Therole offamiliar unitsin perceptionof words andnonwords

Perception & Psychophysics 1977, Vol. 22 (3), 249-261

The role of familiar units in perception of words and nonwords

JAMES L. McCLELLAND University of California, San Diego, La Jolla, California 92093

and

JAMES C. JOHNSTON Bell Telephone Laboratories, Murray Hill, New Jersey 07974

This paper investigates the effects of familiarity with whole-word units and letter-cluster units in perceptual encoding of letter strings. Subjects viewed brief, masked presentations of words and pronounceable pseudowords differing in letter cluster frequency. Identification of both display types was compared to control single letters. Perceptual accuracy was indexed by probe forced-choice responses and full verbal reports of the displays. Evidence that familiarity of whole-word units facilitated encoding was mixed but, on balance, favorable. Evidence that familiarity of letter-cluster units facilitated encoding was completely absent. This negative finding is surprising in view of the fact that we did obtain a large advantage of letters in pseudowords as well as words over single letters. The discussion section considers an alternative to the view that perceivers use detectors for familiar letter-cluster units in the process of forming representations of pronounceable, orthographically regular letter strings.

Since the pioneering studies of Cattell (1886), research into the role of familiarity in perception has focused on the perception of printed words. In early studies, Cattell and others found that subjects were able to report the contents of a brief tachistoscopic display more accurately if words were shown than if the display consisted of unrelated-letter (UL) strings (Huey, 1908; Neisser, 1967). More recently, Reicher (1969) and other have shown that forced-choice identification of stimulus letters is more accurate. for letters presented in words than for letters in unrelatedletter (UL) strings or even single letters presented alone (Kruger, 1975; Smith & Spoehr, 1974). Since post perceptual guessing and memory overload interpretations are ruled out by Reicher's testing procedure, it appears that familiarity affects the formation of a reportable representation of the presented stimulus and not merely the overt verbal response (Baron, in press). We will refer to the process of

The data presented here were collected while both authors were NSF Predoctoral Fellows at the University of Pennsylvania; the research was supported by the Spencer Research Foundation. Preparation of the manuscript and some of the data analyses were funded by Grant PHS MH 15828 to the Center for Human Information Processing, University of California at San Diego, by Grant BNS76-14830 to the first author, and by Bell Laboratories. We would like to thank Ken Steele for his assistance in preparing the experiment and running subjects, Shelly Milestone for her assistance with some aspects of the data analysis, and Mark Jackson, Kevin O'Regan, Jeff Miller, Dave Meyer, and Don Scarborough for helpful comments on recent versions of this paper. Many others read earlier versions, and we are grateful to them as well.

forming such representations as the perceptual encoding process, and we will consider what perceivers know about words that facilitates perceptual encoding.

Familiarity with whole-word units is not the only kind of knowledge that can facilitate perceptual encoding. A large number of studies of perception of nonwords suggest that accuracy of perceptual encoding varies with the degree to which a letter string embodies the structural regularities of English. Different researchers have stressed different structural properties. It has been shown that nonwords are encoded (or at least reported) more accurately if they are high in sequential redundancy (Miller, Bruner, & Postman, 1954; Rumelhart & Siple, 1974; Smith, 1969), if they are orthographically regular (Baron & Thurston, 1973), or if they are made up of various types of subunits such as "spelling pattern units" (Aderman & Smith, 1971; Gibson, Pick, Osser, & Hammond, 1962; Gibson, Schurcliff, & Yonas, 1970) or "vocalic center groups" (Spoehr & Smith, 1975). These various properties tend to be correlated, and all tend to correlate with pronounceability. In addition, the frequency of occurrence of letter clusters is almost certainly correlated with these indices of structural regularity as well.

One hypothesis which might account for these findings states that the perceptual encoding of letter strings is influenced by familiarity with letter groups of various sizes, including whole words, component letter clusters in words, and of course single letters. Indeed, several investigators have suggested that

249

250 McCLELLAND AND JOHNSTON

perceptual encoding of words and nonwords might be accomplished by a system of detectors for familiar word, letter-cluster, and single-letter units (Estes, 1975; LaBerge & Samuels, 1974; Smith, 1971). We call this hypothesis thefamiliar units hypothesis.

A system of familiar unit detectors could operate in a variety of ways that would facilitate perceptual encoding of letter strings. (1) Detection of only a few features might be sufficient for unambiguous activation of the detector for a higher order unit because of the redundancy of English words and letter clusters (Rumelhart & Siple, 1974; Smith, 1969, 1971). (2) Detectors for units of different sizes might operate at least partially independently of each other so that detectors for larger units would provide additional opportunities to pick up stimulus information not encoded by single-letter detectors. (3) Higher-order unit detectors might receive input from features not used for individual letters (cf', Wheeler, 1970). (4) Higher-order unit detectors might be able to maintain their activation in the presence of masking longer than single-letter detectors, permitting sufficient time for recoding of information into a form suitable for overt responding (Johnston, in press; McClelland, 1976).

Our approach is to look for evidence supporting the familiar-units hypothesis which is independent of any specific proposals. Instead, we rely only on the assumption that probability of activating the detector for a given unit is a monotonic increasing function of the frequency of exposure of the unit (Morton, 1969; Smith, 1971). From this assumption, two predictions follow: (1) Strings containing letter clusters of high familiarity should have an advantage over strings containing clusters of low familiarity, and (2) holding cluster familiarity constant, words should have an advantage over pronounceable and orthographically regular pseudowords.

Evidence Concerning Letter Cluster Detectors Gibson et al. (1962) and Gibson et al. (1970) have

shown that letter strings made up only of familiar letter clusters are reported more accurately if the clusters occur in their normal positions in a word. For example, GL at the beginning of GLURCK is reported more accurately than CK at the beginning of CKURGL (see also Aderman & Smith, 1971). To accommodate these findings, it appears necessary to modify the familiar units hypothesis to state that letter cluster detectors are position specific, receiving activation only from displays containing the given cluster in its typical position. Once this modification of the familiar units hypothesis is accepted, the fact that a number of studies (Gibson et .al., 1970; Spoehr & Smith, 1975) have failed to find a correlation of accuracy of report and overall letter cluster frequency ceases to provide a source of embarassment for the familiar units hypothesis.

All three of the studies mentioned which manip-

ulated position-specific letter cluster frequency (Aderman & Smith, 1971; Gibson et al., 1962; Gibson et al., 1970) confounded this factor with the overall orthographic regularityIpronounceability of the letter string. Thus, these studies do not provide direct support for position-specific cluster detectors over and above any other possible explanation of perception of nonword strings.

One way to try to separate position-specific cluster frequency from other potential factors is to use regression analyses. Gibson et al. (1970) tried this approach. They found that string length was a powerful predictor of report accuracy, accounting for 72lJ1o of the variance in item scores. An index of position- and word-length-specific bigram frequency in conjunction with string length accounted for 83lJ1o of the variance, but rated pronounceability in conjunction with length accounted for 87lJ1o of the variance, and inclusion of the bigram frequency index with both string length and pronounceability only added an additional, insignificant I lJIo. An index of position- and word-length-specific trigram frequency was a poorer predictor than the bigram frequency measure, and did not account for any of the variance after bigram frequency was taken into account. Gibson et al. concluded that pronounceability accounted for all of the variance that letter-cluster frequency accounted for, and more besides.

While suggestive, this pattern of results is hardly conclusive. Several difficulties are associated with Gibson et al.'s (1970) use of the Mayzner and Tresselt (1965) count of the frequency of occurrence of bigrams in running text: (1) The count was based on a sample of only 20,000 words, a small data base for establishing reliable estimates for hundreds of different bigrams in each of scores of position and word-length combinations, (2) If letter-cluster frequency is represented psychologically in a similar manner as word frequency, raw frequencies should have been transformed either logarithmically or exponentially (Catlin, 1971; Oldfield & Wingfield, 1965) to give more weight to differences between moderately frequent clusters and infrequent clusters. (3) Sensitivity of detectors for letter clusters may vary, not with the total number of occurrences of the cluster (token count), but with the number of different words in which it appears (type count; Furth, 1972),

Another way to disentangle cluster familiarity and regularity/pronounceability is to examine effects of bigram frequency within sets of word stimuli or sets of pseudoword stimuli which are pronounceable and regular. In one such study, Biederman (1966) reported a slight report advantage for words containing frequent letter sequences. However, he used a very small number of words, and it is quite possible that his effect was simply due to idiosyncracies of his population of words, since significance levels were not computed over stimuli.

FAMILIAR UNITS IN PERCEPTION 251

Evidence for Word Detectors Since Howes and Solomon (1951), a host of studies

have shown that accuracy of verbal report of words varies with word frequency (Catlin, 1971; Neisser, 1967). However, this effect has often been attributed to postperceptual guessing (e.g., Pierce, 1963) and the effect may be partially contaminated by structural differences between high- and low-frequency words (Landauer & Streeter, 1973). Evidence for a word advantage over pronounceable, orthographically regular pseudowords has now been obtained by three investigations using Reicher's (1969) forced-choice test to control for guessing biases (Juola, Leavitt, & Choe, 1974; Manelis, 1974; McClelland, 1976), but other investigations which have used the same procedure (Baron & Thurston, 1973; Spoehr & Smith, 1975) have not found consistent word-pseudoword differences. All of these studies attempted to equate for pronounceability and regularity, but several did not equate for cluster frequency. Failure to control for this factor may partially account for the inconsistent pattern of results obtained.

EXPERIMENT 1

Our first experiment attempted to provide improved tests of the two predictions we derived from the familiar-units hypothesis: (1) higher performance on items composed of more familiar letter clusters, and (2) higher performance on words than on regular/ pronounceable pseudowords. Our strategy was to manipulate the word-pseudoword factor and high vs. low letter-cluster familiarity orthogonally. Cluster familiarity was indexed by a position- and wordlength-specific type count of bigrams in English words, and this measure was cross-validated by a type count and a token count based on the entire Kucera and Francis (1967) corpus of one million words. We did not investigate trigram frequency, but this factor is highly correlated with bigram frequency and did not contribute to the regression analyses of Gibson et al. (1970).

Our experiment also compared perceptual encoding of letters in words and pseudowords to single letters. We used a postdisplay patterned mask known to produce a word superiority over single letters (Johnston & McClelland, 1973). Since single letters and pseudowords have not been previously compared under such conditions, our experiment is the first to test directly whether the word advantage over single letters depends solely on whole-word familiarity.

As dependent measures, we used a verbal report of the letters presented, in addition to the probe forced-choice test of Reicher (1969). The verbal reports provide a more sensitive measure than the forced choice, because of the 50070 performance floor for the latter measure. In addition, verbal reports permit an analysis of interdependencies in process-

ing letters in multi letter groups not possible with forced-choice data. It has been suggested that verbal reporting undermines the word advantage over single letters (Mezrich, 1973). However, we will show that forced-choice performance is not disturbed by making a verbal report before viewing the forcedchoice alternatives.

Method

Materials

Data were collected from a total of 288 experimental stimuli: 96 words, 96 pronounceable pseudowords, and 96 single letters. The stimuli formed 24 matrices like the one in Table I. In all matrices, the items consisted of two pairs of four-letter words, two pairs of four-letter pseudowords, and two single-letter pairs. Each pair of items differed by a single critical letter, and the critical letters were the same in all six pairs in the same matrix.

All word stimuli occurred at least once in the Thorndike-Lorge G count (Thorndike, 1944), though some may not have been familiar to all subjects. All the pseudowords were initially judged to conform to English orthography by both authors and an undergraduate assistant, and none occurred anywhere in the Thorndike-Lorge corpus.

One word pair and one pseudoword pair in each matrix were high in bigram frequency (high-BF), while the other pairs were low in bigram frequency (low-BF). Within each matrix, the mean bigram frequency of the high-BF members was at least double the mean frequency of the low-BF members of each matrix. Over all stimuli, the bigram-frequency distributions of words and pseudowords were closely matched. As many' of the context letters were kept the same between the matched word and pseudoword pairs as possible, given the interlocking set of contraints on string construction.

The bigram-frequency count used in the selection of the stimuli (Rafferty type count) was derived from a crossword-puzzle dictionary (Rafferty, 1960). Words outside normal English usage (as judged by 1.1.) were not counted. The bigram frequencies of the items selected were validated using two objective measures, a type and a token count derived from the four-letter words in the Kucera and Francis (1967) tabulation of one-million words of English text. The high-BF and low-BF stimuli differed by a 3: I ratio on both type counts and more than a 5: I ratio (.7 log unit) on the token count, but the mean frequencies for words and pseudowords were very similar at the same level of bigram frequency (Table 2).

Within the word stimuli, the high-BF and low-BF items were matched for number of words occurring more than 100 times per million ("AA" words), number between 50 and 100 times per million ("A" words), and for distribution of actual frequencies of occurrence for less frequent words (Thorndike-Lorge G count).

In order to maximize the difference in cluster frequency between the high-BF and low-BF items in each stimulus matrix, it was sometimes necessary to use a different critical letter position for the critical letter in the high-BF and low-BF pairs. Over the 24 matrices, however, the critical letter occurred in each position six times at each level of bigram frequency. To control for positionspecific perceptibility differences against the background of the nonuniform masking pattern used in the experiment, the two

Table 1 One of the 24 Stimulus Matrices Used in Experiment 1

Bigram Frequency

Words

Pseudowords Single Letters

High DILL D~LL VILL V~LL _1_ _ E _

Low ITCH ETCH ILCH ELCH

E_ _

Note-Underlines were not present in the stimuli but were present in the choice alternatives to denote the critical letter position.

252 McCLELLAND ANDJOHNSTON

Table 2 Statistics for Bigram Frequency Measures

Critical Bigram Frequency"

Mean Frequency of All Three Bigrams

Words

Pseudowords

Words

Pseudowords

M

SE

M

SE

M

SE

M

SE

Rafferty Type Countb

High BF

18.4

1.2

18.9

1.3

23.8

.9

22.7

.7

LowBF

6.9

.9

6.8

.9

8.5

.6

8.2

.6

K-F Type Countb

HighBF

20.2

1.4

20.9

1.4

25.6

1.0

25.1

.8

LowBF

7.4

1.0

8.1

1.1

8.9

.7

9.1

.6

K-F Token Count''

HighBF

2.97

.06

3.00

.06

3.09

.04

3.08

.03

LowBF

2.16

.12

2.18

.13

2.36

.08

2.18

.08

aFor critical letters in the second and third positions, critical bigram frequency is the mean of the frequencies of the two bigrams containing the critical letter. bNumber of four-letter-word types containing the bigram in the same position. CLoglO of the number offour-letter-word types containing the bigram in the same position.

single-letter pairs in each matrix occupied the same positions as the corresponding critical letters in the word and pseudoword stimuli.

Design The two members of each item pair were each placed into one

of two separate lists, each of which was presented to 12 subjects. Within each list, the words, pseudowords, and single letters were kept segregated, and subjects knew which kind of item was being presented. However, the items were arranged in random order with respect to critical letter position and bigram frequency. Each list was then divided into two cycles of three 24-trial blocks, and each cycle contained one block of items of each material type. Order of blocks within cycles was counterbalanced over subjects.

The dependent measures obtained in the experiment were based on verbal reports of the letter(s) presented and forced choices between the stimulus and its pairmate (see procedure). To assess the impact of the verbal report on forced-choice performance, reports preceded forced choice on half the trials in each block (first or second half, counterbalanced over subjects).

Subjects The subjects were 24 paid University of Pennsylvania students

with normal or corrected-to-normal vision.

Visual Conditions Stimuli were typed on white cards using an IBM Model 12

electric typewriter with carbon ribbon and presented in a twofield Polymetric tachistoscope. At the viewing distance of 39 ern, stimulus letters were approximately .42 deg high and up to .33 deg wide. A four-letter stimulus subtended approximately 1.40 deg. The pre- and postexposure field consisted of a white card with black, pen-drawn contours in a rectangular array approximately 3.4 deg wide and 1.8 deg high. Both curved and jagged contours were present in an irregular pattern with approximately the same grain size as a capital letter (for a replica of the type of mask used, see Johnston & McClelland, 1973). The luminance of the stimulus field was approximately 1.6 log fL and that of the masking field 1.5 log fL, measured by an SEI photometer.

Procedure At the beginning of each session, the experimenter explained

the forced-choice task and defined the types of stimulus material that would appear. Pseudowords were described as strings of letters that were like words without being familiar or meaningful. Each subject was instructed to fixate the middle of the display area and to try to see all the letters presented, giving equal im-

portance to all posiuons. Trials began with a signal from the experimenter. When ready, the subject pressed a foot switch which started the stimulus presentation after a 300-msec delay. Five seconds after stimulus offset, a click cued the subject to look outside of the tachistoscope and choose between the two alternatives displayed just above the top edge of the tachistoscope at the same viewing distance as the stimuli. Critical letter position was indicated as in Table 1.

On trials in the report condition, the subject gave his report during the interval between the stimulus presentation and the forced-choice test. Each subject was told to report the letter or letters which best represented what he had seen, in the appropriate order. The letters reported did not have to correspond to the type of material presented but did have to contain the correct number of letters.

Each subject viewed 12 blocks of trials grouped into four cycles of three blocks. The first two cycles (20 trials per block) contained a separate list of stimuli and were used for practice and to find an exposure duration at which the subject made close to 750/0 correct forced choices. Cycles 3 and 4 contained the six 24-trial blocks of experimental items. Each block was preceded by two trials used to absorb transition effects. Exposure duration remained fixed within cycles in Cycles 3 and 4 but was adjusted between these cycles if performance on Cycle 3 was much above or below 75% correct forced choice averaged over all three material types.

Results and Discussion Replication and Extension of tbe Word Superiority over Letters

Our experiment replicated the word superiority over single letters obtained previously in the forcedchoice test by Johnston and McClelland (1973, pattern mask condition), Reicher (1969), Wheeler (1970), and others. Further, it extended this effect to pseudowords. Subjects were correct on 80070 of the trials in the word condition, 78070 in the pseudoword condition, and 66070 in the single-letter condition. The main effect of material type was reliable both

Fovi(e2r,9s4u)bj=ect4s3, .1F,s(p2,4 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download