Lexical factors in the word-superiority effect

Memory & Cognition 1995,23 (1), 23-33 Lexical factors in the word-superiority effect NANCY HILDEBRANDT, DAVID CAPLAN, SCOTT SOKOL, and LISA TORREANO...
4 downloads 2 Views 1MB Size
Memory & Cognition

1995,23 (1), 23-33

Lexical factors in the word-superiority effect NANCY HILDEBRANDT, DAVID CAPLAN, SCOTT SOKOL, and LISA TORREANO Massachusetts General Hospital, Boston, Massachusetts and Harvard Medical School, Cambridge, Massachusetts In the Reicher-Wheeler paradigm, fluent readers can identify letters better when they appear in a word than when they appear in either a pronounceable pseudoword (a lexicality effect) or a single letter (a word-letter effect). It was predicted that if both of these effects involve a lexical factor, then adult acquired dyslexic subjects whose deficit prevents access to visual word form should show disruptions of the normal effects on the Reicher-Wheelertask. The results were that dyslexic subjects as well as matched control subjects showed a lexicality effect; however, while the control subjects showed a normal word-letter effect, the dyslexic subjects showed a reverse letter-superiority effect. Both effects, however, showed a systematic variation: As performance on lexical decision improved, the subjects' performance on words in the Reicher-Wheelertask was better than that for all the other conditions. These subject correlations were replicated by using data from a second lexical decision experiment, which utilized the same words and pseudowords that were used in the Reicher-Wheeler task. In addition, an item analysis showed that the words that the subjects had discriminated correctly in lexical decision showed a significant advantage over those that they had not, as well as an improvement relative to the other conditions. These results suggest that there is a lexical factor underlying the lexicality and word-letter effects, and it is proposed that the abnormal letter-superiority effect can be accounted for as the manifestation of other competing factors. tion model, feedback from the word level to the letter level was responsible for producing lexicality as well as word-letter effects. However, to our knowledge, it has not been empirically demonstrated that the word-letter effect has a component that is specifically due to the lexical status of an item. Orthographically familiar pseudowords (e.g., boke) also show a greater advantage over single letters than unpronounceable nonwords do (e.g., kboe) (Estes, 1975), by constraining the possibilities for the occurrence of letters in particular serial positions. Hence, the factor of orthographic familiarity could account for some or even all of the advantage for words over letters, without requiring full activation of a specific lexical item. In addition to the lexicality and word-letter effects, there is also an advantage for orthographically familiar pseudowords over unpronounceable nonwords in the Reicher- Wheeler task. Some researchers have found this advantage for pseudowords over unpronounceable nonwords, but no additional advantage for words over unpronounceable nonwords (Baron & Thurston, 1973), and have claimed that the lexicality effect is in fact an artifact of a slightly greater degree of orthographic familiarity for words than for pseudowords in the stimulus set. Although other researchers have reported a lexicality effect with well-matched stimuli (e.g., Manelis, 1974), it would be of interest to show that the lexicality effect and the word-letter effect have a truly lexical component that cannot be attributed to orthographic familiarity. In this paper, we present evidence from subjects with acquired dyslexia that the lexicality effect and the wordletter effect have a common underlying factor that re-

The term word-superiority effect has been used to refer to two different phenomena that can be observed with very brief tachistoscopic presentations of words and letters followed by a mask (the Reicher-Wheeler paradigm). First, subjects tend to be better at identifying letters presented in the context of a word than those presented in the context of an orthographically familiar pseudoword (Adams, 1979; Juola, Leavitt, & Choe, 1974; Manelis, 1974; McClelland & Johnston, 1977; Spoehr & Smith, 1975). We will refer to this phenomenon as the lexicality effect. Second, subjects can identify letters better when they are presented in the context of a word than when they appear as a single-letter target (Reicher, 1969; Wheeler, 1970). We will refer to this phenomenon as the word-letter effect. These two effects have generally been attributed to a common lexical factor-namely, some advantage gained by activating specific lexical items.' For example, in Rumelhart and McClelland's (1982) interactive activa-

This work was supported by First Investigator's Award DC00462 from the National Institute on Deafness and Communicative Disorders to the first author. A portion of the work was presented as a poster at the meeting ofthe Psychonomic Society, November 1993. The authors wish to thank Betty Jaros and Terry Appleton for their help in testing subjects and analyzing the data. The authors also wish to thank Lori Buchanan, David Gow, Lester Krueger, Ken Paap, William Prinzmetal, and Jay Rueckl for helpful comments and discussion on earlier versions of the paper. L.T. is now in the Department of Psychology at Princeton University. Requests for reprints should be addressed to N. Hildebrandt, Neuropsychology Lab, Massachusetts General Hospital, VBK 827, Boston, MA 02114 (e-mail: [email protected]. harvard.edu).

23

Copyright 1995 Psychonomic Society, Inc.

24

HILDEBRANDT, CAPLAN, SOKOL, AND TORREANO

suits from the activation of specific visual word forms, as measured independently on a lexical decision task. We will also argue that the pattern of results supports the view that this lexical factor is the sole factor producing the advantage for words over orthographically matched pseudowords, but is only one of a number of competing factors underlying the word-letter effect. We first consider in more detail the relationship between orthographic familiarity and specific lexical activation and then discuss various factors that have been proposed to underlie the word-letter effect. Orthographic Familiarity and Specific Lexical Activation It is not clear to what extent the processes involved in orthographic familiarity are independent from the activation of specific visual word forms for fluent readers. Even though the development of a sensitivity to orthographic familiarity is parasitic on the development ofthe ability to recognize specific orthographic patterns representing words, the fully fluent reader's use of this implicit knowledge does not necessarily require accessing specific visual word forms. for example, if a fluent reader of English did not know french, he or she would not be expected to show a lexicality effect, but should still show an effect of orthographic familiarity (expressed as an advantage for orthographically familiar pseudowords over non-french-like nonwords), since french and English have many of the same orthotactic constraints. The relationship and overlap between processes involved in orthographic familiarity and lexicality vary in different models ofword recognition. In Besner and Johnston's (1989) dual-route-based model, there is a stage of orthographic familiarity assessment that can occur without the identification of a specific visual word form. On the other hand, in Seidenberg and McClelland's (1989) connectionist model, there is no lexicon, and words as well as pseudowords set up a pattern ofactivation across distributed representations, resulting in the activation of what might be thought of as "orthographic neighborhoods." The difference between words and orthographically familiar pseudowords in this model is essentially a frequency effect: Patterns of activation for specific words are higher than those for orthographically familiar pseudowords, because the patterns for words have been previously activated for fluent readers. In this model, a lexicality effect would mean little more than a slight quantitative increase in activation for specific visual word forms.? Hence, it is important to show that the lexicality and orthographic familiarity effects can be clearly distinguished from each other.

Prinzmetal (1992) argued for a third factor in the word-letter effect. He found that single letters embedded in symbols (e.g., #w##) were more accurately identified than single letters presented alone in a concurrent pattern mask with unlimited exposure duration. He concluded that the word-letter effect is caused by the greater difficulty of finding a single letter in a mask. Prinzmetal and Silvers (1994) also showed that a word-nonword effect (an advantage of words over unpronounceable nonwords) and the word-letter effect arose in different perceptual contexts. The advantage for words over nonwords occurred not only with a pattern mask, but also under many types of conditions in which the stimulus was visually degraded but presented without a mask. On the other hand, the word-letter effect occurred only in the presence of a pattern mask, and it occurred even when the mask was presented simultaneously with the target with an unlimited exposure duration. Under conditions of visual degradation but no masking, the word-letter effect reversed, yielding a significant advantage for letters over words. The dependence of the wordletter effect on a backward pattern mask had been shown previously under conditions of brief display (Johnston & McClelland, 1973; Juola et aI., 1974; Marchetti & Mewhort, 1986; Massaro & Klitzke, 1979). Prinzmetal and Silvers concluded that the word-nonword effect (and, by extension, the lexicality effect) and the word-letter effect must be caused by different factors. We would argue that the word-letter effect involves even more competing factors, which accounts for its occurrence under perceptual conditions that are different from those for the lexicality effect, even though both effects share a lexical component. With regard to the wordletter effect, we propose that the deleterious effect of lateral masking for words competes with the facilitatory effects of orthographic familiarity and a lexical factor. In the presence of a concurrent or backward-pattern mask, the extra difficulty of finding single letters in the mask predominates, making letters more difficult to detect than words; without a mask, the effect of lateral masking predominates, making letters more recognizable than words (as shown by Prinzmetal & Silvers, 1994). for unpronounceable nonwords in conditions with a concurrent or backward-pattern mask, there is no facilitation from either lexicality or orthographic familiarity, resulting in competition of the remaining effects-finding a single letter in a mask (deleterious to letters) and lateral masking (deleterious to nonwords). This is consistent with our present findings and with other studies in which fluent readers had accuracies that did not differ between single letters and unpronounceable nonwords.

Factors Underlying the Word-Letter Effect The word-letter effect is counterintuitive because the isolated letter is presumed to be perceptually simpler than a letter string. This means that the advantage due to lexical information from words counteracts the perceptual disadvantage of four-letter strings due to lateral masking.

Potential Contribution From Acquired Dyslexia Evidence from neuropathology is potentially valuable for providing information concerning the factors underlying the word-superiority effects. In the brain-damaged population, if it can be shown that there is a specific impairment of a factor hypothesized to playa role in a particular effect, the effect should be altered. With respect

WORD-SUPERIORITY EFFECTS IN DYSLEXIA to the word-superiority effects, there is a straightforward prediction that can be made to test whether there is a common lexical factor underlying both effects: If activation of specific visual word forms underlies not only the lexicality effect but also a portion ofthe word-letter effect, then for dyslexic subjects, the size of the wordletter effect and the lexicality effect should be related to the subjects' ability to activate specific visual word forms. A second prediction concerns the distinction between a lexical factor and a factor of orthographic familiarity: If the processes underlying the effect of orthographic familiarity are the same or very similar to those involved in the activation of specific visual word forms, then the advantage for pseudowords over single letters as well as unpronounceable nonwords should be closely tied to the ability to activate specific visual word forms. In the following study with acquired dyslexic subjects, the dependent variable taken to represent knowledge of specific visual word forms was the ability to discriminate between words and orthographically familiar pseudowords on a written lexical decision task.' In Experiment 1, subjects participated in a Reicher-Wheeler experiment, and the results were compared with their previous lexical decision performance. In Experiment 2, the same subjects participated in a new, written lexical decision task, with the same words and orthographically familiar pseudowords that were used in Experiment 1, in order to enable an item-by-item comparison between the two tasks. PRELIMINARY TESTS FOR DYSLEXIC SUBJECTS DyslexicSubjects We selected subjects who had sustained neurological lesions as adults, either through cerebrovascular accident or trauma, and who were shown to have an impairment specific to written word recognition. Accordingly, we selected those who showed at least one of the following patterns of impairment on a general language screen, a subset of tests from the Psycholinguistic Assessment of Language (Caplan, 1993): (1) difficulty in discriminating written words from nonwords on a lexical decision task, since lexical decision does not require output-side production processes; (2) greater difficulty with either word or nonword naming than with word or nonword repetition, to show that the difficulty with oral reading involved word recognition processes over and above a production deficit; or (3) difficulty with written word-picture matching along with good performance on auditory word-picture matching, to show that the reading difficulty was not due to a more general semantic impairment. These criteria resulted in 12 eligible subjects, ranging from 28 to 80 years old, with a mean age of 58.6 years. Level of education ranged from slightly below a high school degree to PhD, with the median level of education at some post-high-school study. Information about individual subjects is provided in the Appendix.'

25

Tests The dyslexic subjects were tested on a variety of tasks, a subset of which we report here. They performed a casematching task and a physical matching task to assess their letter-recognition abilities as well as their ability to scan a string of letters. They also were given a written lexical decision task to determine their ability to activate visual word form. Case matching. The subjects were asked to decide whether or not pairs of vertically presented letters with differing case had the same name. They scored a mean of .95 in accuracy, with a standard deviation of .070, and a range from .75 to 1.00. Hence, they were able to recognize abstract letter identities, abstracted away from case. Physical matching. The subjects were asked to decide whether or not a pair of letters or letter strings were physically identical. There were four stimulus conditions: (1) single letters (e.g., c), (2) four-letter words (e.g., coin), (3) four-letter pronounceable pseudowords (e.g., crin), and (4) four-letter unpronounceable nonwords (e.g., ncoi). Mismatched pairs were constructed by changing one letter in the target, which produced an item of the same stimulus type (e.g., mint-mind, bace-bice). The mean accuracy was .97, with a standard deviation of .013 and a range from .94 to .98. The mean RT was 2,498 msec, with a standard deviation of 1,097.7 and a range from 1,091 to 4,501 msec. On the basis of these results, the subjects were deemed able to recognize letters and scan four-letter strings. Lexical decision. Waters and Seidenberg (1985) and Seidenberg and McClelland (1989) argued that when sets of words and pseudowords do not sufficiently overlap in orthographic familiarity,normal subjects will make lexical decisions without full lexical activation of the words. Their evidence was that subjects had shown an effect ofspelling-to-sound regularity for words only when irregularly pronounced and orthographically unusual words were included, such as eye (Waters & Seidenberg, 1985, Experiments 1 and 2). For this reason, we used the lexical decision experiment described by Waters and Seidenberg (1985, Experiment 1). There were 72 monosyllabic words, orthogonally varied with respect to frequency (high, low) and spelling-to-sound regularity (regular, exception, orthographically strange). Regular and exception words were matched in bigram frequency, whereas orthographically strange words (an orthographic pattern that had a unique spelling-to-sound correspondence, such as eye) had very low bigram frequencies. There were also 48 pronounceable pseudowords, which consisted of monosyllabic words with the initial letter changed. As Waters and Seidenberg argued, and as Seidenberg and McClelland showed computationally with their connectionist model, there was an overlap in orthographic familiarity between the orthographically strange words and the pseudowords in this list; hence, correct judgments on this test could not be made purely on the basis of orthographic familiarity. Because dyslexic subjects tend to show a positive response bias in lexical decision tasks, their ability to dis-

26

HILDEBRANDT, CAPLAN, SOKOL, AND TORREANO

criminate between words and nonwords was calculated by using the nonparametric signal detection A' discrimination scores.> These scores ranged from .65 to .98, with a median of .90. Ten of the dyslexic subjects showed A I scores of at least .80, reflecting performance that was better than chance, though in some cases it was substantially lower than normal. 6 These results show that, within this group of dyslexic subjects, performance on lexical decision ranged from normal to not significantly different from chance.

Table 2 Sample Trials in Reicher-Wheeler Experiment (Experiment 5)

Single-letter target: +

f

****

f

get ready (1,000 msec)

target

mask (150 msec)

(displayed until response entered)

f

****

f

target

mask (150 msec)

(displayed until response entered)

w

Four-letter target: +

w

EXPERIMENT 1 The Reicher-Wheeler Task

get ready (1,000 msec)

Method Subjects. In addition to the 12 dyslexic subjects, a matched group of 12 control subjects was tested in order to replicate the word-superiority effects for a population older than the usual college freshmen, and to replicate the effects in this particular instantiation of the Reicher-Wheeler experiment. The control group subjects' ages ranged from 27 to 80 years, with a mean age of 58.25 years. Levels of education ranged from slightly below a high school diploma to an MBA degree, the median at a two-year associate degree. Materials. Forty-eight sets of stimulus pairs were created in each of four conditions: words, orthographically familiar pseudowords, unpronounceable nonwords, and isolated letters. An example set is shown in Table I. The mismatched letter in the postcue choice, if substituted for the correct letter, would form another item of the same stimulus type. For example, if a subject incorrectly said that the "w" had appeared inform,ferm, orfrmo, the result would be the word worm, the pseudoword werm, and the nonword wrmo, respectively. For each stimulus set, the postcue letter choices were identical across the four stimulus conditions, in order to control for visual feature similarities of the two letters. For the four-letter conditions, the target letter occurred in each of the four letter positions an equal number of times. Other constraints specific to each condition are described below. Table! Properties of Stimuli in Experiment 1

Stimulus

Post-Cue Choice

Word form worm

f/w f/w

Pronounceable pseudoword ferm werm

f/w f/w

Unpronounceable nonword frmo wrmo

f/w f/w

Single letter

f w

Word frequencies (Francis & Kucera, 1982) Subset A (> 100 occurrences per million) Range, 100-12,458 M,710 Median, 337 Subset B (mixed frequency) Range, 1~80 M,33.7 Median, 16.5

f/w f/w

Word pairs (see Table I) consisted of four-letter words varying in one letter. One member of each word pair (Subset A in Table I) had greater than 100 occurrences per million in the word frequency count of Francis and Kucera (1982), while the frequency of the second member of the pair (Subset B) was uncontrolled. This was done so that the subset of very high-frequency words could later be analyzed apart from the entire set, because of Henderson's (1980) proviso that only high-frequency words might show a lexicality effect. Orthographically familiar pseudoword pairs were constructed by changing one letter in a word pair in a position other than the one being probed (see Table I). Bigram frequencies of the pseudowords were matched to those for the word stimuli by using the norms of Solso and Juel (1980). Bigram frequencies of the two conditions were compared with an analysis of variance (ANOVA), with items as a random factor in a 2 X 2 x 3 design. The independent variables were lexicality (word, pseudoword), word frequency (very high vs. mixed for words, plus their matched pseudowords), and serial position of the big ram (Letters 1-2, 2-3, or 3--4). There were no main effects, and no interactions with a probability less than .58. Unpronounceable non word pairs were constructed by rearranging the four letters from each word to form an unpronounceable combination. The letters in the probe position were the same as those for the matched words and pseudoword conditions. Bigram frequencies were considerably lower than for the word and pseudoword conditions. Single-letter pairs in each set consisted of the same pair that occurred as the poststimulus cue in the other three conditions. Each item of each pair in each condition occurred once in the experiment. There were thus eight stimuli in each of the 48 sets, making a total of384 experimental trials. The stimuli were divided into four testing sessions so that no two of the paired items in each condition occurred in adjacent sessions. Conditions were blocked and ordered in a Latin-square design across the four sessions. Within each block, the stimuli were randomized separately for each subject. An additional 51 letters, 39 words, 39 orthographically familiar pseudowords, and 39 unpronounceable non words were created for practice trials. Practice stimuli were distributed in such a way that the first session contained 27 single letters and 15 of each of the four-letter conditions. The practice for the second, third, and fourth sessions each included 8 items of each condition. The practice trials were blocked in the same Latin-square design that was used for the experimental stimuli for each session. Procedure. The partial-report version of the Reicher-Wheeler task was used, following Reicher (1969). The experiment was run on an IBM portable 386-70 computer with a gas plasma screen, using Micro Experimental Laboratory software (Schneider, 1988). Exposure duration was established during the practice trials at the beginning of each of the four sessions by presenting the four dif-

WORD-SUPERIORITY EFFECTS IN DYSLEXIA

ferent types of stimuli in practice blocks. Within each practice block, exposure duration was adjusted so that the subject would average 75% correct; this was done to avoid ceiling and floor effects. The exposure duration at the end of each of the four practice blocks was then summed and divided by four to calculate a single exposure duration for all four blocks of experimental stimuli in that session. The structure of each trial is shown in Table 2. Following a centrally presented fixation point, the target letter or letter string was presented for the exposure duration established for that session, followed by a mask containing four asterisks. The two-letter response choice then appeared, presented vertically, and remained on the screen until a response was entered. For the four-letter conditions, the serial position was denoted by showing the letter choices accompanied by hyphens to represent the other serial positions. At the beginning of the experiment, the subjects were shown a sample of what the target and the letter choices would look like and were asked to push the upper or lower response key to match their choice of which letter had appeared in the target.

Results The mean exposure duration across the four sessions for the control subjects ranged from 35 to 208 msec, with a median of 70 msec; for the dyslexic subjects, duration ranged from 48 to 456 msec, with a median of 197 msec. Figure I shows a graph of the mean accuracy and standard error for each condition for the acquired dyslexic and control subjects.' An ANOVA was performed with the between-subjects factors of group (dyslexic, control) and condition (letters, words, pseudowords, nonwordsj.f There was a main effect of group [for subjects, F(l,22) = 5.81,p = .025; for items, F(l,380) = 117.73, P < .001];overall, the controls were more accurate than the dyslexics. There was a main effect of condition [for subjects, F(3,66) = 11.19,p < .001; for items, F(3,380) = 14.70,p < .001]. There was also an interaction between group and condition [for subjects, F(3,66) = 22.33,p < .001; for items, F(3,380) = 66.27,p < .001). Simple effects showed that the control subjects were more accurate than the dyslexic subjects for discriminating words, pseudowords, and nonwords, and words were more accurately recognized than pseudowords and nonwords for both subject groups. For the comparisons involving sin-

1.0 Controls Dyslex ic 5 0.9 .>,

u

2 ~

0.8

u u

«

0.7

0.6 L - - - . , - - - - r - - - - - - - . Letters

Words

-t-t-

_

Pseuooworc s Nonwords

Stimulus Type

Figure 1. Accuracy on the Reicher-Wheeler task (Experiment 1).

27

Table 3 Correlations of Lexical Decision Ability (Baseline) With Differences Between Reicher-Wheeler Conditions (Experiment 1)

Condition Correlation Words-letters .699* Words-pseudowords .700* Words-nonwords .711* Pseudowords-letters .299 Nonwords-Ietters .204 Pseudowords-nonwords .246 *Significant correlation with lexical decision task A'.

gle letters, the dyslexics showed better performance in the single-letter condition than in all the other conditions, while the controls showed better performance for words and pseudowords than for single letters, but showed no difference in performance between nonwords and single letters. All simple effects were significant for subjects as well as items. Because of the predicted systematic variation across individual dyslexic subjects, their individual accuracies for each condition are graphed in Figure 2. It can be seen that the dyslexics' performance ranged from that of I subject, whose accuracy was at chance for each condition, to those of several subjects, who more nearly approximated the normal results, showing slightly better performance for words than for letters and better performance for words than for pseudowords. Chi-square values were calculated between conditions for the individual subjects. It was found that although 10 out of 12 control subjects showed a significant word-letter effect, none of the dyslexic subjects showed a significant advantage for words over single letters; in fact, 5 of the dyslexic subjects showed a significant advantage for letters over words. In order to test the prediction that the subjects' ReicherWheeler effects would approach normal as lexical decision ability improved, the size of the difference between pairs of conditions on the Reicher-Wheeler task was correlated with the A' score in lexical decision for each subject. In order to better equate the differences between conditions in different parts of the scale, the accuracy scores for the dyslexics for each condition in the ReicherWheeler experiment were transformed by using an arcsine transformation of the square root of each value (Ferguson, 1981). The accuracy on the single-letter condition was then subtracted from that for each ofthe fourletter conditions, and the accuracy for pseudowords and nonwords was subtracted from that for words. The delta values for each of the six pairwise comparisons between Reicher-Wheeler conditions were then correlated with the A' value obtained in lexical decision for each subject. A scattergram was plotted for each of the six comparisons to make sure that the correlations that did not reach significance did not have some systematic but nonlinear relationship. The Pearson product-moment correlations are shown in Table 3. It can be seen that all of the comparisons involving words reached significance, while all other

28

HILDEBRANDT, CAPLAN, SOKOL, AND TORREANO

········MF 1.0 Rccuracy

-JC --- CC

_.. _. ' •. ;"'_-.. "'-o-:-:-o-,..::":."-.,

Suggest Documents