remaining lists were presented on the second day. Upon completion of each day’s task, the children were given a small bag of M&M candies and informed that they had done very well despite the difficulty of the task. RESULTS
Experiment
I
The intrusion error data, combined for the two groups, were analyzed in a manner that permitted an unbiased assessment of the extent to which distinctive feature codes were used for encoding and recalling the six stop consonants. For purposes of describing this analysis and the manner in which the sources of potential bias were removed, the data of Expt I have been presented in a series of tables which depict the various transformations of the data. Table 2 presents the frequencies with which correct responses, intrusions (rank ordered), and omissions occurred for each of the presented sounds: the six different stops. The omissions category also includes extra-list intrusions, that is, the intrusions of syllables beginning with consonantal sounds other than the six stop consonants. Examining the first column of Table 2, we can see that when [ba] TABLE THE FREQUENCIES OMISSIONS
WITH
(INCLUDING
WHICH
CORRECT
EXTRA-LIST SIX
PRESENTED
2 RESPONSES,
INTRUSIONS) SOUNDS
INTRUSION OCCURRED
(EXPT
Lb1
Recalled correctly
bl 284
Intruded sounds
[tl
[tl
&I
[tl
117
150
154
122
[tl
[kl
[PI r :P
Omissions
kl
[PI
Dl
[kl
[dl
kl
225
I48
[PI 297
[tl 408
[kl 388
[kl 106
109
bl
Lb1
119 bl
103
101
ril
r;;
[dl 64
GP 65
if 70
[PI 42
kl 46 [dl 39
79
85
81
85
AND OF THE
I)
Presented sounds
[dl
ERRORS, FOR EACH
246
PETER
D.
EIMAS
TABLE THE
RANK
3
ORDER OF INTRUSION ERRORS FOR EACH BASED ON THE CONDITIONAL PROBABILITIES
Presented sounds
bl
Intruded sounds
[tl .2863 [PI .2771
[kl
rt1
.2116
.2309
kl .1950
kl
[dl
PRESENTED (EXPT I)
kl
[PI
[tl
[kl
[tl
[kl
.2994
.3263
[tl .2857 [kl .2787 bl .2365
kl .3453
[tl .2915
[PI .1772 bl .1711 [dl .1652 kl .I411
kl .2070
Lb1 .1856
bl .2182
LPI
[dl
.1637
.1356
kl .1077
[dl
w
.1491
.1397
[PI .0890
[dl .0913
.1606
SOUND,
LPI .2041 Lb1 .2012 IdI .0962
was presented, [ba] was correctly recalled 284 times, whereas, for example, [ta] intruded 117 times, and there were 85 omissions. Although it would be possible to estimate the use of a distinctive feature code from the data of Table 2, the results would be of questionable validity due to biasing effects arising from differences in frequency with which each of the six sounds was correctly recalled and/or omitted in recall and from differences in frequency with which each of the sounds was emitted during recall independently of its strength in memory. To adjust for the former, each of the five intrusion frequencies for a given presented sound was divided by the total number of intrusion errors for that presented sound. The resulting conditional probabilities, in rank order, are shown in Table 3. These conditional probabilities are simply the probabilities that a particular sound intruded, given the presented sound and an error in recall. To adjust for the factor of response bias, that is, the different frequencies with which the sounds were emitted during recall irrespective of their availability in short-term memory, the conditional probabilities of Table 3 were reordered. As shown in Table 4, for each intruded sound the rank orderings of presented sounds (in conditional probabilities) are listed. For example, examination of the first column in Table 4 indicates that when [b] was the intruded sound, [p] was most often the presented sound and [t] was least often the presented sound. From Table 4 it is possible to evaluate the extent to which distinctive feature analyses were used in coding and recall. Consider the first column of Table 4. According to CPA and the shared feature hypothesis, when [b] is the intruded sound, it should occur more often when [p] is presented than when [t] or [k] is presented, inasmuch as [b] differs from
DISTINCTIVE
FEATURE
TABLE THE
RANK THAT
ORDER OF PRESENTED THE INTRUDED SOUND
Intruded sounds
Lb1
Presented sounds
[PI .2365
kl
.2182
kl
.2012
4
SOUNDS, BASED ON THE CONDITIONAL OCCURRED IN PLACE OF THE PRESENTED (EXPT I)
[dl
[PI
[tl
&I
[tl
[kl
.1652
.2070
bl .2271 bl .2041 [tl .I772 [dl .1637 kl .0890
[dl .2994 [kl .2915 LPI .2857 Lb1 .2683 kl .2309
[tl .3453 kl .3263 [PI .2787 [dl .2116 [bl .1950
bl .I491
kl
.1356
Lb1 .1606
[tl
.1411
[dl
M
[dl
.0962
.1397 [PI .1077
[tl
PROBABILITIES SOUND
kl
.1856 .1711
247
CODES
LPI
.0913
[p] in voicing alone, whereas [b] differs from [t] and [k] in both voicing and place of articulation. Similar reasoning leads to the predictions that [b] should intrude more frequently when either [d] or [ gl is presented than when either [t] or [k] is presented. From this analysis it is possible to derive six binary predictions for each of the six intruded sounds. In the example discussed, five of the six predictions are confirmed. A similar logic permits the derivation of binary predictions using Halle’s system of distinctive feature analysis. Table 5 presents the percentages of correct binary predictions for both the CPA and Halle’s system and indicates the percentages of correct predictions that were statistically reliable as determined by the binomial test corrected for continuity. Chance or p was of course SO. Considering the CPA first, it is evident that this description of phonetic features plus the shared feature hypothesis predicts the pattern of intrusion errors considerably better than would be expected by chance alone. Indeed, when all of the data are considered together, 94% of the binary predictions were confirmed. The pattern of results is very similar for the Halle system. The use of all three features is typically as good as or better than the use of a single feature or a combination of two features. Inspection of Table 5 indicates that the CPA was a more powerful predictor of error patterns than was the Halle system. However, a comparison of the percentages of correct predictions for the combined data (using all features of the Halle system) did not reveal a reliable difference between the two systems k2 (1) = 2.89, p > .05). In summary, the errors in the short-term recall of lists of stop consonants by first-
248
PETER
D. EIMAS
TABLE PERCENTAGES
OF CORRECT
EXPERIMENTS
BINARY
AND
5 PREDICTIONS
DISTINCTIVE
FEATURE
Expt 1 Auditory
AS A FUNCTION
OF
SYSTEMS
Expt II Auditory
Visual
56 63 78** 60 78** 80** 80**
72** 56 64* 68* 75** 68* 73**
CPA Halle Grave Diffuse Voicing Gr & Di Gr & Vo Di & Vo All
75**
66” 64* 75** 78** 73** g-l** A
* p < .05. ** p < .Ol.
grade children are not random in nature, but rather have distinctive tures in common with the presented sound.’ Experiment
fea-
II
Table 5 presents the percentages of correct binary predictions for the combined auditory data as well as for the visual group. Once more, utilization of the CPA and the shared feature hypothesis resulted in predicted error patterns that were reliably better than would be expected by chance. The Halle system of phonetic analysis was most successml when all three features were used in deriving the predicted order of in’ It is possible that not all of the errors are a result of processes in short-term memory; that is, some of the errors may be due to perceptual processes. Wickelgren (1965, 1966) attempted to control for perceptual errors by including for his final analysis only those items that were correctly perceived immediately after presentation. The present studies did not employ a control of this nature for a number of reasons: (I) It was ourjudgment that to ask children to reproduce each item immediately after presentation would add greatly to the difficulty of the task and thereby interfere with performance; (2) There is evidence that subject produced responses interfere with short-term memory (e.g.. Waugh & Norman, 1965); (3) The perceptual error rate obtained by Wickelgren (1%5) was relatively small, 3% and 8%. Moreover, the pattern of perceptual intrusion errors is similar to that of the intrusion errors from short-term memory. Hence, the feature codes are similar for both systems; (4) To the extent that the codes for perception and memory do in fact differ, our estimate of the use of distinctive feature codes in memory is a conservative one. In toto, the small gain in precision that might have come from the use of a control for perceptual errors did not seem to outweigh the potentially disruptive effects.
DISTINCTIVE
FEATURE
CODES
249
trusion errors. The individual features of grave and diffuse and the combination of these two features were for the most part only slightly better predictors than chance in this experiment. However, it is important to note that, despite these shortcomings, Halle’s system did predict the order of intrusion errors at better than the .Ol level of significance when all three features were used. As in Expt I, the CPA was the better predictor of intrusion errors. However, the difference again failed to attain an acceptable level of statistical significance h2 (1) = 2.28, p > .05).
Of interest is the finding that even when children were presented the stimuli visually, the pattern of intrusion errors was reliably predicted by a linguistic feature analysis. Obviously the visual stimuli were coded linguistically prior to recall. Whether they were also coded in a visual manner cannot be determined from the present data. The evidence again strongly indicates that the presented sounds were encoded in a manner adequately described by a linguistic feature analysis. DISCUSSION
The evidence from the present studies demonstrates the ability of children to encode lists of stop consonants into sets of distinctive features, each set of which uniquely describes a particular stop. In attempting to recall the presented sounds, the children referred to the code of distinctive features (undoubtedly not at a conscious level) and recalled the features at least relatively independently, in that frequently only some of the features of the recalled sounds were correct. Thus the pattern of erroneous responses was in agreement with the hypothesis that the probability of an intrusion error would be a positive function of the number of distinctive features shared by the presented and intruded sounds. That the visually presented syllables were recalled in essentially the same manner as were the aurally presented syllables permits the inference that at some level of processing the visual information was transformed into a linguistic code, a finding obtained previously with adults (Sales, Haber, & Cole, 1969). It is of interest to note the linguistic level at which this transformation occurred, namely the phonetic. It is not necessary that visual information (nor for that matter acoustic information) contain syntactic and semantic information in order for linguistic analysis to proceed. Rather it is only necessary that there be sufficient information for representation in linguistic form. Although the CPA system of distinctive feature an.alysis generally resulted in a greater number of confirmed predictions, the difference was not sufficiently large to be able to state with statistical confidence that
250
PETER
D.
EIMAS
this system is the more likely phonetic code for short-term storage.2 Nor is it possible to state whether the code is best conceived of as a set of abstract features, related in complex and as yet not fully understood ways to the acoustic and articulatory levels of speech. Whatever the ultimate nature of the phonetic code, we can be reasonably certain that it is comprised of distinctive features. Moreover, the phonetic code, or at least portions of it, are available very early in life for purposes of perception and without any apparent active process of tuition. Evidence from recent studies on the perception of speech have demonstrated that infants under the age of 4 months are capable of perceiving acoustic information about voicing (Eimas, Siqueland, Jusczyk, & Vigorito, 1971) and place of articulation (Morse, 1972; Eimas, in press) in a linguistically relevant manner. With increasing maturity, the phonetic code, represented by distinctive features, becomes available for other functions such as the production of speech and short-term storage. REFERENCES Atkinson, R. C., & Shiffrin, R. M. Human memory: A proposed system and its control processes. In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation (Vol. 2). New York: Academic Press, 1968. Chomsky, N., & Halle, M. The sound pattern of English. New York: Harper & Row, 1968. Cole, R. A., Haber, R. N., & Sales. B. D. Mechanisms of aural encoding: I. Distinctive features for consonants. Perception and Psychophysics, 1968, 3, 281-284. Cole, R. A., Sales, B. D., & Haber, R. N. Mechanisms of aural encoding: II. The role of distinctive features in articulation and rehearsal. Perception and Psychophysics, 1969, 6, 343-348. Eimas, P. D. Speech perception in early infancy. In L. B. Cohen & P. Salapatek (Eds.), Infant Perception. New York: Academic Press, in press. Eimas, P. D., Siqueland, E. R., Jusczyk, P., & Vigorito, J. Speech perception in infants. Science, 1971, 171, 303-306. Haith, M. M. Developmental changes in visual information processing and short-term visual memory. Human Development, 1971, 14, 249-261. Halle, M. On the basis of phonology. In J. A. Fodor & J. J. Katz (Eds.), The structure of language. Englewood Cliffs, NJ: Prentice-Hall, 1964. 2 There are other distinctive feature systems that might have been used, for example, the systems proposed by Miller and Nicely (1955), Wickelgren (1966), and Chomsky and Halle (1968). The Miller and Nicely system reduces to the CPA features for the present set of sounds. The Wickelgren features, which are a variation of the CPA system, include the distance between items along the place of articulation feature. Use of the Wickelgren system with the present data resulted in a 3-8% reduction in correct predictions compared with the CPA system. The Chomsky and Halle features, which have one more dimension than the very similar Halle system, yield lower levels of correct predictions than the Halle feature system. The difference is about 8% in each instance. The two systems selected for detailed analysis are not only good examples of distinctive feature systems but also the better predictors of intrusion errors.
DISTINCTIVE
FEATURE
CODES
251
Hintzman,
D. L. Articulatory coding in short-term memory. Journal of Verbal Learning Behavior, 1967, 6, 312-316. Liberman, A. M., Cooper, F. S., Shankweiler, D. P., & Studdert-Kennedy, M. Perception of the speech code. Psychological Review, 1967, 74,431-461. Melton, A. W., & Martin, E. Coding processes in human memory. New York: Winston, 1972. Miller, G. A., & Nicely, P. E. An analysis of perceptual confusions among some English consonants. Journal of the Acoustical Society of America, 1955. 27, 338-352. Morin, R. E., Hoving, K. L., & Konick, D. S. Are these two stimuli from the same set? Response times of children and adults with familiar and arbitrary sets. Journal of and Verbal
Experimental
Child
Psychology,
1970, 10, 308-318.
P. A. The discrimination
of speech and nonspeech stimuli in early infancy. Journal of Experimental Child Psychology, 1972, 14, 477-492. Norman, D. Models ofhuman memory. New York: Academic Press, 1970. Sales, B. D., Haber, R. N., & Cole, R. A. Mechanisms of aural encoding: IV. Hear-see, say-write interactions for vowels. Perception and Psychophysics, 1969, 6, 385-390. Shiffrin, R. M., & Atkinson, R. C. Storage and retrieval processes in long-term memory. Psychological Review, 1969, 76, 179-193. Waugh, N. C., & Norman, D. A. Primary memory. Psychological Review, 1965, 72, 89-107. Wickelgren, W. A. Distinctive features and errors in short-term memory for English Morse,
vowels.
Journal
of the Acoustical
Society
of America,
1965. 38, 583-588.
Wickelgren, W. A. Distinctive features and errors in short-term memory for English consonants. Journal of the Acoustical Society of America, 1966, 39, 388-398. Wickelgren, W. A. Auditory or articulatory coding in verbal short-term memory. Psychological Review, 1969, 76, 232-235. RECEIVED:
February 7, 1974;
REVISED:
March 27, 1974
OF
Distinctive
EXPERIMENTAL
Feature
CHILD
PSYCHOLOGY
Codes
19, 241-251 (1975)
in the
Short-Term
Memory
of
Children PETER
D. EIMAS
Brown University The present studies investigated the use of distinctive feature codes in shortterm memory tasks by young, school-aged children. Each child was presented with 30 lists of consonant-vowel (CV) syllables for recall. In each list, the initial consonants were different, being drawn from the set of six English stop consonants. The data of interest were the errors in recall, which were not random as would be the case had the encoding unit been the entire consonantal phone. Rather the errors had distinctive features in common with the presented sound. Thus, children encode the consonantal phones into sets of unique features which are recalled independently at times.
Recent experimental and theoretical work in the area of memory has been directed toward an understanding of the structures and control processes associated with the reception, encoding, storage, and retrieval of information in the human adult (e.g., Atkinson & Shiffrin, 1968; Norman, 1970; Shiffrin & Atkinson, 1969). One aspect of memory which appears to be receiving increasing attention is the manner in which information is encoded for storage. Indeed, a very recent book on memory is devoted almost exclusively to the problems of coding (Melton & Martin, 1972). The encoding process also has special importance for developmental psychologists concerned with memorial processes, in that there is evidence indicating that at least some of the age-related deficiencies in memorial tasks are associated with the inability of younger children to select and use an efficient code (e.g., Haith, 1971; Morin, Hoving & Konick, 1970). The present studies were undertaken to investigate one aspect of the encoding process in children, namely, the extent to which they use a distinctive feature code in the storage of phonetic information. Such studies are relevant not only to the understanding of memorial processes in children but also to the development of the phonological component of This research was supported by Grant HD 05331 from the National Institute of Child Health and Human Development. The author wishes to thank Ms. Jane Carlstrom for her capable assistance in the collection and analysis of the data, Ms. Mary O’Brien, Principal of the Martin Luther King, Jr., School, and Mr. William Peacock, Director of the Summer Elementary School Program in Seekonk, Massachusetts, for their cooperation and support. 241 Copyright All rights
0 1975 by Academic Press, Inc. of reproduction in any form reserved.
242
PETER
D.
EIMAS
language. Children were presented with lists of consonant-vowel (CV) nonsense syllables differing only in the initial stop consonant: [b, d, g, p. t, or k]. The data of interest are the errors in recall. Wickelgren (1966). in a similar study with adults, found that the errors in recall were not random as would be expected had the most elementary unit of storage been the consonant phones, themselves. Rather he found that “Some of the features of a consonant can be recalled when others cannot, producing a systematic tendency for the errors in short-term recall to have distinctive features in common with the correct consonant” (Wickelgren, 1966, p. 397). Additional evidence for this conclusion is found in the research of Hintzman (1967), Cole, Haber, and Sales (1968), and Cole, Sales, and Haber (1969). Table 1 shows two exemplars of distinctive feature systems that are capable of describing each of the stop consonants uniquely and that are capable of making at least some differential predictions regarding the pattern of errors in short-term recall. Consider first the system based on a conventional phonetic analysis (CPA), found in most elementary linguistics textbooks. The six stop consonants can be described by two features or dimensions: the manner of production, voiced vs. voiceless, and the point of articulation or in this case the major point of closure. Thus, the stop consonant [d] is a voiced, alveolar stop. It differs from [t] only with respect to voicing, from [b] and [g] only in place of articulation, and from [p] and [k] in both voicing and place of articulation. On TABLE DISTINCTIVE
FEATURE
DESCRIPTIONS
1 OF THE SIX STOP CONSONANTS
Conventional phonetic analysis Place of articulation Bilabial
Alveolar
Velar
Voiced
bl
[dl
kl
Voiceless
LPI
[tl
[kl
Halle’s phonetic analysis Stop consonants Feature
bl
[dl
kl
[PI
it1
kl
Grave Diffuse Voicing
f + +
+ +
+ +
+ + -
+ -
+ -
DISTINCTIVE
FEATURE
CODES
243
the basis of a shared feature hypothesis, given [d] as the presented stimulus and an erroneous response, then [t] , [b] and [g] should intrude more frequently than either [p] or [k], The second system of analysis by distinctive features is one developed by Halle (1964). In this analysis each of the six stop consonants may be uniquely described by a set of three features: grave, diffuse, and voiced. As Wickelgren (1966) has noted, the values of the features are described in articulatory terms, but there is no attempt to validate the system on these grounds. Hence, it is most likely better to consider these features as abstract in that there may be complex relationships between the features and their acoustic and articulatory representations (in addition, see Wickelgren, 1969; Liberman, Cooper, Shankweiler, & Studdert-Kennedy, 1967). Using this system and the shared feature hypothesis, the predictions are that [b] and [t] will intrude more frequently than [g] or [p] which in turn will intrude more frequently than [k], given [d] as the presented sound and an error in recall. From these analyses it is possible to decide whether distinctive features provide an adequate description of the short-term encoding process of the stop consonants and, if so, which of the two feature systems has the greater predictive power. METHOD
Subjects In Expt I, 46 boys and girls served as subjects. The children attended a prefirst-grade summer school in Seekonk, MA, and were on the average 6 years of age at the time of testing. They were randomly assigned to one of two experimental groups of 23 subjects each. The subjects in Expt II were 60 boys and girls attending second grade in the Providence, RI, school system. They were approximately 7.5 years of age at the time of testing. The children were randomly assigned to one of three treatment conditions, each of which had 20 subjects. Stimulus
Materials
and Apparatus
The stimulus elements for all groups in both experiments were the syllables [ba, da, ga, pa, ta, and ka]. Each child was tested on a set of 30 lists composed of these syllables, the length of the lists in a particular set being dependent upon the age of the child and the particular treatment group to which the child had been assigned. In Expt I, the list length was either three or four syllables. One of the two groups received lists of length four, while the second group received lists of length three. In Expt II one group received lists of length five and the second group, lists of length four. In addition, a third group, which received the syllables in the visual mode, were tested on lists containing five syllables.
244
PETER
D.
EIMAS
List length was manipulated at each age level in order to determine whether the extent to which a feature code was used varied with increased demands on the short-term memory system. Although there were small differences in degree to which the distinctive feature code was used as a function of list length, the differences never reached statistically acceptable levels. As a consequence, list length and its effects upon the storage of phonetic information will not be considered further. In order to construct the 30 lists of varying length, the following procedure was employed. Thirty lists of five items each were randomly selected, with the restrictions that each of the five items be a different syllable and that each of the six syllables appear equally often (five times) in each of the five serial positions. From this master set of lists of length five, the two remaining sets of lists, one with four items per list and the other with three items per list, were constructed by merely eliminating the final item or the final two items of each of the 30 master lists. In this manner, each set of 30 lists had each of the six syllables appearing equally often: 25 times per syllable in the 5-item list, 20 times per syllable in the 4-item lists, and 15 times each in the 3-item lists. With the exception of the single group in Expt II that received the lists visually, the lists were recorded on magnetic tape in a monotone voice and played to the children on a Sony portable tape recorder. Each list was preceded by the word “ready” and the individual syllables were recorded at the rate of one syllable every 1.5 sec. In order to present the stimuli visually, the six syllables were printed on separate 3 x 5 cards in large capital letters. The rate of stimulus presentation was approximately equal to that for the auditory groups. Each stimulus was exposed for approximately .75 set, after which each card was turned over and left in front of the child, forming a row of five cards. Procedure
Each child was tested individually for two days either in a small room provided by the school or in a mobile laboratory. After instructions that informed the child that he was to play a game in which he has to try and remember in the same order what the machine had told him (or what he had seen on the cards), a short series of practice trials was given. The trials were intended to verify the child’s comprehension of the instructions and to permit the child some practice in recalling items in the order specified. The child’s spoken answers were recorded immediately by the experimenter on prearranged answer sheets. The child was encouraged to guess when he was not sure of an item, although omissions were permitted. A response was scored as correct only if it was recalled in the correct order. The instructions regarding the recall of the items were simply to begin
DISTINCTIVE
FEATURE
245
CODES
recall as soon as the machine stopped talking or as soon as all of the cards had been presented. Fifteen of the 30 lists were presented on the first day of testing and the remaining lists were presented on the second day. Upon completion of each day’s task, the children were given a small bag of M&M candies and informed that they had done very well despite the difficulty of the task. RESULTS
Experiment
I
The intrusion error data, combined for the two groups, were analyzed in a manner that permitted an unbiased assessment of the extent to which distinctive feature codes were used for encoding and recalling the six stop consonants. For purposes of describing this analysis and the manner in which the sources of potential bias were removed, the data of Expt I have been presented in a series of tables which depict the various transformations of the data. Table 2 presents the frequencies with which correct responses, intrusions (rank ordered), and omissions occurred for each of the presented sounds: the six different stops. The omissions category also includes extra-list intrusions, that is, the intrusions of syllables beginning with consonantal sounds other than the six stop consonants. Examining the first column of Table 2, we can see that when [ba] TABLE THE FREQUENCIES OMISSIONS
WITH
(INCLUDING
WHICH
CORRECT
EXTRA-LIST SIX
PRESENTED
2 RESPONSES,
INTRUSIONS) SOUNDS
INTRUSION OCCURRED
(EXPT
Lb1
Recalled correctly
bl 284
Intruded sounds
[tl
[tl
&I
[tl
117
150
154
122
[tl
[kl
[PI r :P
Omissions
kl
[PI
Dl
[kl
[dl
kl
225
I48
[PI 297
[tl 408
[kl 388
[kl 106
109
bl
Lb1
119 bl
103
101
ril
r;;
[dl 64
GP 65
if 70
[PI 42
kl 46 [dl 39
79
85
81
85
AND OF THE
I)
Presented sounds
[dl
ERRORS, FOR EACH
246
PETER
D.
EIMAS
TABLE THE
RANK
3
ORDER OF INTRUSION ERRORS FOR EACH BASED ON THE CONDITIONAL PROBABILITIES
Presented sounds
bl
Intruded sounds
[tl .2863 [PI .2771
[kl
rt1
.2116
.2309
kl .1950
kl
[dl
PRESENTED (EXPT I)
kl
[PI
[tl
[kl
[tl
[kl
.2994
.3263
[tl .2857 [kl .2787 bl .2365
kl .3453
[tl .2915
[PI .1772 bl .1711 [dl .1652 kl .I411
kl .2070
Lb1 .1856
bl .2182
LPI
[dl
.1637
.1356
kl .1077
[dl
w
.1491
.1397
[PI .0890
[dl .0913
.1606
SOUND,
LPI .2041 Lb1 .2012 IdI .0962
was presented, [ba] was correctly recalled 284 times, whereas, for example, [ta] intruded 117 times, and there were 85 omissions. Although it would be possible to estimate the use of a distinctive feature code from the data of Table 2, the results would be of questionable validity due to biasing effects arising from differences in frequency with which each of the six sounds was correctly recalled and/or omitted in recall and from differences in frequency with which each of the sounds was emitted during recall independently of its strength in memory. To adjust for the former, each of the five intrusion frequencies for a given presented sound was divided by the total number of intrusion errors for that presented sound. The resulting conditional probabilities, in rank order, are shown in Table 3. These conditional probabilities are simply the probabilities that a particular sound intruded, given the presented sound and an error in recall. To adjust for the factor of response bias, that is, the different frequencies with which the sounds were emitted during recall irrespective of their availability in short-term memory, the conditional probabilities of Table 3 were reordered. As shown in Table 4, for each intruded sound the rank orderings of presented sounds (in conditional probabilities) are listed. For example, examination of the first column in Table 4 indicates that when [b] was the intruded sound, [p] was most often the presented sound and [t] was least often the presented sound. From Table 4 it is possible to evaluate the extent to which distinctive feature analyses were used in coding and recall. Consider the first column of Table 4. According to CPA and the shared feature hypothesis, when [b] is the intruded sound, it should occur more often when [p] is presented than when [t] or [k] is presented, inasmuch as [b] differs from
DISTINCTIVE
FEATURE
TABLE THE
RANK THAT
ORDER OF PRESENTED THE INTRUDED SOUND
Intruded sounds
Lb1
Presented sounds
[PI .2365
kl
.2182
kl
.2012
4
SOUNDS, BASED ON THE CONDITIONAL OCCURRED IN PLACE OF THE PRESENTED (EXPT I)
[dl
[PI
[tl
&I
[tl
[kl
.1652
.2070
bl .2271 bl .2041 [tl .I772 [dl .1637 kl .0890
[dl .2994 [kl .2915 LPI .2857 Lb1 .2683 kl .2309
[tl .3453 kl .3263 [PI .2787 [dl .2116 [bl .1950
bl .I491
kl
.1356
Lb1 .1606
[tl
.1411
[dl
M
[dl
.0962
.1397 [PI .1077
[tl
PROBABILITIES SOUND
kl
.1856 .1711
247
CODES
LPI
.0913
[p] in voicing alone, whereas [b] differs from [t] and [k] in both voicing and place of articulation. Similar reasoning leads to the predictions that [b] should intrude more frequently when either [d] or [ gl is presented than when either [t] or [k] is presented. From this analysis it is possible to derive six binary predictions for each of the six intruded sounds. In the example discussed, five of the six predictions are confirmed. A similar logic permits the derivation of binary predictions using Halle’s system of distinctive feature analysis. Table 5 presents the percentages of correct binary predictions for both the CPA and Halle’s system and indicates the percentages of correct predictions that were statistically reliable as determined by the binomial test corrected for continuity. Chance or p was of course SO. Considering the CPA first, it is evident that this description of phonetic features plus the shared feature hypothesis predicts the pattern of intrusion errors considerably better than would be expected by chance alone. Indeed, when all of the data are considered together, 94% of the binary predictions were confirmed. The pattern of results is very similar for the Halle system. The use of all three features is typically as good as or better than the use of a single feature or a combination of two features. Inspection of Table 5 indicates that the CPA was a more powerful predictor of error patterns than was the Halle system. However, a comparison of the percentages of correct predictions for the combined data (using all features of the Halle system) did not reveal a reliable difference between the two systems k2 (1) = 2.89, p > .05). In summary, the errors in the short-term recall of lists of stop consonants by first-
248
PETER
D. EIMAS
TABLE PERCENTAGES
OF CORRECT
EXPERIMENTS
BINARY
AND
5 PREDICTIONS
DISTINCTIVE
FEATURE
Expt 1 Auditory
AS A FUNCTION
OF
SYSTEMS
Expt II Auditory
Visual
56 63 78** 60 78** 80** 80**
72** 56 64* 68* 75** 68* 73**
CPA Halle Grave Diffuse Voicing Gr & Di Gr & Vo Di & Vo All
75**
66” 64* 75** 78** 73** g-l** A
* p < .05. ** p < .Ol.
grade children are not random in nature, but rather have distinctive tures in common with the presented sound.’ Experiment
fea-
II
Table 5 presents the percentages of correct binary predictions for the combined auditory data as well as for the visual group. Once more, utilization of the CPA and the shared feature hypothesis resulted in predicted error patterns that were reliably better than would be expected by chance. The Halle system of phonetic analysis was most successml when all three features were used in deriving the predicted order of in’ It is possible that not all of the errors are a result of processes in short-term memory; that is, some of the errors may be due to perceptual processes. Wickelgren (1965, 1966) attempted to control for perceptual errors by including for his final analysis only those items that were correctly perceived immediately after presentation. The present studies did not employ a control of this nature for a number of reasons: (I) It was ourjudgment that to ask children to reproduce each item immediately after presentation would add greatly to the difficulty of the task and thereby interfere with performance; (2) There is evidence that subject produced responses interfere with short-term memory (e.g.. Waugh & Norman, 1965); (3) The perceptual error rate obtained by Wickelgren (1%5) was relatively small, 3% and 8%. Moreover, the pattern of perceptual intrusion errors is similar to that of the intrusion errors from short-term memory. Hence, the feature codes are similar for both systems; (4) To the extent that the codes for perception and memory do in fact differ, our estimate of the use of distinctive feature codes in memory is a conservative one. In toto, the small gain in precision that might have come from the use of a control for perceptual errors did not seem to outweigh the potentially disruptive effects.
DISTINCTIVE
FEATURE
CODES
249
trusion errors. The individual features of grave and diffuse and the combination of these two features were for the most part only slightly better predictors than chance in this experiment. However, it is important to note that, despite these shortcomings, Halle’s system did predict the order of intrusion errors at better than the .Ol level of significance when all three features were used. As in Expt I, the CPA was the better predictor of intrusion errors. However, the difference again failed to attain an acceptable level of statistical significance h2 (1) = 2.28, p > .05).
Of interest is the finding that even when children were presented the stimuli visually, the pattern of intrusion errors was reliably predicted by a linguistic feature analysis. Obviously the visual stimuli were coded linguistically prior to recall. Whether they were also coded in a visual manner cannot be determined from the present data. The evidence again strongly indicates that the presented sounds were encoded in a manner adequately described by a linguistic feature analysis. DISCUSSION
The evidence from the present studies demonstrates the ability of children to encode lists of stop consonants into sets of distinctive features, each set of which uniquely describes a particular stop. In attempting to recall the presented sounds, the children referred to the code of distinctive features (undoubtedly not at a conscious level) and recalled the features at least relatively independently, in that frequently only some of the features of the recalled sounds were correct. Thus the pattern of erroneous responses was in agreement with the hypothesis that the probability of an intrusion error would be a positive function of the number of distinctive features shared by the presented and intruded sounds. That the visually presented syllables were recalled in essentially the same manner as were the aurally presented syllables permits the inference that at some level of processing the visual information was transformed into a linguistic code, a finding obtained previously with adults (Sales, Haber, & Cole, 1969). It is of interest to note the linguistic level at which this transformation occurred, namely the phonetic. It is not necessary that visual information (nor for that matter acoustic information) contain syntactic and semantic information in order for linguistic analysis to proceed. Rather it is only necessary that there be sufficient information for representation in linguistic form. Although the CPA system of distinctive feature an.alysis generally resulted in a greater number of confirmed predictions, the difference was not sufficiently large to be able to state with statistical confidence that
250
PETER
D.
EIMAS
this system is the more likely phonetic code for short-term storage.2 Nor is it possible to state whether the code is best conceived of as a set of abstract features, related in complex and as yet not fully understood ways to the acoustic and articulatory levels of speech. Whatever the ultimate nature of the phonetic code, we can be reasonably certain that it is comprised of distinctive features. Moreover, the phonetic code, or at least portions of it, are available very early in life for purposes of perception and without any apparent active process of tuition. Evidence from recent studies on the perception of speech have demonstrated that infants under the age of 4 months are capable of perceiving acoustic information about voicing (Eimas, Siqueland, Jusczyk, & Vigorito, 1971) and place of articulation (Morse, 1972; Eimas, in press) in a linguistically relevant manner. With increasing maturity, the phonetic code, represented by distinctive features, becomes available for other functions such as the production of speech and short-term storage. REFERENCES Atkinson, R. C., & Shiffrin, R. M. Human memory: A proposed system and its control processes. In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation (Vol. 2). New York: Academic Press, 1968. Chomsky, N., & Halle, M. The sound pattern of English. New York: Harper & Row, 1968. Cole, R. A., Haber, R. N., & Sales. B. D. Mechanisms of aural encoding: I. Distinctive features for consonants. Perception and Psychophysics, 1968, 3, 281-284. Cole, R. A., Sales, B. D., & Haber, R. N. Mechanisms of aural encoding: II. The role of distinctive features in articulation and rehearsal. Perception and Psychophysics, 1969, 6, 343-348. Eimas, P. D. Speech perception in early infancy. In L. B. Cohen & P. Salapatek (Eds.), Infant Perception. New York: Academic Press, in press. Eimas, P. D., Siqueland, E. R., Jusczyk, P., & Vigorito, J. Speech perception in infants. Science, 1971, 171, 303-306. Haith, M. M. Developmental changes in visual information processing and short-term visual memory. Human Development, 1971, 14, 249-261. Halle, M. On the basis of phonology. In J. A. Fodor & J. J. Katz (Eds.), The structure of language. Englewood Cliffs, NJ: Prentice-Hall, 1964. 2 There are other distinctive feature systems that might have been used, for example, the systems proposed by Miller and Nicely (1955), Wickelgren (1966), and Chomsky and Halle (1968). The Miller and Nicely system reduces to the CPA features for the present set of sounds. The Wickelgren features, which are a variation of the CPA system, include the distance between items along the place of articulation feature. Use of the Wickelgren system with the present data resulted in a 3-8% reduction in correct predictions compared with the CPA system. The Chomsky and Halle features, which have one more dimension than the very similar Halle system, yield lower levels of correct predictions than the Halle feature system. The difference is about 8% in each instance. The two systems selected for detailed analysis are not only good examples of distinctive feature systems but also the better predictors of intrusion errors.
DISTINCTIVE
FEATURE
CODES
251
Hintzman,
D. L. Articulatory coding in short-term memory. Journal of Verbal Learning Behavior, 1967, 6, 312-316. Liberman, A. M., Cooper, F. S., Shankweiler, D. P., & Studdert-Kennedy, M. Perception of the speech code. Psychological Review, 1967, 74,431-461. Melton, A. W., & Martin, E. Coding processes in human memory. New York: Winston, 1972. Miller, G. A., & Nicely, P. E. An analysis of perceptual confusions among some English consonants. Journal of the Acoustical Society of America, 1955. 27, 338-352. Morin, R. E., Hoving, K. L., & Konick, D. S. Are these two stimuli from the same set? Response times of children and adults with familiar and arbitrary sets. Journal of and Verbal
Experimental
Child
Psychology,
1970, 10, 308-318.
P. A. The discrimination
of speech and nonspeech stimuli in early infancy. Journal of Experimental Child Psychology, 1972, 14, 477-492. Norman, D. Models ofhuman memory. New York: Academic Press, 1970. Sales, B. D., Haber, R. N., & Cole, R. A. Mechanisms of aural encoding: IV. Hear-see, say-write interactions for vowels. Perception and Psychophysics, 1969, 6, 385-390. Shiffrin, R. M., & Atkinson, R. C. Storage and retrieval processes in long-term memory. Psychological Review, 1969, 76, 179-193. Waugh, N. C., & Norman, D. A. Primary memory. Psychological Review, 1965, 72, 89-107. Wickelgren, W. A. Distinctive features and errors in short-term memory for English Morse,
vowels.
Journal
of the Acoustical
Society
of America,
1965. 38, 583-588.
Wickelgren, W. A. Distinctive features and errors in short-term memory for English consonants. Journal of the Acoustical Society of America, 1966, 39, 388-398. Wickelgren, W. A. Auditory or articulatory coding in verbal short-term memory. Psychological Review, 1969, 76, 232-235. RECEIVED:
February 7, 1974;
REVISED:
March 27, 1974