Journal of Experimental Psychology: Human Perception and Performance 1978, Vol. 4, No. 2, 224-231

Geometric and Semantic Similarity in Visual Masking J. Zachary Jacobson and Garry Rhinelander Queen's University, Kingston, Canada Semantic and geometric or physical similarity were manipulated separately in a backward-masking situation. When the target was a word to be read aloud, formal similarity between the letters of target and mask facilitated target recognition, as did associative similarity. Masking a target word by its own anagram also facilitated whole word report. In contrast, formal similarity was inhibitory rather than facilitatory of report when the target was spelled letter-by-letter, rather than read whole. This was true even for the same target words whose whole report was facilitated by formal similarity. A model to account for this reversal in the broader context of the neural substrate of reading is advanced. It is proposed that letter and word processing are fundamentally different in that letters are recognized by hierarchical feature analysis while words are stored and recognized wholistically by diffuse and redundant networks. Implications of the results for the study of reading are discussed. Experiments involving backward masking have shown that similarity of various sorts •between target and mask affects how strongly the mask operates. It is usually reported that there is greater interference with a target by a mask when they are similar in some way. Turvey (1973) suggested that masking is basically different (and more complete, all else being equal) when letters are masked by pieces of letters than when they are masked by random dot patterns. Jacobson (1974) reported that when the task is whole-word reading, words are more strongly masked by other words or pseudowords than by random dots or pieces of letters or whole letters twisted away from upright. In a metacontrast situation (where the target and mask do not coincide spaThis research was supported in part by a grant from the National Research Council of Canada to the first author. We thank Marni McCall, Susan Gordon, and Carol Timney for their assistance in running subjects. We also thank Janet Jacobson, Susan Lederman, and D. J. K. Mewhort for comments and suggestions. Requests for reprints should be sent to J. Zachary Jacobson, Psychology Department, Queen's University, Kingston, Ontario, Canada K7L 3N6.

tially), Merikle (Note 1) found that letter and number second stimuli were more disruptive of letter and number targets than were dissimilar symbols. On another level, words masked by associates are read more easily than they are when masked by nonassociated words (Jacobson, 1973; McNicol & Howes, 1971; Samuels, 1970; Dallas & Merikle, Note 2). Because word association connotes similarity of meaning between target and mask, this runs counter to the trend of similarity increasing the strength of masking. Of course there are major differences in the levels of meaning among the various stimuli used in the above studies. Nevertheless, similarity at one level is mainly facilitatory, while at another it is mainly inhibitory. The present series of studies was done to examine the reversal of the effects of similarity in masking and to relate that to the study of reading. Experiment 1 Walley and Wieden (1973) proposed a mechanism to account for cognitive effects in visual masking. They stated, after Konor-

Copyright 1978 by the American Psychological Association, Inc. 0096-1523/78/0402-0224$00.7S

224

SIMILARITY IN MASKING ski (1967), that visual information is processed into specific patterns by various levels of feature analyzers, and that these are hierarchically organized. A to-be-perceived pattern produces activity that ultimately impinges upon one crucial cell, or gnostic unit, which accomplishes the perception. These cells are antagonistic to each other due to lateral inhibition. Drawing from reports of the cytoarchitecture of the visual cortex in other animals (Hubel & Wiesel, 1962), Walley & Weiden further suggest that gnostic units for similar percepts tend to be located closer to each other than those related to dissimilar percepts, and strength of inhibition thus increases with increasing similarity among visual patterns. Visual masking is suggested to be the result of lateral inhibition among gnostic units. Therefore the degree of cognitive masking increases with increasing similarity of targets and masks. A different model has been advanced to account for association effects in masking. Whereas Walley and Weiden (1973) suggested that single units are the crucial units of perception, Jacobson (1976) suggested that it was more reasonable to think of the perception of targets such as words as being mediated by extensive redundant networks of cortical neurons, very much like Hebb's (1949) "phase sequences." A mask causes activity in networks other than those perceiving the target. Words that are associated share many of the same elements in their memory traces; because associates share many of the same connections, an associated mask has the effect of partly restimulating the target's trace with the mask, thus reducing the mask's efficacy. Walley and Weiden's (1973) theory makes some predictions that are countervailed by other data: When target stimuli are words, mask similarity should be more inhibitory than dissimilarity. However, the results of target and mask association are opposite to this, as are others of the relations of target and mask structure when targets are words (Jacobson, 1973, 1974). Carrying the latter analysis one step further, masking words by similarly shaped

225

letters or by identical letters placed in random order (anagrams should be facilitatory). Under the gnostic cell theory of masking, that same sort of similarity should cause greater inhibition. Method Subjects. The subjects were 20 third- and fourth-year undergraduate psychology students, 9 men and 11 women. Stimuli. Twelve words ranging in length from four to seven letters served as the target stimuli. Three different classes of letter masks were constructed, and there was a mask from each class for each target used. Each masking stimulus contained the same number of letters as the target word for which it was used. One class of masks were geometrically similar in that they contained contour features similar to the letters contained in the target words; each letter of the mask was angular or curved as determined by the curvature or angularity of the particular letter in the target it coincided with. A second set of masks was made of randomly chosen letters not contained in the target stimuli. The remaining masks were anagrams of the target word, that is, the letters that made up the target word were randomly rearranged to form the mask. The same letters never occupied identical locations in both the target and mask. The stimuli used are shown in Table 1. Apparatus. A Gerbrands two-field projection tachistoscope was used to present the stimuli. When projected, the words were made of bright capital letters (27 c/m2) with a height of 2.5 cm and a width of 2 cm on a dark background (7 c/m 2 ) shown on a screen 3 m in front of the subject. Thus a single letter subtended slightly less than .5° of visual arc. There was no fixation point, but subjects were instructed to look at that region in the center of the screen where the stimuli appeared; that was not difficult, since the room was dimly lit (producing the 7 c/m2 background on the projection screen). Procedure. The order of stimuli was randomized separately for each subject. In each presentation, the target word was shown for 8 msec, and the mask was subsequently presented in the same place for 120 msec. A word pair was first shown with a lag (stimulus onset asynchrony; SO A) of 10 msec, the next time at 20 msec, and so on. Target and mask were changed at each exposure and replaced with an as yet unidentified word pair with the SOA appropriate for that word pair at that point in the process. This was done in order to preclude spillover priming from one exposure to the next. Had that not been done, it might have been possible to explain any effect of similarity of form as being due to priming or repeti-

226

ZACHARY JACOBSON AND GARRY RHINELANDER

clearly in the direction opposite to physical similarity or letter identity producing greater masking.

Table 1

Targets and Masks in Experiment 2 Masks

Experiment 2

Target words

Geometrically similar

Randomly selected

Anagram

ADVANCE BEGIN DRAG LARGE HARD CROWD NOISE THROW STUDENT WRONG BREAK FLAT

MBWMHFG SPCTH BPMC DMPCF NMPB GPQVB HQTBF INPQV BIHBFHI VPQHC SPFMB EDMI

JZXIJGB AOFLU TZYV FKVDY GFQW IMSXZ PTRMP PNEMQ DZOVOLQ THZVF FNLTN URPA

NVACEDA NIBEG AGDR AEGLR DRHA WODRC ENSIO RWOHT NDTEUTS GNWRO EBKRA LATF

tion clarity (Dodwell, 1971; Haber, 1965) or guessing of letters similar to the previous mask. The subject's task was to read the target word aloud when he could. Exposures continued until all targets were read. The dependent variable was the mask's lag, or stimulus onset asynchrony, when the target word was identified. Any particular target could appear in any mask condition. However, each target was only given to each subject in one condition, and that selection was made randomly for each subject. Each subject received four stimulus pairs from each masking condition.

Results and Discussion The mean stimulus onset asynchrony when the target stimulus was read correctly was 39.5 msec for the geometrically similar mask trials, 44.3 msec for the randomly selected mask condition, and 38.7 msec for the anagram masks. These results are shown in Figure 1. There was no reliable difference between words masked by anagrams or by similar letters, but random letters were significantly more disruptive than the two other sorts of mask (randomization test, p = .01). The only effect of changing the stimuli on each trial was to reduce critical SOAs in all conditions by about 20 msec as compared with earlier studies we have done in which target and mask were not changed at each exposure. The differences between mask conditions —both with and without continuous replacement—are small, but they are reliable and

Jacobson (1973) reported that association is facilitatory of word reading in masking, but in that situation it might again be argued-—as above—that the facilitation is due to a carryover of information from one exposure of the mask to the next of the target. If that were the case, the facilitation effect would not take place in backward masking as such; rather it too might be a case of priming or repetition clarity or guessing. Therefore, because we wished to study similarity in masking only, the second study was a replication of Jacobson's (1973) study on association, with the difference that target and mask were changed at each exposure. This was done as in Experiment 1 and for the same reason, namely, to prevent carryover facilitation. Critical stimulus onset asynchronies of tachistoscopically presented words were noted when they were backward-masked by words that either were or were not associates of the target. Method Twenty undergraduate psychology students (10 men and 10 women) who had volunteered served as subjects. 50r-

40

30 ANAGRAM

GEOMETRICALLY SIMILAR LETTERS SORT

RANDOM

OF MASK

Figure 1. Mean critical stimulus onset asynchrony (SOA, in msec) when targets were read as a function of whether masks were anagrams, similar letters, or random letters (Experiment 1).

SIMILARITY IN MASKING The same word pairs were used as in Jacobson's first study (1973, Table 1). Five associated word pairs (as defined by Arthur's list of free associates, Note 3) and five unassociated word pairs were chosen for each subject from the 32 pairs available. The same tachistoscope, luminances, and procedures were used as in Experiment 1; the subject's task was to read the target when he could, and the dependent variable was the mask's lag when that happened. Also as in the previous study, targets and masks were changed at each exposure to prevent spillover priming.

227

Subjects were asked to spell real words, redundant (fourth-order) pseudowords, and nonredundant (zero-order) pseudowords when they were masked by upright letters, letters twisted away from upright, and pieces of letters. If similarity causes greater masking, then those masks should be respectively less disruptive when subjects are asked to spell the targets. This is because of the decreasing similarity between these masks and the targets. In fact, that is what happened.

Results and Discussion Associated pairs were identified at a mean critical SOA of 40.6 msec as opposed to a mean of 47.5 msec for unassociated pairs (randomization test, one-tailed /> = .001). No other trend was observed reliably in the data. Previous performance of subjects in this situation without stimulus replacement on each exposure is about 40 msec on associated word pairs and 70-80 msec for unassociated controls. If stimulus replacement had any effect in comparison with previously reported results, it was to ease the control trials, rather than to make more difficult the associated trials, as a priming or response-bias interpretation of the association effect would have predicted.1 In light of these previous results and those of Experiment 2, then, association facilitation seems to be a perceptual effect in backward masking. Experiment 3 When targets are read as words, similarity of the shapes of target and mask letters aids recognition, as does association. These results notwithstanding, the notion that similarity increases masking may not concern anything so advanced as reading a whole word. Perhaps it is "unfair" to expect notions involving formal similarity in particular to make predictions as far as word reading, wherein the very event of reading might alter the result. For this reason a study was done on the effect of mask similarity, wherein the subject's task was to spell word and pseudoword targets rather than to read them as complete words.

Method Twenty undergraduates at Queen's University were tested. The same apparatus was used as in the two previous studies. Target stimuli were four-letter uppercase words and zero- and fourthorder pseudowords. Masks were four upright letters, four letters twisted away from upright, and four "chimeric" letters (half of one letter, half of another), which were also randomly twisted from upright. Mask letters were all randomly chosen; in this way any difference among masks would necessarily be ascribed to the differences in mask efficiency among letter fragments and skewed and upright letters. Subjects were asked to spell the targets letter-by-letter from left to right. Stimulus onset asynchronies were recorded when all of the target letters had been correctly identified, but unlike the previous studies, stimuli were not changed on each trial—a pair was shown repeatedly, with increasing SOA, until it was spelled correctly. Each subject was tested three times in each of the nine conditions.

Results and Discussion The results are shown in Figure 2. Unlike the previous studies, mask similarity produced increased masking, F(2, 38) = 24.1, p < .0001. There was also a target effect such that words were easier than redundant 1 Another possible explanation of the result is that the presence of any mask word triggers associate responding willy-nilly, thus interfering with report in unassociated trials. If this were the case, words masked by unredundant pseudowords, which produce fewer associations than a word, ought to be read about as easily as those masked by associates. This was tested in a between-subjects study, and pseudowords were almost exactly as deleterious masks as unrelated words were, and both were more deleterious than associates (Jacobson, 1973).

228

ZACHARY JACOBSON AND GARRY RHINELANDER NON-REDUNDANT PSEUDOWORDS

REDUNDANT PSEUDOWORDS

.O REAL WORDS

o 100

z

LETTER PIECES

LETTERS TWISTED

UPRIGHT LETTERS

MASKS

Figure 2. Mean critical stimulus onset asynchrony (SOA) for spelled targets as a function of target and mask content (Experiment 3).

pseudowords and redundant pseudowords were easier than nonredundant pseudowords, F(2, 38) = 123.47, p < .00001. Subjects were faster on the third trial in any condition (142 msec mean critical SOA) than on the first trial (163 msec mean critical SOA), F(2, 38) = 6.79, p < .004. In the third study, then, the effect of mask similarity reversed and became inhibitory. This is the more usual report, and it is what feature-extractive notions of target identification predict. Such pyramidal networks have been hypothesized as the basis of perception (e.g., Uttley, 1958, as well as Walley & Wieden, 1973, and Konorski, 1967), and the primitive first few steps of these seem to exist in other mammals. However, the present studies taken together indicate that such networks do not extend throughout memory, or even to the perception or storage of whole words, and that words are recognized as wholes, at least in adult readers. (Such wholistic recognition was seen to occur in randomly connected perceptions; Block, 1962; Block, Knight, & Rosenblatt, 1962.) Experiment 4 Experiment 3 found progressively greater inhibition due to increasing similarity of

mask to target when the subject's task was to spell the target. However, the physical similarity of individual letters was not manipulated between targets and words. If indeed there are two sorts of recognition operating—feature analysis for letters (where similarity is inhibitory) and direct recognition of whole words (where similarity is facilitatory)—then geometric similarity among letters ought also to be inhibitory when words are spelled rather than read. A study was done using the same target stimuli as in Experiment 1, masked by geometrically similar, random, and deliberately dissimilar letter stimuli. The main difference between this and the first study was that subjects were asked to spell the targets rather than to read them. This study was designed on the assumption that if the effect of similarity were to reverse for the same targets spelled as were read in Experiment 1, then this would provide support for the dual processes of recognition suggested above. Deliberately dissimilar letters were now included as a further test; obviously these should produce even less masking than randomly chosen letters. Method Eighteen Queen's University undergraduates served as subjects. Apparatus and procedure were as described for the previous studies. Target words were masked by similarly shaped letters, randomly chosen letters, and dissimilarly shaped letters. Like Experiment 3, stimuli were not changed at each exposure; rather, they were continued until the target was reported correctly. (Any carryover effect would have worked to facilitate targets masked by similarly shaped letters.) Each subject received 30 trials, 10 each of words masked by similar letters, words masked by dissimilar letters, and words masked by randomly chosen letters. Words were different in each condition for subjects, but targets were counterbalanced across subjects such that if one third of the subjects received a particular target word masked by dissimilar letters, another one third of the subjects received the same word masked by randomly chosen letters, and one third of the subjects received the same word masked by similar letters. Order of trials was randomized and was different for each subject. The dependent variable was the mask delay when the subject first correctly spelled the target word.

SIMILARITY IN MASKING

229

•5200

Results

c

Spelling SOA was considerably longer than word recognition when the same stimuli were used (see Figure 3). Words were spelled at a mean delay of 179 msec, which is much longer than is the case when subjects have to read the word whole. Words masked by similar letters were spelled at a mean delay of 189 msec, words masked by dissimilar letters were spelled at a mean delay of 171 msec, and words masked by randomly chosen letters were spelled at a mean delay of 178 msec; the differences were reliable, F(2, 34) = 5.44, p < .01, and in the predicted direction. The data again suggest the existence of a feature extraction system which identifies letters but not words. It is most parsimonious to consider that the two systems operate simultaneously, word identification not depending on letter recognition among skilled readers. General Discussion The third and fourth studies show that words and separated letter targets are processed differently. While there is strong neurological evidence that the beginning of a feature analysis network exists in the mammalian brain, nevertheless it is clear that we cannot expect it to converge upon single cells for the perception of particular words (Uttal, 1971, as well as the present data; also see Anderson, 1974). The situation must change at some level of processing, and this may be in the region of letter recognition, between angle and line analyses on the one hand and word recognition on the other. The current data support this interpretation, and no other one comes easily to mind. The idea that there are parallel letter and word identification processes may 'be surprising at first glance, but it is not entirely new. Good (1965) speculated that the human brain must not be just a parallel processor, it must be "ultra-parallel," with a very great number of syntheses, analyses, actions, and reactions all happening at one

O Q.

8 H 160 SIMILAR

DISSIMILAR

MASKING

RANDOM

LETTERS

Figure 3. Mean critical stimulus onset asynchrony (SOA, in msec) when target words are spelled as a function of whether masks are comprised of similar, random, or dissimilar letters. (Compare with Figure 1.)

time. What Good did not anticipate was that there might be two sorts of processes, one neurally pyramidal and extractive of features, the other neurally diffuse and integrative of information. Thus, it is easy to imagine a situation like Selfridge's (1959) Pandemonium, in which a number of recognizer "demons" operate simultaneously, each responding in its own circumscribed manner. The organism's total behavior is determined by which demon "shouts" the loudest, and that in turn is determined by the consonance of the various recognizers with regard to the stimulus as well as by the task demands. If the task is spelling or letter recognition, then the organism responds like a set of letter recognizing demons. Otherwise (and in most reading situations) the organism responds like a set of reading demons, which act differently. If it is easy to imagine how two such systems can coexist peacefully, it is not as easy to imagine how they might originate separately. Obviously, letter recognition must precede word reading in children learning or about to learn to read. Perhaps, with practice, letter identification is bypassed in favor of processing of units at least as large as familiar five-, or six-, or seven-letter words. Such large quantitative changes are frequently accompanied by apparently qualitative changes, and possibly that is a sufficient cause for the basic differ-

230

ZACHARY JACOBSON AND GARRY RHINELANDER

ence that we suggest exists between letter and word identification. Perhaps this implies that there is a discernible discontinuity in the process of learning to read a particular word, as the neurology for recognizing that word goes from the first to the second, more direct, mode. The foregoing is problematic only if it is presumed that the recognition process is necessarily serial. Gaugh (1972) suggested that reading is serial, proceeding from letter recognition to reconstruction of the word (perhaps through syllabic units, see Mewhort&Beale, 1977). The present data, however, suggest that such a reconstruction is not the process of normal reading. Indeed, the much greater SOA needed to spell a target word than to read it seems to hint that spelling of a word in this situation takes place itself as a reconstruction after the word is read. However, that does not seem to have been the case in the present Experiment 4—if subjects had read the word and then spelled it, they would have had much less difficulty with the task than they did have. The subjects did not appear to fall into this easier strategy of report; why they did not do so is a puzzle. Whatever the reason for that, it seems that reading of familiar words proceeds wholistically, and to this extent, theories of reading that make that presumption (e.g., Kolers, 1968) are supported over serial theories. It follows that in the study of reading, single letters and (at least unpronounceable) pseudowords are largely inappropriate as stimuli, and studies using only those may produce misleading results. Comparison of Experiments 1, 3, and 4 shows that even spelling real words is fundamentally different from reading them. Another conclusion that may be reached is that the effect of mask-target similarity is subtler than a simple increase or decrease of masking efficacy. Reference Notes 1. Merikle, P. M. Cognitive backward masking: A general type of masking by nonoverlapping

material. Paper presented at the Annual Meeting of the Eastern Psychological Association, Philadelphia, April 1974. 2. Dallas, M., & Merikle, P. M. Associative masking: Logogen interaction or criterion bias? Paper presented at the Annual Meeting of the Canadian Psychological Association, Quebec City, 1975. 3. Arthur, A. Z. Queen's norms for responses to 100 words from the Kent-Rosanoff Word Association Test. National Research Council of Canada project no. APA-0269, Queen's University, Kingston, Ontario, Canada, 1969.

References Anderson, R. M., Jr. Wholistic and particulate approaches in neuropsychology. In W. B. Weiner & D. S. Palermo (Eds.), Cognition and the symbolic processes. Hillsdale, N.J., Erlbaum, 1974. Block, H. D. The perceptron: A model for brain functioning. Reviews of Modern Physics, 1962, 34, 123-135. Block, H. D., Knight, D. W., Jr., & Rosenblatt, S. Analysis of a four-layer, series-coupled perceptron. II. Reviews of Modern Physics, 1962, 34, 135-142. Dodwell, P. C. On perceptual clarity. Psychological Review, 1971, 78, 275-289. Gaugh, P. B. One second of reading. In J. F. Kavanaugh & I. G. Mattingly (Eds.), Language by ear and by eye. Cambridge, Mass.: MIT Press, 1972. Good, I. J. Speculations on the first ultra-intelligent machine. Advances in Computers, 1965, 24, 247-275. Haber, R. N. Effect of prior knowledge of the stimulus on word-recognition processes. Journal of Experimental Psychology, 1965, 69, 282-286. Hebb, D. O. The organisation of behavior. New York: Wiley, 1949. Hubel, D. H., & Wiesel, T. N. Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. Journal of Physiology, 1962,160, 106-154. Jacobson, J. Z. Effects of association upon masking and reading latency. Canadian Journal of Psychology, 1973, 27, 58-69. Jacobson, J. Z. Interaction of similarity to words of visual masks and targets. Journal of Experimental Psychology, 1974, 102, 431-434. Jacobson, J. Z. Relative possibilities of loops and redundant connections in random neural nets. Journal of Mathematical Psychology, 1976, 13, 148-162. Kolers, P. Introduction. In E. G. Huey, The psychology and pedagogy of reading. Cambridge, Mass.: M.I.T. Press, 1968. Konorski, J. Integrative activities of the brain:

SIMILARITY IN MASKING An interdisciplinary approach. Chicago: University of Chicago Press, 1967. McNicol, D., & Howes, P. M. The effects of context on the perception of speech. Australian Journal of Psychology, 1971, 23, 305-310. Mewhort, D. J. K., & Beale, A. L. Mechanisms of word identification. Journal of Experimental Psychology: Human Perception and Performance, 1977, 3, 629-640. Samuels, S. J. Recognition of flashed words by children. Child Development, 1970, 41, 1089-1094. Selfridge, O. G. Pandemonium: A paradigm for learning. In The mechanization of thought processes. London: Her Majesty's Stationery Office, 19S9. Turvey, M. T. On peripheral and central processes in vision: Inferences from an information-pro-

231

cessing analysis of masking with patterned stimuli. Psychological Review, 1973, 80, 1-52. Uttal, W. R. The psychobiological silly season— or—what happens when neurophysiological data become psychological theories. Journal of General Psychology, 1971, 84, 151-166. Uttley, A. M. A theory of the mechanism of learning based on the computation of conditional probabilities. Proceedings of the 1st International Congress of Cybernetics (Namur, 1956). Paris: Gauthier-Villars, 1958. Walley, R. E., & Wieden, T. D. Lateral inhibition and cognitive masking: A neuropsychological theory of attention. Psychological Review, 1973, 80, 284-302. Received October 14, 1977 •

Erratum to Ebbesen et al. In the article, "Laboratory and Field Analyses of Decisions Involving Risk," by Ebbe B. Ebbesen, Stanley Parker, and Vladimir J. Konecni (Journal of Experimental Psychology: Human Perception and Performance, 1977, Vol. 3, No. 4, pp. 576-589), Equation 1 on p. 577 was printed incorrectly. The equation reads :

c,

(1)

+ c,

(1)

It should be changed to read :

R - u(v) + b

Geometric and semantic similarity in visual masking.

Journal of Experimental Psychology: Human Perception and Performance 1978, Vol. 4, No. 2, 224-231 Geometric and Semantic Similarity in Visual Masking...
616KB Sizes 0 Downloads 0 Views