JOURNAL OF VERBAL LEARNING AND VERBALBEHAVIOR 11, 794--800 (1972)
Semantic Organization and Retrieval of Information from Sentences R I C H A R D C . ANDERSON
University of Leeds, Leeds, England I Seventy-two subjects a t t e m p t e d to recall the object n o u n s of seven once-exposed sentences given as retrieval cues the subject n o u n of each sentence a n d four other w o r d s related to the subject n o u n . T h e related cues were a r e m o t e superordinate, a close superordinate, a r e m o t e c o h y p o n y m , a n d a close c o h y p o n y m of the subject n o u n . F o u r predictions a b o u t differences in effectiveness o f these cues, derived f r o m the K a t z a n d F o d o r a n d the Quillian theories o f
semantics, were confirmed. There was one case about which the theories made different predictions. Consistent with Quillian's theory, but contrary to the Katz and Fodor theory, significantly more words were recalled when the cue was a close rather than a remote superordinate.
The purpose of the research described in this paper was to test some predictions derived from two theories of semantics. The first theory has linguistic origins. Katz and F o d o r (1963, 1964) have argued in an influential paper that meanings can be conceived as sets of minimal elements called semantic markers. The meanings of all of the words in a language are represented in a vast semantic tree. The meaning of any one word is the string of markers needed to chart a path from the top to the bottom of the tree. For example, the semantic markers for the most common reading of bachelor have been given by Katz (1966, p. 155) as: Physical Object, Living, Human, Male, Adult, and Never Married. Words with related meanings are defined with similar strings of markers. For instance, the markers for person would be: Physical Object, Living, and Human. In general, according to this theory, the string of markers for a subordinate term will include all of the markers for its 1 D u r i n g the time this research was completed, the a u t h o r held a F u l b r i g h t - H a y e s Fellowship a n d was on leave f r o m the University of Illinois. T h e a u t h o r is grateftd for the assistance of the staff o f the C o m p u t e r Based L e a r n i n g Project at the University o f Leeds, especially Patricia W o o d s . T h a n k s are due to A n d r e w O r t o n y for his critical c o m m e n t s . 794 Copyright © 1972 by Academic Press, Inc. All rights of reproduction in any form reserved.
superordinates, plus additional ones (see Katz & Fodor, 1964, p. 497). A somewhat different theory of semantic organization has been influenced by work in artificial intelligence. Quillian (1968, 1969) has theorized that the meanings of terms can be represented as interconnecting nodes in a semantic network. The first node in the unit defining a term indicates class membership or the superset. For example, immediately under canary is the indication that a canary is a bird, followed by other properties of canaries. Under bird is the indication that a bird is an animal, along with the attributes possessed by all birds; and so on. The nodes are allowed a fairly rich internal structure and different nodes are interconnected in various ways. What Quillian (1968, p. 227) calls the full meaning of a term consists not of a single node, but of all the nodes that can be reached from the node naming the to-be-understood term. In the present experiment, subjects were exposed to a series of sentences containing concrete subject nouns. Then they attempted to recall the last word of each sentence given as the retrieval cue, the subject noun or a word related to it. The related cues were remotely superordinate to, closely superordinate to,
SEMANTIC ORGANIZATIONAND RETRIEVAL remotely cohyponymous with, or closely cohyponymous with the subject nouns. For example, one of the sentences was, The bungalow stood near the river. The cues were bungalow, building, dwelling, cathedral, and cabin. Cohyponyms are words that name members of a common class (cf., Lyons, 1968). Cabin, but not cathedral, is cohyponymous with bungalow under dwelling. In other words, building names a broad class that includes cathedral; dwelling names a narrower class that excludes it. Predictions about performance on the task described above cannot be made from the Katz and Fodor theory alone. Also required are assumptions about retrieval strategies, and Katz and Fodor have not discussed the mechanisms by which the human agent performs. Several possible mechanisms will be considered. The first is based on strong assumptions. It is assumed that the markers defining a word form an ordered set; the marker indicating the most general distinction, for example, Physical Object--is at the top, and the marker, or distinguisher, indicating the most particular distinction is at the bottom. 2 While Katz and Fodor have said that it is not necessary to assume that markers comprise ordered sets, all of their exposition and examples are presented in an ordered fashion. Furthermore, Weinreich (1966) has argued persuasively that markers must form ordered sets if a theory of this type is to function properly. This is, therefore, not an unreasonable assumption. Under the assumption that markers form ordered sets, a sensible retrieval strategy would be to start with the bottom marker for the retrieval cue and search for a recently stored item coded with the same marker. If the search is successful, the person responds with the information stored in connection with 2 A distinguisher can be regarded as a marker that carries the idiosyncratic meaning of a term. According to Katz and Fodor, entries in the lexicon also include grammatical markers and selection restrictions, but these are irrelevant to the present discussion.
795
this item; in the case of the present experiment, the last word in the sentence containing that item. I f the search is unsuccessful, the person moves to the next marker in the string for the retrieval cue, so to speak collapsing the distinction embodied in the bottom marker. Again, the stored items are searched for the presence of this marker. The process continues until either the stored item is retrieved or the markers defining the retrieval cue are exhausted. It is assumed that each search operation entails some probability of error. Hence, the larger the number of markers that have to be searched the lower is the likelihood of retrieval. The following predictions from the Katz and Fodor theory of semantics can now be made. These predictions are rationalized in terms of the assumption that markers under the retrieval cue are matched against stored items one at a time in bottom-to-top order. The predictions do not change if it is assumed whole path segments are matched after discounting markers at the bottom of the string. N o r is the assumption that markers form ordered sets required if it is assumed that one or more markers under the retrieval cue are selected for matching at random. (1) The probability of recall is greater when the cue is a close cohyponym than when the cue is a remote cohyponym. The explanation is simply that a remote cohyponym entails fewer markers in common with the to-be-retrieved item than a close cohyponym. Therefore, a greater number of markers must be examined in the former case than in the latter before finding a marker in common with the to-be-retrieved item. (2) The probability of recall is greater given a close superordinate than a close cohyponym, and (3) greater given a remote superordinate than a remote cohyponym. The last marker of any superordinate is included in the string of markers for the stored item whereas every cohyponym entails at least one marker distinct from any under the stored item. (4) There is a greater difference between remote superordinates and remote co-
796
ANDERSON
hyponyms in probability of recall than between close superordinates and close cohyponyms. This is a corollary of what has already been set forth. Finally, (5) the probability of recall is the same given a close or remote superordinate. This prediction, which turns out to be the crucial one, depends least on strong retrieval assumptions. No matter whether the markers under a term are ordered or unordered, and no matter which markers are selected to match with stored items, the prediction is the same. This is true because of the fact that every marker under a superordinate, close or remote, is a marker under each subordinate of these terms. Hence, retrieval is a one step operation given any superordinate term as a retrieval cue. Consider next the predictions that follow from Quillian's theory of semantics. As already mentioned, Quillian's network theory provides for various interconnections between nodes. This permits what might be called horizontal search as well as the vertical search mediated by class membership markers. However, in the hierarchical task employed in the present experiment it seemed safe to ignore the former kinds of paths (cf., Collins & Quillian, 1969). It is assumed that when presented with a retrieval cue the subject first searches all nodes below the cue. If he is unsuccessful in finding a match with a stored item, he then moves to the node above the cue, and searches all the as-yet-unsearched nodes below that node. If still unsuccessful, the subject again moves up a node. When he finds a match, the subject responds with the information placed in memory in connection with that item, namely, in the present instance, the last word from the sentence of which the retrieved item was the subject noun. Again, it is assumed that each search operation entails some likelihood of error and that the more operations that must be completed, the lower the probability of recall. Based on these assumptions about retrieval, several predictions can be made from the Quillian theory. The first four are the same as the
ones derived from the Katz and Fodor theory. (1) The probability of recall is greater when the cue is a close cohyponym than when the cue is a remote cohyponym. Obviously, a larger number of nodes must be searched in the latter case than in the former. (2) The probability of recall is greater given a close superordinate than a close cohyponym, and (3) greater given a remote superordinate than a remote cohyponym. The shortest distance from a cohyponym to a stored item always entails a search of every node along the path of the cohyponym to the first node from which there is a path directly down to the to-beretrieved item. When starting from the superordinate, the person may immediately take the path directly to the stored item. Hence, while the maximum number of nodes that may have to be searched is the same, the average number will be smaller in the case of a superordinate than a cohyponym (provided the latter term is cohyponymous at the node marked by the superordinate). (4) There is a greater difference between remote superordinates and remote cohyponyms in probability of recall than between close superordinates and close cohyponyms, which follows from what has been explained above. We come now to the case about which the theories make different predictions. According to the Quillian theory, (5) the probability of recall is greater given a close superordinate than a remote superordinate. The minimum, maximum, and average number of nodes that will have to be searched are all smaller in the former case than in the latter. The minimum is the number of nodes separating the superordinate and the to-be-retrieved item. The maximum includes all of the nodes along wrong branches from the superordinate. The number and length of wrong branches are greater for remote than close superordinates. METHOD Subjects The subjects were 72 university students working for master's degrees in education. Four additional
797
SEMANTIC ORGANIZATION AND RETRIEVAL
beverage as well as liquid. The same consideration applied to the choice of close cohyponyms. So, duck was selected instead of, for instance, meadowlark as the cohyponym of canary. Several native English speakers
people were excluded because they failed to complete the test within the allotted time. All subjects held bachelor's degrees, in various fields. They ranged from 21 to 24 years in age. There were about equal numbers of men and women.
of English informally judged the semantic relationships among the words within sets.
TABLE 1
Procedure
SENTENCES USED IN THE EXPERIMENT
The whisky stained the carpet. The rice waved in the breeze. Rugby interested the officer. The canary escaped from the cage. The taxi appeared through the fog. The butcher opened the door. The bungalow stood near the river.
Materials Tables 1 and 2 contain the sentences and the retrieval cues that were employed in the experiment. The first retrieval cue within each set was the subject noun of one of the sentences. The remaining four words were, respectively, a remote superordinate, a close superordinate, a remote cohyponym, and a close cohyponym of the subject noun. The remote superordinates were chosen to be as distant from the subject nouns as possible without overlapping with the words from other sets. A satisfactory test of the second, third, and fourth predictions outlined earlier requires superordinates and cohyponyms in corresponding positions on a scale of remoteness-closeness. Accordingly, the remote cohyponyms were selected to be cohyponymous with the subject nouns at the level of generality represented by the remote superordinate, and not at some lower, or narrower, level. Thus, for instance, petrol was selected as the remote cohyponym for whisky instead of, say, milk, which would have been cohyponymous under
The sentences were presented by slide projector at a 7-second rate to groups of from 12 to 22 subjects. There were two random presentation orders. In order to increase the level of learning, the subjects were instructed to form a vivid mental image of the event described in each sentence. After one exposure of the sentences, subjects completed addition, subtraction, and multiplication problems for 1 minute to preclude recall from short term and therefore, perhaps, nonsemantic memory. Subjects then completed a test consisting of five retrieval cues for each sentence, the subject noun of the sentence and the four other words related to the subject noun. In every case, the response was the last word of the sentence that contained that subject noun. The cues appeared in a mimeographed booklet with one cue per page. The instructions indicated that subjects should neither skip ahead nor look back to items previously covered. To prevent systematic position or sequence effects, there were eight different random orders of retrieval cues, completed by approximately equal numbers of subjects. Six and one-half minutes were allowed for the test. RESULTS
Table 3 contains the mean proportions of words recalled given each type of cue. A fully within-subject analysis of variance of number of words recalled showed a significant effect
TABLE 2 SEVEN SETS OF RETRIEVAL CUESa
Subject noun
Remote superordinate
whisky rice rugby canary taxi butcher bungalow
liquid plants game animal vehicle person building
Close superordinate
Remote cohyponym
Close cohyponym
spirits corn teamsport bird automobile tradesman dwelling
petrol daisies chess tiger carriage aunt cathedral
gin barley hockey duck sportscar grocer cabin
a Note: Spirits, petrol, and corn correspond, respectively, in American usage to liquor, gasoline and grain. 29
798
ANDERSON TABLE
3
MEAN PROPORTION OF WORDS RECALLED AS A FUNCTION OF THE TYPE OF CUE a
Subject noun
Remote superordinate
Close superordinate
Remote cohyponym
Close cohyponym
.903
.702
.813
.438
.726
" Note: Each proportion is based on (7 words) × (72 subjects) = 504 observations. for type of cue [F(4, 2 8 4 ) = 90.97, p < .01]. Bonferroni t tests (Kirk, 1968) were employed for the planned nonorthogonal comparisons between remote and close cohyponyms, remote superordinates and remote cohyponyms, close superordinates and close cohyponyms, remote superordinates--remote cohyponyms difference and close superordinates--close cohyponyms difference, and, finally, remote and close superordinates. Each difference in recall was significant (a = .01). The types of related cue were compared to the subject noun cue using Dunnett's procedure. Again, all comparisons were significant (~ = .01). The detail of the data was consistent with the overall results. In six out of seven cases involving the important comparison between remote and close superordinates, the close superordinate proved to be the more effective retrieval cue. The one exception entailed the recall of the last word of the sentence for which the subject noun was taxi. In this case, equal recall was obtained with vehicle (the supposed remote superordinate) and automobile (the close superordinate). Looking at the data from the perspective of the individual subject, 34 of 72 recalled more words when the cue was a close superordinate, 35 recalled equal numbers of words given the two types of superordinate, and only three recalled more when the cue was a remote superordinate. The data from contrasts between other pairs of cues were similarly consistent. While position of cue in the test and type of cue were partially confounded, it was possible to make rough checks for position and sequence effects. The test was scored in four
sections according to the position of the cues. At each position the means were rank ordered as they are in Table 3 and the differences between types of cue were of roughly the same size. Compared next were the proportions of correct responses to superordinate cues depending upon whether these cues appeared in the test before or after the subject noun. There was a suggestion of a priming effect. In other words, given that a correct response had been made to the subject noun, recall was slightly higher when the superordinate cues followed the subject noun cue. There was a slighly larger difference between the remote and close superordinates when both appeared before the subject noun, but the difference when both appeared after the subject noun was still substantial (.09). It appears safe to conclude that the results were not complicated by position or sequence effects. Because of the severe constraints the design imposed upon the selection of cue words, it was not feasible to control in advance extraneous properties of the words that could have influenced performance; however, it was possible to consider some of these properties retrospectively. It is known that frequency of usage of cue words is negatively related to recall following one exposure of paired associates or sentences (Schnorr & Atkinson, 1970; Klemt & Anderson, 1972). For the present data, correlations between ThorndikeLorge frequency and recall were obtained by computing the variances and covariances separately for each type of cue (except the subject noun cue). The correlation coefficients were then calculated from pooled variances
SEMANTIC ORGANIZATION AND RETRIEVAL
and covariances. Five cues were not included in this analysis because either English usage deviated substantially from American usage or the cue was a compound of two words. There were correlations o f - . 2 5 between cue frequency of usage and mean number of words recalled, and - . 2 7 between frequency of usage and mean proportion of words recalled given correct recall to the subject noun cue. The later coefficient estimates the relationship free from the influence of characteristics of the sentences that might have affected learnability-memorability. It happens that there were mean frequency of usage differences between the various types of cues, generally biased in favor of the hypotheses set forth in the introduction. The close superordinates, for instance, were of lower mean frequency than the remote superordinates. However, following an adjustment to statistically equate frequency of usage, all of the comparisons reported above were still significant (~ = .01). Thus, the present results cannot be explained in terms of frequency of usage. Concreteness-abstractness of cues shows a strong relationship to learning and recall (Paivio, 1969). It is not u n c o m m o n to find recall two or three times as high for concrete cues than it is for abstract cues (Paivio & Foth, 1970). Obviously cohyponyms will tend to be more concrete than superordinates. I f concreteness-abstractness were the only factor at work, one would expect to find that cohyponyms were more effective retrieval cues than superordinates. The fact that, at a given level of closeness or remoteness, superordinates evoked better recall than cohyponyms indicates the overriding importance of the relationship between the cue and the to-beretrieved item. There were 223 intrusions, which is 11 percent of all overt responses and 31 percent of total errors. The remaining errors were errors of omission. Most of the intrusions could be classified in one of three ways. There were, first, 77 intralist intrusions in which the per-
799
son produced a legitimate response but paired it with the wrong cue. The most frequent error of this type entailed giving the response fog to the cuepetrol. Fog was the response term in the sentence whose subject noun was taxi. Evidently these Ss found a semantic path from petrol to taxi. Most of the remaining intralist intrusions seemed to involve a direct semantic connection between the cue and the error word (e.g., responding officer to the cue person, responding door to the cue cabin), instead of the correct but indirect connection mediated by the subject noun of a sentence. Second, there were 63 extralist intrusions in which the error word bore an obvious semantic relationship to the correct response (e.g., floor instead of carpet, wind instead of breeze, mist instead of fog). Third, there were 54 cases in which the S responded with the subject noun of a sentence rather than the object noun. Usually he had the sentence right; only he supplied the wrong word. Discounting the case of petrol and fog, generally when an S made an incorrect response to one cue within a set, h e made the same error with the other cues within this set if he responded at all. DISCUSSION Predictions from two theories in which meanings are conceived as hierarchically organized sets of features or properties were experimentally investigated. Four predictions that both theories make were confirmed. One of the results could have been expected from common sense or direct extrapolation from previous research, namely, the finding that close cohyponyms were more effective cues than remote cohyponyms. The other three are not intuitively obvious. As predicted, remote superordinates were more effective than remote cohyponyms; close superordinates were more effective than close cohyponyms; and there was a greater difference in the number of words recalled between remote superordinates and remote cohyponyms, on the one hand, than between close superordinates and close
800
ANDERSON
cohyponyms, on the other. If anything, simple extrapolation from previous research would give rise to the expectation that cohyponyms would be more effective retrieval cues than superordinates, since the former tend to be more concrete. And, there is no clear alternative way to reason from previous research about the effects of remoteness on superordinate-cohyponym differences in recall. These findings, then, must be regarded as persuasive support for the view that meanings are hierarchically arranged sets of semantic elements. There was one case about which the theories make different predictions. Quillian's theory predicts that close superordinates will be more effective cues than remote superordinates whereas the Katz and Fodor theory predicts that they will be equally effective. The data supported the Quillian theory. The conclusion is that under any term there is a marker for only the immediately superordinate class rather than markers for all superordinate classes. Predictions about recall depend upon assumptions about retrieval strategy as well as the semantic theory itself. In the case of the Katz and Fodor theory, three different retrieval assumptions were considered. These were that markers are processed one at a time in bottom-to-top order, that whole path segments are matched after discounting markers at the bottom of the string, and that one or more markers from an unordered set are sampled at random. Each of these retrieval assumptions gives rise to the same predictions about performance on the task employed in this experiment. Indeed, the author was unable to think of a retrieval strategy that would rescue the Katz and Fodor theory from a wrong prediction. With respect to the Quillian theory, different assumptions about retrieval would lead to
different predictions. However, the fact that the results matched very closely the predictions which were made from the Quillian theory may be regarded as indirect evidence for the felicity of the retrieval assumptions. REFERENCES COLLINS,A. M., & QUILLIAN,M. R. Retrieval time from semantic memory.Journalof VerbalLearning and Verbal Behavior, 1969, 8, 240-247. KATZ, J. J. The philosophy of language. New York: Harper & Row, 1966. KA:rZ,J. J., & FODOR,J. A.The structure of a semantic theory. Language, 1963, 39, 170-210. [Reprinted in J. A. Fodor & J. J. Katz (Eds.) The structure of language. Englewood Cliffs, N J: Prentice-Hall, 1964.] KIRK, R. E. Experimental design: Procedures for the behavioral sciences. Belmont, CA: Brooks/Cole, •968. KLEMT,L. L., & ANDERSON,R. C. Effects of sentence elaboration and frequency of usage on noun pair learning. Mimeographed paper. Training Research Laboratory, University of Illinois, 1972. LYONS, J. Introduction to theoretical linguistics. London: Cambridge University Press, 1968. PAwIo, A. Mental imagery in associative learning and memory.Psychological Review, 1969,76, 241-263. PAIvIo,A., & FOTH,D. Imaginal and verbal mediators and noun concreteness in paired associate learning: The elusive interaction. Journal of Verbal Learning and Verbal Behavior, 1970, 9, 384-390. QUILLIAN, M. R. Semantic memory. In M. Minsky (Ed.) Semantic informationprocessing. Cambridge, MA: MIT Press, 1968. QUILLIAN, M. R. The teachable language comprehender: A simulation program and theory of language. Communications of the ACM, 1969, 12, 459-476. SCHNORR,J., & ATKINSO~q,R. C. Study position and item differences in the short- and long-term retention of paired associates learned by imagery. Journal of Verbal Learning and Verbal Behavior,
1970, 9, 614-622. WEINREICH, U. Explorations in semantic theory. In T. A. Sebeok (Ed.) Current trends in linguistics, IlL The Hague: Mouton, 1966.
(Received May 12, 1972)