Journal of Memory and Language 48 (2003) 255–280
Journal of Memory and Language www.elsevier.com/locate/jml
Spelling–sound consistency effects in disyllabic word naming Dan Chateau and Debra Jared* Department of Psychology, University of Western Ontario, London, Ontario, Canada N6A 5C2 Received 10 July 2001; revision received 4 February 2002
Abstract The present study investigated the role of spelling–sound consistency in naming printed disyllabic words. Participants in Experiment 1 named 1000 monomorphemic six-letter disyllabic words. Spelling–sound consistency measures for 11 orthographic segments were used to predict the naming latencies and error rates on the words. The consistency of vowel segments, particularly the one in the second syllable, contributed significantly to the prediction of naming latencies and error rates. In addition, the consistency of the BOB (body-of-the-BOSS, which is the orthographic segment containing the first vowel grapheme and as many following consonants as make an orthographically legal word ending) was also a significant predictor. The effect of the spelling–sound consistency of BOB and V2 segments was replicated in factorial experiments. These findings suggest that readers learn spelling–sound relationships not only for individual letters of graphemes but also for larger orthographic segments in disyllabic words, likely those that provide information about pronunciations beyond that of the individual letters of which they are composed. This study provides the kind of information that is needed to extend current models of word recognition beyond their current focus on monosyllabic words to more complex words. Ó 2002 Elsevier Science (USA). All rights reserved. Keywords: Word recognition; Word naming; Spelling–sound consistency; Phonology; Phonological processing; Disyllabic words; Polysyllabic words; Multisyllabic words
Psycholinguistic research has enabled cognitive psychologists to understand many of the processes involved in reading words. The current understanding of word naming, however, has been gained primarily through research conducted with monosyllabic words. As a result, theories and computational models of word recognition, primarily, or exclusively, address the processing of monosyllabic words (e.g., Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001; Plaut, McClelland, Seidenberg, & Patterson, 1996; Zorzi, Houghton, & Butterworth, 1998). Whether the results of studies using monosyllabic words generalize to polysyllabic words, which in fact make up the majority of words we read, has not yet been explored in much detail. If models of
* Corresponding author. Fax: +519-661-3961. E-mail address:
[email protected] (D. Jared).
English word recognition and naming are to be extended to account for the processing of polysyllabic words, there are many immediate questions that arise. For example, aspects of polysyllabic word recognition that are simply not relevant for processing monosyllabic words include the assignment of stress and the role of the syllable. The likelihood that words contain multiple morphemes, or that a word may be a derivation from another word class, increases as the number of syllables in a word increases. A reader must also determine whether consonants appearing between two vowels should be articulated with the preceding or the following vowel. These additional factors will increase the complexity of current theories and models of word recognition. The most influential theories and computational models of word recognition tend to focus on phonological processing, or the conversion of print to sound (Coltheart et al., 2001; Plaut et al., 1996; Zorzi et al.,
0749-596X/02/$ - see front matter Ó 2002 Elsevier Science (USA). All rights reserved. PII: S 0 7 4 9 - 5 9 6 X ( 0 2 ) 0 0 5 2 1 - 1
256
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
1998). Recent evidence indicates that phonological processing is an important process for activating word meanings, particularly for less familiar words (Jared, Levy, & Rayner, 1999). Because a wordÕs phonological representation may be a major contributor to the activation of its meaning, the factors that influence phonological processing will play an important role in reading comprehension. The aim of the present study is to increase our understanding of how phonological representations of disyllabic words are derived from print. There are two main theoretical approaches to the representation of spelling–sound knowledge. In the dual-route view, spelling–sound knowledge is represented as a set of grapheme–phoneme conversion (GPC) rules that are applied to letter strings one grapheme at a time, from left to right (Coltheart, Curtis, Atkins, & Haller, 1993; Coltheart & Rastle, 1994; Coltheart et al., 2001; Rastle & Coltheart, 1999). An additional lexical look-up system is necessary to correctly name words whose pronunciations do not follow the rules (exception words such as HAVE). The difficulty associated with naming a printed word depends on whether the wordÕs pronunciation is in agreement with the rules. Regular words (words whose pronunciations follow the rules) will be named more quickly and accurately than exception words because both the GPC assembly route and the lexical route produce the correct pronunciation for regular words whereas the two routes produce conflicting pronunciations for exception words. The current computational version of this model does not have polysyllabic words in its vocabulary, although Rastle and Coltheart (2000) have suggested how it might handle the assignment of stress for disyllabic words. Connectionist theorists, on the other hand, view the phonological processing mechanism as one in which readers learn the statistical relationships between spelling and sound, rather than rules (e.g., Plaut et al., 1996; Seidenberg & McClelland, 1989; Zorzi et al., 1998). In this type of theory, every word in the language has its own distributed pattern of activation in each of three pools of units. One pool represents a wordÕs orthography, one pool represents its phonology, and a third represents its meaning. The pools of units are connected by way of hidden units. The hidden units help the system learn generalizations about the relationships between two pools of units. For example, the hidden units mediating between orthographic and phonological units allow the system to detect the sort of orthographic subpatterns predictive of pronunciations. In this view, spelling–sound knowledge is represented as weights on connections between orthographic units and hidden units, and these hidden units and phonological units. Because there are no representations for individual words, and because the same set of units and connections is used for all words, the naming of a given word is influenced by knowledge of other similarly spelled
words. This view predicts that the computation of the correct pronunciation of a word from its orthography is affected by the consistency of the spelling–sound translations in that word, or how often the letters map to one pronunciation versus another (Plaut et al., 1996; Seidenberg & McClelland, 1989). The connectionist approach has not yet been extended to deal with polysyllabic words, although Plaut et al. have stated that, ‘‘the most obvious and natural extension [of this approach] is to the reading of polysyllabic words,’’ (p. 106).
Regularity and consistency effects in polysyllabic word naming There has been one study that attempted to extend findings concerning phonological processing in monosyllabic words to polysyllabic words (Jared & Seidenberg, 1990). Previous research on monosyllabic words had shown that exception words take longer to name than regular words, but only when they are low in frequency (Seidenberg, Waters, Barnes, & Tanenhaus, 1984). Jared and Seidenberg (Expt. 2) examined whether this frequency by regularity interaction occurs for polysyllabic words as well. They defined an exception word as one that had a syllable that was pronounced differently in the word than when presented in isolation (e.g., DANGER). Their results indicated that the regularity effect was stronger for low-frequency disyllabic words than it was for high-frequency words, replicating the studies using monosyllabic words. Jared and Seidenberg (1990, Expt. 1) also examined whether spelling–sound consistency affected disyllabic word naming. Extending GlushkoÕs (1979) work with monosyllabic words, regular-inconsistent words were included in the experiment along with exception words and matched regular-consistent words. Regular-inconsistent words (e.g., DIVINE) contained a syllable that is regular in the target word, but that is given an exception pronunciation in another word (e.g., RAVINE). Regular-inconsistent words and exception words took longer to name than their matched regular-consistent words, and exception words took longer to name than regularinconsistent words. The findings of these two experiments indicate that the typicality of spelling–sound correspondences affects disyllabic word naming, just as it does monosyllabic word naming. Two further experiments examined whether polysyllabic words are parsed into syllables and pronunciations determined syllable by syllable. Experiment 3 matched pairs of short, medium, and long words for number of letters, but varied the number of syllables (e.g., CRUISE/ CROCUS, APPROACH/ADDITION, COMMISSION/ COMPETITION). If pronunciations are computed syllable by syllable, then for each word length, words with
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
one more syllable should take longer to name. An effect of the number of syllables was found, but only for low-frequency words. This effect should also have been found for high-frequency words if pronunciations are computed syllable by syllable. In Experiment 4, the initial syllable of disyllabic words was presented 250 ms prior to, and above, the second syllable. If polysyllabic words are routinely parsed into syllables in the course of processing, naming latencies should be affected very little as a result of the unusual display. The results of the study, however, demonstrated that naming latencies for exception words were significantly longer than in Experiment 2, in which the same words were presented intact. The increase in naming latencies was similar for both high- and low-frequency exception words. Importantly, latencies for regular words were unaffected by the change in display, indicating that the procedure did not make reading harder in general. These results suggest that the translation of print to sound does not occur on a syllable by syllable basis. Rather it appears that information from outside the syllable can be used to generate the correct pronunciation of polysyllabic words, particularly exception words. A remaining question, however, is what factor was responsible for the effect of number of syllables reported for low-frequency words in Experiment 3. Jared and Seidenberg (1990) suggested that words with a greater number of syllables also had more vowels to process, and because the pronunciations of vowels are typically more ambiguous than those of consonants, it was the increased number of vowels that was responsible for the syllable-length effect. For exception words, the context of consonants surrounding or following a vowel might be necessary to produce the correct pronunciation.
Consistency effects in monosyllabic words Since Jared and SeidenbergÕs (1990) experiments were run in the mid-1980s, much has been learned from studies with monosyllabic words about the nature of spelling–sound consistency effects. In particular, evidence suggests that consistency should not be viewed as a categorical variable, but rather that its effects are graded. In GlushkoÕs (1979) studies and in other studies that followed, words were categorized as inconsistent if they had one or more word-body neighbors with a conflicting pronunciation (e.g., WASTE is inconsistent because of neighbor CASTE). Using this definition, some researchers found that regular-inconsistent words took longer to name than regular-consistent words (e.g., Andrews, 1982), and others did not (e.g., Stanhope & Parkin, 1987). However, some inconsistent words have very few and/or very infrequent neighbors with conflicting pronunciations (hereafter referred to as ÔenemiesÕ), while other inconsistent words have many and/or very frequent enemies. The same could be said for the
257
neighbors of a word that share a pronunciation (hereafter referred to as ÔfriendsÕ); some inconsistent words have many and/or very frequent friends while others have few and/or infrequent friends. Jared, McRae, and Seidenberg (1990) explored whether the conflicting results of previous studies occurred because the degree of consistency is relevant for processing, and not simply whether or not a word has enemies. Jared et al. (Expt. 2) used four sets of low-frequency inconsistent words, factorially manipulating the summed token frequency of the target wordsÕ enemies (high, low) and the target wordÕs friends (high, low). When compared to control groups of matched consistent words, the largest consistency effects were obtained with words having a low summed frequency of friends and a high summed frequency of enemies. No consistency effect was obtained for inconsistent words having a high summed frequency of friends and a low summed frequency of enemies. Words with a high summed frequency of both friends and enemies, or a low summed frequency of both, produced small effects between the two extremes. A summary of the results of 15 experiments by other researchers demonstrated that significant effects of consistency were obtained only when the words included in the experiment had a low mean summed frequency of friends and a high mean summed frequency of enemies (Jared et al., 1990, Table 5). The results of this experiment demonstrate the importance of a continuous rather than a categorical view of spelling–sound consistency. Subsequent research provided evidence that the naming of high-frequency words is also affected by their word-body consistency (Jared, 1997, 2002). Spelling–sound consistency effects were observed for highfrequency words when they had the same neighborhood characteristics as low-frequency words that produce consistency effects. In the studies conducted by Jared (1997, 2002) and Jared et al. (1990), measures of consistency were based on monosyllabic word neighbors only, ignoring the possible impact of longer words on monosyllabic word recognition. There has been one study that did examine spelling–sound consistency effects in monosyllabic words using measures of consistency that included polysyllabic words as potential neighbors (Treiman, Mullenix, Bijeljac-Babic, & Richmond-Welty, 1995). The database used to calculate consistency statistics was the computerized version of the Merriam-Webster Pocket Dictionary, which contains 19,750 words, most of which are polysyllabic. Using monosyllabic words that had a CVC structure, Treiman et al. not only examined the impact of word-body consistency as Jared et al. (1990) had, they also examined the impact of the consistency of other orthographic segments. Spelling– sound consistency was calculated for onsets ðC1 Þ, vowels (V), codas ðC2 Þ, as well as the word-body ðVC2 Þ, and the onset–vowel segment ðC1 VÞ. The consistency of a
258
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
particular pronunciation of an orthographic segment was defined as the number of friends divided by the number of all words with the letter pattern. Orthographic segments that are pronounced the same way in all of the words in which they appear had a value of 1.0 (e.g., UST, in MUST, DUST, COMBUST). Orthographic segments that have more than one pronunciation had values somewhere between 0 and 1.0 (e.g., AVE, in GAVE, BEHAVE, HAVE, OCTAVE). In essence, this consistency measure is the proportion of similarly pronounced neighbors. Treiman et al. (1995) first presented an analysis that showed which of the orthographic segments had the most reliable spelling-to-sound mappings. For each monosyllabic CVC word, the consistency of each of the five segments was calculated as described above. To determine the reliability of an orthographic segment as a class, the mean of the consistency values for that segment across the entire set of CVC words was calculated. For example, to determine the reliability of C1 segments in CVC words, the proportion of consistent neighbors for the C1 segment in each CVC monosyllabic word was summed and divided by the total number of CVC words in the set. The C1 and C2 segments in the CVC word set had the most reliable pronunciations when all of the words in the dictionary were considered as potential neighbors of the CVC words. Using token counts, the mean proportions of similarly pronounced neighbors were .95 and .84, respectively. This indicates that one could almost always be certain of the pronunciation of the consonants in the CVC word set. The calculation of reliability for the vowel segments in CVC monosyllabic words was a bit more complicated. Because polysyllabic words were included in the database, monosyllabic wordsÕ vowel consistencies were calculated twice. In one case, the vowels that appeared in the initial syllable of polysyllabic words were used to calculate consistency, and in the second case, the vowels that appeared in the final syllable of polysyllabic words were used to calculate consistency. Vowels that appeared in syllables in the middle of a polysyllabic word were not considered. The V segments in CVC monosyllabic words had a low consistency no matter which vowel position of polysyllabic words was used to determine potential neighbors (.46 for initial syllable vowels, .47 for final syllable vowels). On average, then, a vowel segment in a CVC monosyllabic word is pronounced the same in only about half of the other the words in which the segment appears. The C1 V segment (word initial in polysyllabic words) was not consistent in its translation to phonology either (.49). This indicates that the initial consonant in a monosyllabic word does little to constrain the pronunciation of the vowel. Only when combined with the C2 segment was the vowel pronunciation somewhat consistent, with the VC2 segment (final syllable position in
polysyllabic words) having a higher average proportion of similar pronunciations (.72) than the V alone, or the C1 V. In other words, the final consonants in a monosyllabic word constrained the possible pronunciations of the vowel. Somewhat surprisingly, the vocabulary used to determine the consistency of spelling–sound segments did not affect the measures of consistency. When only the monosyllabic words in the dictionary were considered as possible neighbors, the proportions of consistent pronunciations for the VC2 and the other orthographic segments were quite similar. Treiman et al. (1995) then conducted regression analyses on the naming latencies and percentage of errors from two mega-studies, each of which required participants to name in excess of one thousand words. The goal was to determine which, if any, of the orthographic segmentsÕ consistencies predicted performance on the naming task. The five consistency measures ðC1 ; V1 ; C2 ; C1 V1 ; V1 C2 Þ were included as predictors, as well as several other relevant variables, such as printed frequency, characteristics of the initial phoneme, homophony, and interactions between the predictors. Treiman et al. (1995) presented the results of the regression analyses using the consistency measures based on type counts and with only monosyllabic words as potential neighbors. The pattern of results was apparently quite similar when polysyllabic words were also considered as potential neighbors, and when token counts were used rather than type counts. Overall, the regression equations were able to account for a significant proportion of the variance in naming latencies in the two mega-studies (R2 ¼ :38 and .50). Not surprisingly, the printed frequency of the words contributed significantly to the prediction of naming latencies and error rates in both studies. The consistency of the onset considered alone ðC1 Þ and the consistency of the wordbody ðVC2 Þ were the best consistency predictors of naming latencies and errors for both of the mega-studies. As mentioned, the previous analysis indicated that the C2 consonants constrain the pronunciation of the vowel, and these regression results indicate that readers learn about this relationship and use it when computing a wordÕs pronunciation. This result also fits well with the findings from factorial studies examining monosyllabic word-body consistency effects (e.g., Jared et al., 1990).
Spelling–sound segments in disyllabic words It is quite apparent from this work on monosyllabic words that readers learn not only about relationships between individual graphemes and their pronunciations but also between word bodies and pronunciations. Knowledge of the relationships between the spellings and pronunciations of word bodies appears to be particularly useful for naming monosyllabic words because
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
the consonants following the vowel constrain the possible pronunciations of the vowel (Treiman et al., 1995). Readers may also learn spelling–sound relationships for orthographic segments in disyllabic words that are larger than individual graphemes. However, it is unclear whether there is an orthographic segment that provides information about pronunciations in disyllabic words beyond that provided by simple segments, as does the word body in monosyllabic words, and if so whether it respects syllable boundaries. Assuming there is such a segment or segments, it is also unclear whether readers could pick up on such complex relationships. It could be that readers learn the relationship between spelling and sound for segments that are within the same phonological syllable. Although Jared and Seidenberg (1990) argued that the pronunciations of polysyllabic words are not computed syllable by syllable, other evidence suggests that polysyllabic words are parsed into syllables. Ferrand, Segui, and Humphreys (1997), for example, used a syllable priming procedure to determine whether orthographic representations are structured according to phonological syllables. Facilitation was obtained for a target word only when the prime word corresponded to the first syllable of clearly syllabifiable words, and not when the prime was one letter more or one letter less than the syllable. This is an important result, because an orthographic parse limits the consonants that could be used to constrain vowel pronunciations in the phonological processing of disyllabic words. However, a closer look at the target words in the experiment where primes had one letter more than the phonological syllable reveals that the phonology of the prime (e.g., REP) and the phonology of the target (e.g., REPORT) often did not match. The pattern of priming obtained by Ferrand et al. (1997), therefore, could in fact be due to the degree of phonological overlap between the prime and target. Furthermore, Schiller (2000) failed to replicate Ferrand et al.Õs (1997) syllable priming effect using the same materials and method. Nevertheless, given the importance of understanding the role of syllables in word naming, the impact of the spelling–sound consistency of segments constrained by syllable boundaries was examined in the present study. Another possibility is that readers learn the relationship between spelling and sound for orthographic segments that are larger than a syllable, or that are not limited by syllable boundaries. There has, in fact, been some investigation into an orthographic segment that includes consonants from beyond the syllable boundary. The BOSS, or Basic Orthographic Syllable Structure, was first proposed as a lexical access unit by Taft (1979), and was thought to be the orthographic access code to the mental lexicon. The BOSS was based purely on a wordÕs orthography and was defined as follows:
259
‘‘Include in the first syllable as many consonants following the first vowel of the word as orthotactic factors will allow without disrupting the morphological structure of that word.’’ (Taft, 1979, p. 24)
The orthotactic factors referred to by Taft specify which combinations of letters can appear at the end of a word (i.e., ND can appear at the end of a word but DN cannot), and frequently will include consonants that are part of the second phonological syllable. TaftÕs (1992) later research on the BOSS took note of the fact that word naming experiments had identified the word body as an important segment in the derivation of phonology for monosyllabic words. The question addressed by Taft (1992) was what orthographic segment constituted the ÔbodyÕ of a polysyllabic word. TaftÕs argument in this series of experiments was that the bodyof-the-BOSS (BOB; e.g., EAD in MEADOW) was the important unit for the derivation of phonology for polysyllabic words. Evidence for this claim came from an experiment in which he showed that the likelihood of assigning second-syllable stress to a target disyllabic pseudoword (e.g., CAMULK) given a prime word with second-syllable stress was greater when the prime had the same BOB as the target (e.g., LAMENT) than when the prime shared no letters with the target (e.g., DIVERT). No such influence on stress assignment was observed for primes sharing the first syllable with the target (e.g., CAVORT). This finding led Taft to suggest that BOB nodes in the lexical access system are linked to subword units in the phonological lexicon. The phonology of unknown words or pseudowords, such as CAMULK, are generated using these mappings. Because the BOB has been shown to have an effect on disyllabic word naming, the spelling–sound consistency of this orthographic segment was also examined in the present study.
The present study The aim of the present study is to extend our knowledge of the phonological processing of monosyllabic words to disyllabic words. The study examines whether the spelling–sound consistency of any simple or higher-order orthographic segment influences the naming of disyllabic words. Two general approaches have been taken in empirical work on phonological processing: factorial design studies (e.g., Jared et al., 1990), and regression studies (e.g., Treiman et al., 1995). When stimuli are equated on other relevant variables, a factorial design allows an investigator to confidently assert that the manipulation of the independent variables was responsible for significant effects. The regression design allows research to be more exploratory in nature, examining the effect of many different variables with a single data set, including variables that are not identified
260
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
a priori. Given the advantages and disadvantages of each of the factorial and regression designs, some investigators have advocated the use of both (e.g., Treiman et al., 1995). This combination of methodologies was employed in the present study.
Experiment 1 In Experiment 1, analyses quite similar to Treiman et al.Õs (1995) were conducted on disyllabic word naming data. Spelling–sound consistency statistics were calculated for a variety of letters and letter clusters that appeared in 1000 disyllabic words, and these consistency statistics were used to predict naming latencies and error rates on those 1000 words. Because this work is the first to examine spelling–sound consistency in disyllabic words, it was not known at the outset what the influential spelling–sound segments of a disyllabic word would be, and because there are several different possibilities, the regression method allowed for multiple measures of consistency to be evaluated. The most basic step up from monosyllabic word naming was taken in this research. Only monomorphemic words were included as target words in these studies so that effects due to morphological complexity were avoided. All words were six letters long to eliminate effects of word length and to ensure that a single fixation would be sufficient to encode the stimuli. The main question for Experiment 1, then, was what are the orthographic segments in disyllabic words whose spelling–sound consistency affects naming latencies and errors? In the present study, the pool of neighbors that was used to determine the consistency of an orthographic segment in a disyllabic word included all other disyllabic words of any length with the same orthographic segment in the same syllable position. Disyllabic words tend not to be stressed on the second syllable in English, and this alters the pronunciation of vowels that appear in the second syllable compared to when they appear in monosyllabic words or the initial syllables of disyllabic words. It is unlikely that readers fail to learn this correlation between the position of a vowel segment and its typical pronunciation, so considering only segments in the same position in disyllabic words seems appropriate. Even if there is some influence of words where the same segment appears in a different position, this may not substantially alter consistency statistics. Limiting neighbors to disyllabic words makes the computations of consistency much easier and may also have little impact on consistency statistics. For monosyllabic words, Treiman et al. (1995) found no substantial difference in the predictive power of the consistency measures when they were based only on monosyllabic words compared to when they were based on all words.
The orthographic segments whose spelling–sound consistencies were evaluated included simple segments in the first syllable (e.g., C1 ; V1 ; C2 ), simple segments in the second syllable (e.g., C3 ; V2 ; C4 ), orthographic segments that include the vowel segment and any consonant(s) prior to the vowel in the same phonological syllable ðC1 V1 ; C3 V2 Þ, orthographic segments that include the vowel segment and any consonant(s) after the vowel that are in the same phonological syllable ðV1 C2 ; V2 C4 Þ, and the BOB, which includes the first vowel and all following consonants that constitute legal word endings. The symbols C and V refer not just to single consonants and vowels, but rather to all letters of that type in a syllable before the next type is encountered. For example, the SP in SPIDER is its C1 and the EA in BEAKER is its V1 . Method Participants Twenty-nine undergraduate students at the University of Western Ontario participated in the study. Participants volunteered for the study as one method of earning credit in an introductory Psychology class. All participants were native English speakers and had normal or corrected to normal vision. Apparatus An Apple MacIntosh LC III computer with a MacIntosh Color Display monitor was used for this study. The computer was interfaced with a model Mk6 button box, developed by MacWhinney, Laxman, and Taylor, to which a Realistic 33-1060 electret condensor microphone was connected. The button box was used to time participantsÕ naming latencies to the nearest millisecond. The application program PsyScope (version 1.0.2; Cohen, MacWhinney, Flatt, & Provost, 1993) was used to display stimuli and record participantsÕ naming latencies. Materials One thousand disyllabic words were taken from the CELEX lexical database (Baayen, Piepenbrock, & Gulikers, 1995). There were two constraints on the choice of words. First, the target words had to be monomorphemic (according to the CELEX database). In addition, the target words had to be six letters in length. Any disyllabic word in the database that matched these criteria was included in an initial list. Words that first year undergraduate students were unlikely to know were then removed from this initial list. The remaining list was trimmed to 1000 words. Procedure Participants in the experiment were tested individually in a session that lasted approximately 55 min.
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
Participants were first shown standard naming instructions displayed on the computer monitor. When the participant was ready to begin, they were given 10 practice trials. During this time, the experimenter adjusted the sensitivity of the microphone in order to ensure that it would pick up the participantÕs voice, but not be activated due to incidental noises. Each individual trial began with a word presented in the center of the computer monitor in lowercase Chicago 14-point font. The word remained on the screen until the participant voiced a response. A timer in the button box timed the naming latency in milliseconds from the onset of the stimulus to the onset of the participantÕs response. The intertrial interval was 1500 ms. The experimenter recorded participantsÕ incorrect pronunciations by hand. The list of 1000 words was presented in a different random order for each participant. Participants were given a short break at intervals of 250 trials. Independent variables The predictor variables used in this regression study are similar to those employed by Treiman et al. (1995). The log CELEX frequency of each word, the characteristics of the initial phoneme, and the stress pattern of the word (e.g., first-syllable or second-syllable) served as an initial set of predictors. Ten dummy variables for the initial phoneme were chosen based on Treiman et al.Õs (1995) analysis: one for voiced versus unvoiced, four for manner of articulation (nasal, fricative, liquid/semivowel, affricate), and five for place of articulation (bilabial, labiodental, alveolar/alveopalatal, velar, and glottal).1 Stress pattern was included as a predictor because studies have found that English words with stress on the first syllable are named faster than those with stress on the second syllable (e.g., Brown, Lupker, & Colombo, 1994). Characteristics of the initial phoneme and log frequency are not of theoretical interest in the present study, but were included to remove variance associated with those variables. In addition, Taft (personal communication, April, 2000) has suggested that fewer intervening consonants between vowels could make processing more difficult, regardless of the particular spelling–sound consistency for the segments. The number of intervocalic consonants (i.e., consonants between the two vowel segments) was also included as a predictor in the regression analyses. The inclusion of these predictors allows for a more accurate estimation of the impact of spelling–sound consistency.
1
We did not include dummy variables for the first vowel, but because recent evidence (Kessler, Treiman, & Mullennix, 2002) suggests that the triggering of voice keys may also be affected by vowel characteristics, future research will need to develop a way to code vowel quality for regression analyses.
261
Spelling–sound consistency measures were calculated for each of the following segments: C1 , V1 , C2 , C3 , V2 , C4 , C1 V1 , V1 C2 , BOB, C3 V2 , and V2 C4 . An example of the division of a word into the segments is presented for the word VERTEX in Fig. 1. Examples of words with high- and low-consistency segments are given in Table 1. Because the structure of disyllabic words varies, not all words have a value for all of the segments. For instance, BELONG has no C2 segment, and BRANDY has no C4 segment. Each of the Cs and Vs could include one or more letters. Those with more than one letter (e.g., TH, AI) were considered to be a different C or V than those with a single letter (e.g., T or A). Therefore, the pool of words used to determine the consistency of a multi-letter C or V did not include words containing only one of the letters. For example, the consistency of the spelling– sound mapping for TH, as in METHOD, was not affected by words such as MOTIVE, and vice-versa. The identification of some of the complex orthographic segments in disyllabic words is not as simple as for monosyllabic words. For disyllabic words, the identification of V1 C2 segments was based on the syllable boundary as given by the CELEX database. If words did not have a C2 ðn ¼ 249Þ, the consistency statistic for the V1 C2 was calculated using just the V1 . The consis-
Fig. 1. The division of the word VERTEX into orthographic segments for calculating spelling–sound consistency. In addition to the 10 segments shown above, the consistency of the BOB, which is ERT for VERTEX, was also calculated.
Table 1 Examples of high-consistency and low-consistency words for each orthographic segment Segment C1 C1 V1 V1 V1 C2 C2 C3 C3 V2 V2 V2 C4 BOB
High-consistency word
Low-consistency word
cannon gossip hyphen barber custom banter foster domain canine donkey
circle govern system warden dismal virtue vertex certain famine monkey
262
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
tency of the V1 C2 segment was not the same as V1 consistency for these cases because the pool of neighbors used when calculating V1 consistency included all words with the V1 regardless of the following C2 , whereas the pool of neighbors used when calculating V1 C2 consistency included only words with the V1 and no C2 . Similarly, for many words, the BOB segment and the V1 C2 segment were comprised of the same letters. However, the two consistency measures were always different for a given word. This is because the pool of words used to determine the consistency statistics for each segment were different. For the V1 C2 pool, only other words with the same C2 at the end of the first syllable were included, but this syllable constraint was absent for BOBs. For example, the word CANCEL has the letters AN as both its V1 C2 and its BOB (the BOB is not ANC because NC cannot appear as a word ending in English). Although the word CANINE also has the letters AN, it would be included in the pool of words used to calculate BOB consistency but not V1 C2 consistency. This is because the N is part of the second syllable of CANINE and not the first. The C3 V2 was identified using the C3 segment in most cases, and the second vowel segment. When a C2 segment was ambisyllabic and there was no C3 segment, the C2 segment was treated as a C3 solely for the purposes of the calculating C3 V2 consistency (e.g., MANAGE, n ¼ 128). The V2 segment was identified in the same manner as for the initial vowel, except that in this position, there were a number of cases in which a vowel letter would be followed by a consonant, and then the letter E (e.g., RAVINE). These cases were to be considered a single, but discontinuous, vowel segment. In this manner, the pronunciation of words containing I_E, for instance, did not affect the consistency measures for I with no following E. These discontinuous vowel segments were also used in the identification of C3 V2 and V2 C4 segments. In the case of words ending in a consonant and LE (e.g., MINGLE), the identification of segments was a little more complicated. The initial consonant in the syllable was treated as the C3 (e.g., the G in MINGLE). The L was treated as the C4 segment, and the E was the V2 segment. For the C3 V2 , the LE and the preceding consonant were included in the orthographic segment, with potential neighbors being all words that had these letters beginning the second syllable (e.g., GARGLE, NEGLECT, JUGGLER). For the identification of V2 C4 segments, the LE formed the V2 C4 orthographic segment ðn ¼ 139Þ, and only words that ended in LE could be neighbors. The same procedure was used for words that ended in RE, such as MEAGRE. Words in which a word-final C4 is an R or L also present a problem because the preceding vowel is often not pronounced (e.g., FOSTER). However, two-syllable words must have a V2 . In these cases we considered the
V2 to be a schwa. This may not be an entirely satisfactory solution because schwas not followed by an R or L have a slightly different pronunciation (e.g., WANTED). Future work will need to explore better ways to handle this issue. The pronunciations listed in the CELEX database were altered to reflect Canadian phonology. The various pronunciations for each of the orthographic segments were then identified, and the words containing the orthographic segment were separated into different lists, with one for each pronunciation. In order for two words with the same BOB to be considered to have the same BOB pronunciation, the BOB consonant(s) had to have the same pattern of phonological syllable assignment. Thus, the BOB segment AND was considered to have the same pronunciation in PANDER and SANDAL (both have /n/ pronounced in the first syllable and /d/ pronounced in the second syllable), but a different pronunciation in SANDSTONE (/n/ and /d/ are pronounced in the same syllable). The method for calculating spelling–sound consistency was similar to that used by Treiman et al. (1995). However, rather than calculate a type-based consistency measure and a token-based consistency measure, a single consistency measure was calculated as follows. First, the logarithm of the token count for each word in the database was calculated. To obtain the consistency of a particular pronunciation for an orthographic segment, the log frequencies of all the words in which the segment was pronounced the same were summed. This sum was then divided by the total of the log frequencies for all words in which the orthographic segment appeared, no matter how it was pronounced. A consistent pronunciation, in which all words with the orthographic segment are pronounced the same, had a value of 1.0. As an example, the C1 V1 letter pattern BU has seven possible pronunciations for Canadian speakers. The sum of the logarithm of the token frequencies for each word in which BU is pronounced as in BUGLE is 6.159. This value was then divided by the sum of the logarithm of the token frequencies for all words in which BU appears as a C1 V1 regardless of the pronunciation (330.065). This particular pronunciation for BU, therefore, has a consistency value of .0187. Results The data for 92 words were not included in any of the analyses because the initial phoneme was a vowel, not a consonant. These were excluded because the consistency statistics for the initial vowel when preceded by a consonant are likely to be different from the consistency statistics for words without consonantal onsets. Additionally, the initial-phoneme predictors employed in this study are based on words beginning with a consonant, rather than a vowel. Eight more words were not
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
included because there were no medial consonants between the vowels (e.g., CLIENT). Our consistency statistics for these words may be misleading because monosyllabic words such as FRIEND were not included as potential neighbors. Four other words were removed from the data set because three or fewer participants named the word correctly. For words that had two possible correct stress patterns ðn ¼ 18Þ, the most frequently given stress pattern was treated as correct, and trials from participants responding with the alternate stress pattern were excluded from the data set. Trials in which the naming latency was greater than 1500 ms were not included in the computation of a wordÕs average naming latency. Across all participants this affected approximately 1.3% of the trials, with the maximum for a single subject being 2.2%. Following Treiman et al. (1995), multiple regression analyses in which all predictors were simultaneously entered into the regression equation were conducted on the logarithm of the mean naming latency for each word and the square root of the proportion of errors. Because not all words have a value for every one of the segments, regression analyses using all consistency measures as predictors could not include data from all of the words. This is because a wordÕs data is excluded from a regression analysis if there is a missing value. Only 477 of the 1000 words presented to the participants had a spelling–sound consistency value for all of the orthographic segments (these were the CVCCVC words). The simultaneous regression analyses, therefore, included data from just these words. The simultaneous entry regression analyses have a limitation, however, in that they include many variables in the regression equation that do not contribute significantly to the prediction of naming latencies or errors. The impact of those variables that do affect naming latencies may, then, be underestimated because of shared variance with the non-significant predictors. For this reason, hierarchical stepwise regression analyses were also conducted in which only those variables that contributed significantly to the prediction of the criterion variable were included in the equation. Because the varying structure of disyllabic words means that not all words have a value for each of the consistency predictors, a series of hierarchical regression analyses were conducted in which different subsets of consistency predictors were entered and data from different subsets of words were included. Multiple regression on naming latencies: Simultaneous entry of all predictors The first analyses on naming latencies and error rates forced all 24 predictors into the regression equation: 10 initial phoneme dummy variables, 11 spelling–sound consistency measures, log frequency, the number of intervocalic consonants, and stress pattern. Only data from the 477 words that had consistency values for all
263
segments were included as the criterion. The regression equation predicting log latency was significant, F ð23; 453Þ ¼ 13:97, p < :001, MSe ¼ :001. The equation accounted for 42.0% of the variance in the data set (38.8% adjusted).2 The results for initial phoneme predictors are not presented here because they are not of theoretical interest in the present study. The results for the other predictors are presented in Table 2. Log frequency contributed significantly to the regression equation, with higher frequency words being named faster than lower frequency words. Of the spelling–sound consistency measures, V1 C2 consistency, C3 V2 consistency, V2 C4 consistency, and BOB consistency each contributed a significant unique proportion of variance to the regression equation. Words with higher consistency measures for these segments produced faster latencies than words with lower consistency measures. Stress pattern also contributed significantly to the regression equation, with second-syllable-stress words producing longer latencies than first-syllable-stress words. An alternative analysis described by Lorch and Myers (1990) was also run using the same set of predictors. According to Lorch and Myers, in the repeated measures case, a regression equation should be generated for each individual tested, and then a single sample t test should be conducted on the set of b weights to determine whether they are significantly different from zero. This procedure was run here and the results are also presented in Table 2. Multiple regression on error rates: Simultaneous entry of all predictors The same 24 predictors were entered in an analysis of error rates (see Table 3). The regression equation predicting error rates was significant, F ð23; 453Þ ¼ 8:87, p < :001, MSe ¼ :009, accounting for 31.0% of the variance in the data set (27.5% adjusted). Log frequency contributed significantly to the regression equation, with higher frequency words producing fewer errors than lower frequency words. Stress pattern also contributed significantly to the regression equation, with secondsyllable-stress words producing more errors than firstsyllable-stress words. As in the analysis of naming latencies, BOB consistency contributed significantly to the regression equation. Words with a higher consistency BOB produced fewer errors than words with a lower consistency BOB. 2
Adjusted values for Ôvariance accounted forÕ reflect the fact that in any multivariate regression, the inclusion of more predictors can only increase the R2 value (it can never decrease with the inclusion of more predictors). Therefore, as a matter of routine, there is an downward adjustment to the R2 value that is determined by the number of cases and the number of predictors involved in the analysis.
264
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
Table 2 Results for simultaneous and hierarchical regression analyses predicting naming latencies N Words
Simultaneous 477
Predictor Frequency b t
477 L&M
Hierarchical 477
519
600
647
710
767
833
896
896 L&M
).383 )10.32
)12.42
).389 )10.58
).409 )11.77
).435 )13.13
).443 )14.08
).430 )15.05
).439 )15.82
).463 )17.40
).467 )17.94
)14.86
.087 2.26
1.43
.100 2.66
.051 1.41
.063 1.89
.026 .18
.009 .27
.031 .99
).030 .98
).025 .86
.92
C1 b t
).058 )1.39
).62
).063 )1.60
).052 )1.40
).051 )1.49
).053 )1.64
).019 ).65
).040 )1.41
).022 ).81
).047 )1.81
).89
V1 b t
.013 .27
).95
.008 .17
).073 )2.08
).015 ).36
).079 )2.48
).046 )1.46
).046 )1.49
).069 )2.42
).080 )2.90
)7.64
C2 b t
.004 .09
).006 ).13
).074 )2.08
).022 ).69
).016 ).40
NI
NI
NI
NI
NI
).58
C3 b t
.070 1.73
.038 .97
.054 1.48
NI
NI
NI
)1.62
V2 b t
).070 )1.38
)4.72
).138 )3.00
C4 b t
).051 )1.34
)1.22
C 1 V1 b t
).082 )1.73
V 1 C2 b t
Stress b t
NI
NI
.004 .14
.003 .11
).202 )5.69
).180 )5.37
).175 )5.50
).081 )2.20
).096 )2.68
).097 )2.83
).157 )5.83
)9.36
).064 )1.72
NI
).062 )1.86
NI
).113 )3.89
NI
).111 )4.08
NI
NI
)1.88
).075 )1.99
).051 )1.16
).105 )3.14
).072 )1.83
).037 )1.19
).028 ).92
).038 )1.01
).024 ).66
)2.87
).101 )1.97
)4.90
).084 )2.06
).094 )1.93
).055 )1.48
).077 )1.97
).035 )1.09
).082 )2.63
).008 ).027
).047 )1.56
)2.87
C 3 V2 b t
).098 )2.07
)3.25
).098 )2.16
).080 )1.84
).034 ).84
).015 ).38
).112 )3.23
).106 )3.07
).072 )2.22
).055 )1.71
)1.66
V 2 C4 b t
).091 )2.15
).073 )1.76
NI
).057 )1.59
NI
NI
NI
)4.05
BOB b t
).100 )2.32
)2.15
).106 )2.56
).107 )2.92
).137 )3.70
).077 )2.14
).145 )4.49
).082 )2.34
).120 )3.78
).104 )3.39
)8.63
# intervocC b t
).053 )1.42
)5.50
).047 )1.28
).033 ).94
).092 )2.50
).106 )3.03
).173 )5.19
).180 )5.53
).152 )4.79
).138 )4.54
)3.53
).021 ).66
NI
).012 ).43
Note. L&M, analyses using the Lorch and Myers procedure, df ¼ 28; NI, not included in the analysis; # intervocC, number of intervocalic consonants. The t values in bold have p < :05.
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
265
Table 3 Results for simultaneous and hierarchical regression analyses predicting naming errors N Words
Simultaneous 477
Predictor Frequency b t
Hierarchical 477
519
600
647
710
767
833
896
).186 )4.62
).169 )4.23
).191 )4.83
).228 )6.21
).232 )6.45
).286 )8.75
).294 )9.09
).317 )10.39
).310 )10.31
Stress b t
.303 7.21
.315 7.71
.260 6.39
.218 5.84
.186 5.06
.101 2.83
.071 1.98
.074 2.21
.056 1.71
C1 b t
.095 2.10
.026 .66
.016 .41
.056 1.52
.043 1.18
.073 2.17
.045 1.37
.078 2.47
.042 1.37
V1 b t
.012 .27
).012 ).30
).025 ).61
).004 ).09
).004 ).08
).071 )1.96
).078 )2.21
).078 )2.32
).100 )3.10
C2 b t
).018 ).35
).017 ).42
).040 ).99
.051 1.37
.020 .055
NI
NI
NI
NI
C3 b t
.078 1.77
.042 1.01
.067 1.66
NI
NI
V2 b t
).065 )1.18
).177 )4.29
C4 b t
).056 )1.36
).067 )1.64
C1 V1 b t
).065 )1.26
V1 C2 b t
.009 .26
).005 ).16
).043 ).98
).087 )2.06
).074 )1.84
).120 )3.17
NI
).123 )3.84
NI
NI
NI
).156 )3.80
).165 )4.40
).144 )3.89
NI
).053 )1.42
NI
).028 ).71
).041 )1.04
).088 )2.38
).090 )2.47
.013 .28
.005 .12
).037 ).87
).019 ).47
).040 .73
).003 ).07
).028 ).62
.044 1.07
.021 .53
).001 ).02
).034 ).92
.032 .90
).007 ).20
C3 V 2 b t
).083 )1.61
).060 )1.22
).065 )1.31
.003 .07
).004 ).10
).159 )4.52
).139 )3.47
).105 )2.80
).083 )2.23
V2 C4 b t
).078 )1.71
).068 )1.55
NI
).059 )1.48
NI
).080 )2.26
NI
).060 )1.79
NI
BOB b t
).221 )4.74
).204 )5.08
).185 )4.64
).171 )4.60
).162 )4.45
).168 )4.60
).156 )4.38
).166 )4.98
).158 )4.89
).005 ).13
.011 .28
).009 ).22
).004 ).09
).017 ).41
).016 ).38
).019 ).49
).019 ).50
).049 )1.39
# intervocC b t
).136 )4.05
Note. NI, not included in the analysis; # intervocC, number of intervocalic consonants. The t values in bold have p < :05.
266
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
The only other consistency measure to contribute significantly to the regression equation was C1 consistency, but the effect was in a direction opposite to what one would expect. Words having a higher consistency produced more errors than words with a lower consistency. To explore the reason for this unexpected finding, we examined the C1 consistency value for each of the words in the experiment. C1 consistency was uniformly quite high (>.87) except for 11 words that began with the letter C pronounced as /s/ (e.g., CIRCLE), which had a C1 consistency value of .07. These words, however, had very few errors ( 6 3%). To see whether they were responsible for the positive relationship between C1 consistency and errors, we reran the error rate analysis with these words eliminated. C1 consistency did not contribute to the prediction of errors in the analysis with this small change. Thus, the CI-type words were responsible for producing the positive relationship between C1 consistency and errors. The low C1 consistency for the 11 CI-type words likely did not impair their naming because the pronunciation of the C1 was highly predictable given the following vowel.3 Hierarchical regression analyses Eight hierarchical regression analyses were run on naming latencies and on error rates. The first hierarchical analysis for each included all of the spelling– sound consistency measures, and thus were hierarchical versions of the previous simultaneous regression analyses. The same set of 477 CVCCVC words was used. Six additional analyses were run that entered all possible subsets of the simple segment predictors. This allows for the determination of the impact of spelling–sound consistency of the orthographic segments on naming words of various structures. Each of these analyses predicted naming latencies and errors for a different subset of the 896 wordsÕ naming latencies and error rates. The eighth and final analysis entered only C1 , V1 , and V2 , as simple orthographic segment predictors, and the larger segments other than the V2 C4 . All of the words in the data file included these segments ðN ¼ 896Þ. For each of the hierarchical regression analyses on naming latencies and error rates, three blocks of predictors were entered. In the first block, only the initial phoneme characteristics, log frequency, and stress pattern were considered in a stepwise procedure. These variables were considered first so that the variance as3 The Lorch and Myers (1990) procedure described in the analysis on naming latencies was not run on the error rates. While error rate can be considered to be a continuous variable when averaged across a group of individuals, the variable is dichotomous for an individualÕs data set. A participant either named the word correctly or incorrectly. Therefore, linear regression analyses, which provide the necessary b weight, cannot be run on an individualÕs error data.
sociated with them would be accounted for before examining the impact of spelling–sound consistency predictors. In the stepwise procedure, however, the only variables that are entered in the equation are the ones that contribute significantly to the prediction of the criterion variable, so not all of the predictors that were considered would be part of the regression equation at the end of the first block. In the second block, the spelling–sound consistency measures for simple segments (e.g., C1 , V1 , C2 , C3 , V2 , C4 ) were forced into the regression equation, whether or not they contributed significantly. This would attribute as much variance as possible to the spelling–sound consistency of small orthographic segments before examining the impact of larger orthographic segments on naming latencies. In the third and final block, all of the larger orthographic segment consistency measures were considered in a stepwise procedure (e.g., C1 V1 , V1 C2 , BOB, etc.), as well as the number of intervocalic consonants. In this block, the simple segment consistency measures that were entered on the second block and stress pattern were included again, so that if any were not contributing significantly at this point, the stepwise procedure would remove them from the regression equation. After the final block, therefore, only the variables that contributed significantly to the prediction of naming latencies or errors were included in the regression equations. Naming latencies. The results for all eight hierarchical regression analyses are presented in Table 2. Two are described here. One entered all of the predictors, but included only 477 words. This regression equation was significant, F ð12; 464Þ ¼ 25:52, p < :001, MSe ¼ .071, and accounted for 39.8% of the variance in log naming latencies (38.2% adjusted). V2 consistency was the only simple segment that was a significant predictor; of the higher-order segments, C1 V1 , V1 C2 , C3 V2 , and BOB consistency were significant. The other analysis entered only the consistency measures that all of the target words contained and therefore it included the full set of words ðN ¼ 896Þ. This analysis also resulted in a regression equation which was significant, F ð9; 886Þ ¼ 68:73, p < :001. It accounted for 41.1% of the variance in log naming latencies (40.5% adjusted). V1 and V2 consistency were the only simple segments that were significant predictors, and BOB consistency was the only significant higher-order segment. The Lorch and Myers (1990) analyses were also run on the set of 896 words. The results are presented in Table 2. Error rates. The results for all eight hierarchical regression analyses are presented in Table 3. Again, two are presented here. One entered all of the predictors, but included only 477 words. This analysis resulted in a regression equation which accounted for 26.2% of the variance in error rates (25.6% adjusted), F ð6; 470Þ ¼ 41:92, p < :001. V2 consistency was the only simple segment that was a significant predictor and BOB
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
consistency was the only significant higher-order segment. The other analysis entered only the consistency measures that all of the target words contained ðN ¼ 896Þ. This analysis resulted in a regression equation which accounted for 20.0% of the variance in error rates (19.4% adjusted), F ð8; 895Þ ¼ 31:54, p < :001, MSe ¼ .089. V1 and V2 consistencies were the only simple segments that were significant predictors, and C3 V2 and BOB consistencies were the only significant higher-order segments. Discussion The results of this study provide clear evidence that spelling–sound consistency influences the naming of disyllabic words. Across the regression analyses, the spelling–sound consistency of two segments were particularly good predictors of naming performance. The spelling–sound consistency of the BOB was always a significant predictor, and the consistency of the V2 was entered as a significant predictor in most of the analyses. These results suggest that readers learn spelling–sound relationships for both simple and higher-order orthographic segments. Simple orthographic segments In the present study, the only simple spelling–sound consistency measure that was repeatedly entered in the regression equations was V2 consistency. Although it was not entered in the standard simultaneous entry analyses on naming latencies and errors, V2 consistency contributed significantly to all of the standard hierarchical regression analyses on latencies and all but two of the analyses on error rates. As well, when the Lorch and Myers (1990) procedure was conducted on latency data, V2 consistency was a significant predictor of naming latencies for both the simultaneous entry analysis and the hierarchical analysis. The spelling–sound consistency of the V1 segment was not a significant predictor in the standard simultaneous entry analyses on naming latency or errors, or in the simultaneous entry analysis using the Lorch and Myers (1990) procedure on naming latencies. However, V1 consistency was a significant predictor of either latencies or errors (or both) for all but two of the data sets used in the hierarchical regression analyses. The impact of the consistency of this segment alone, therefore, is not strong. In combination with the results for the V2 segment, however, the conclusion can be drawn that the spelling–sound consistency of vowels affected the naming of disyllabic words. These results support the hypothesis made by Jared and Seidenberg (1990) that the computation of vowel pronunciations may be the cause of the syllable-length effects reported in their Experiment 3. In contrast, variation in C consistency was found to contribute little to variation in naming latencies and
267
errors. This is likely because the pronunciation of most Cs is highly consistent; that is, there are few alternate pronunciations for consonants. With little variation in consistency, it is difficult for a C segment to predict variance in naming performance. Furthermore, some of the variation in consonant pronunciations, in particular that of C and G, is highly predictable from the following vowel. Such predictability would attenuate the influence of the inconsistency of the consonant. It is possible that there may be some cost to naming performance of a word containing a consonant with an atypical and unpredictable pronunciation (e.g., HONEST), but because there are so few of these words, they have little impact on the regression analyses. A factorial study might reveal consonant consistency effects if enough stimuli could be found. Larger orthographic segments In addition to the simple orthographic segments, the consistency of larger orthographic segments also contributed to the prediction of naming latencies and error rates in this study. In the simultaneous entry analyses, and in all of the hierarchical regression analyses on naming latencies and error rates, BOB consistency was a significant predictor. The higher the consistency of the BOB, the more quickly and accurately words were named. The fact that the BOB was a significant predictor of naming performance even after the simple segments were entered into the regression equations indicates that it provides information about pronunciations beyond the information provided by the simple segments of which it is composed, and that readers have learned about the relationships between BOBs and pronunciations. If readers had learned only about relationships between simple segments and their pronunciations, then the consistency of the V1 segment, the C2 segment, and the C3 segment would have captured all of the relevant variance in naming latencies and BOB consistency would not have been entered. A likely reason that the BOB is an influential segment in disyllabic word naming is that it constrains the possible pronunciations of the initial vowel segment and medial consonants in a way similar to the monosyllabic word body. Such a constraining influence may be the reason that V1 consistency was not a stronger predictor of naming. We included BOB consistency in this study because there was some prior evidence that the BOB is relevant for naming disyllabic words (Taft, 1992). It was a good predictor of disyllabic word naming performance in Experiment 1 and so we will further evaluate this segment in Experiments 2 and 3. However, in the General discussion we will raise the question as to whether there is anything special about the BOB segment or whether some other segment that also includes letters from the second syllable would be as good or better a predictor of naming latencies.
268
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
In contrast to the strong influence of BOB consistency, the consistency of the phonological-syllable defined V1 C2 segment was weaker. It was a significant predictor of naming latencies in the standard and in the Lorch and Myers (1990) simultaneous regression analyses, in three of the eight standard hierarchical analyses, and in the Lorch and Myers hierarchical analysis. However, it was not a significant predictor of naming errors in any of the analyses. The spelling–sound consistency of the V1 C2 segment, therefore, may have a small influence on the phonological processing of disyllabic words or may have an influence on just some of them. Experiment 2 examines the relative impact of BOB consistency and V1 C2 consistency on disyllabic word naming in a tightly controlled factorial design to determine whether the consistency of both segments does indeed affect performance. Other larger orthographic segments were not frequently entered in the regression equations on naming latencies. The spelling–sound consistency of the V2 C4 segment, the counterpart of the BOB=V1 C2 for the second syllable, was a significant predictor of naming latencies in both the standard and Lorch and Myers (1990) simultaneous regression analyses, but was a significant predictor in only one hierarchical regression analysis and that was on errors. This evidence suggests that the consistency of the V2 C4 segment does not have much impact on the computation of phonological representations. A possible explanation is that the final consonant does not often help constrain the pronunciation of the second vowel. The relatively rare occurrences of non-lax vowels in the second syllable in English may obviate the need for a phonological processing mechanism to pick up on contingencies between V2 and C4 segments and their pronunciations. The consistency of the C1 V1 segment also did not often contribute significantly to the regression equations in the hierarchical analyses. This is not surprising, because for monosyllabic words C1 V consistency has not been shown to predict either naming latencies or error rates (Treiman et al., 1995). Treiman et al. provided evidence that initial consonants in monosyllabic words provide little constraint on the pronunciation of the following vowel. The fact that the spelling–sound consistency of the C1 V1 segment was not often entered in the regression equations here suggests that initial consonants in disyllabic words also typically provide little information about the pronunciation of the following vowel, or conversely, that vowels provide little information about the pronunciation of the preceding consonant. However, this seems to be less the case for C3 V2 segments. The consistency of the C3 V2 segment was a significant predictor in both the standard and Lorch and Myers (1990) simultaneous regression analyses on naming latencies, in four hierarchical analyses on latencies, and in three hierarchical analyses on errors.
It was a particularly strong predictor in analyses that included words that did not have a C2 . Although consonants may not typically provide information about the following vowels, there are several exceptions to this general trend that may be responsible for the entry of C1 V1 and C3 V2 segments into some regression equations. For example, W provides information about the pronunciation of a subsequent A, as in WALLET, and the vowels I and E provide information about the pronunciation of a preceding C or G. Among the words used in this experiment, there were over twice as many C3 segments as C1 segments with a C or G followed by an I or E (39 versus 17). This may account for the finding that C3 V2 consistency was more often a significant predictor than C1 V1 consistency. There are also additional pronunciations of consonants in the C3 position, conditioned by the following vowels, that simply do not exist for the same consonants in the C1 position (e.g., the C in SOCIAL). The observations here concerning words beginning with CI or with CI appearing in the medial consonant position raise an issue that will need to be explored further. The issue is whether good predictability of the pronunciation of a higher-order segment can compensate completely for low predictability of the pronunciation of a simple segment that it contains. These results also provide some evidence for Jared and SeidenbergÕs (1990) suggestion that the orthographic representation of a disyllabic word is not parsed into syllables and pronunciations computed syllable by syllable. If that were the case, then one would have expected the consistency of the V1 C2 to have been identified as the strongest spelling–sound consistency predictor because it would correspond to the body of the orthographic segment being mapped to phonology. Likewise, the consistency of the BOB should not be a significant predictor, because the word recognition system would not process the BOB as a single orthographic segment. The results from this study point to the opposite conclusion, suggesting that, because the BOB is a significant predictor, the word recognition system does not parse the orthography of polysyllabic words along phonological syllable boundaries. A role for syllables in reading and naming words may arise within the phonological representation, after activation from orthography, or in the later activation of motor programs from phonological representations. Indeed, effects of phonological syllable frequency have been observed in other studies (Carreiras, Alvarez, & de Vega, 1993; Perea & Carreiras, 1998), although these were conducted in Spanish, not in English, and would have to be replicated in English before strong conclusions regarding the syllableÕs role in English can be made from this finding. Levelt and WheeldonÕs (1994) work providing support for the presence of a mental syllabary also suggests a later role for the syllable.
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
There is an additional reason from these findings to believe that a wordÕs orthography is not parsed into unique segments and pronunciations computed segment by segment. Experiment 1 indicated that V1 consistency has some impact on phonological processing. This finding is surprising because the V1 segment is contained in the BOB. It appears, then, that a letter can participate in spelling–sound mappings of several sizes, each of whose consistency can affect the naming of disyllabic words. This type of finding strongly suggests that phonological processing is a matter of a distributed mapping from orthography to phonology. It would be difficult to implement a system with a parse that can account for effects from different segments that both contain the same letters (i.e., V1 and BOB). The focus of the present research was on spelling– sound consistency. However, we included stress pattern as a predictor in our regression equations, and therefore we can make a few comments regarding the influence of stress on word naming. An interesting issue concerns whether there is an influence of stress pattern that cannot be accounted for by spelling–sound consistency measures. Stress pattern was a significant predictor in some of the regression analyses here, so clearly spelling–sound consistency measures do not completely account for stress effects. However, stress pattern was a much weaker predictor of naming latencies than many would suspect. Stress pattern was significant in the latency analyses only when all of the consistency predictors were entered. In this case, just words with a CVCCVC structure were included in the analyses ðN ¼ 477Þ. Stress was not a significant predictor of naming latencies in any of the other analyses, including the Lorch and Myers (1990) analysis on the same set of 477 words. A closer look at the CVCCVC words revealed that there are only nine words with stress on the second syllable (1.9%), and these are named considerably more slowly than the CVCCVC words with stress on the first syllable (710 ms versus 622 ms). The extreme rarity of CVCCVC words with second-syllable stress may have caused participants to hesitate slightly when they encountered one. Secondsyllable stress was more common among the words with other structures. Of the remaining 419 words, 96 are stressed on the second syllable (23.0%). When subsets of these were added to the regression analyses on naming latencies, stress was no longer a significant predictor. Thus, second-syllable stress words do not always incur a naming delay; some likely even have a naming advantage over first-syllable stress words. More fine-grained work will need to determine whether readers learn a relationship between word structure and likely stress pattern. The results for the error analyses are very different from those for latencies. Stress pattern was a significant predictor of errors in all but one analysis. However, this finding must be interpreted with caution. The participants in this study named a very large number of words
269
(1000) in one session, most of which had first-syllable stress. In fact, the ratio of first-syllable stress to secondsyllable stress was almost 8:1. A difference this large could lead to a bias toward stressing words on the first syllable, causing more errors when a word is encountered that has stress placed on the second syllable. Experiments in which the ratio of first- and second-syllable stress words is varied would help determine whether errors are due solely to the stress pattern or whether they are largely influenced by the composition of the stimulus list.
Experiment 2 Because Experiment 1 is the first study to compare the influence of a variety of spelling–sound consistency segments in disyllabic words, replicating any significant findings from the first study with different participants and using a factorial design is an appropriate next step. If effects are replicated, the influence of the orthographic segmentÕs consistency on disyllabic word naming is more clearly established. The regression study provides strong evidence of the importance of the BOB segmentÕs spelling–sound consistency for the naming of disyllabic words. This finding suggests that readers do not learn only about relationships between single graphemes and phonemes. Furthermore, it suggests that pronunciations of disyllabic words are not computed syllable by syllable, but rather that pronunciations of letters in one syllable may be constrained by letters appearing in the other syllable. However, the conclusion of some previous research with disyllabic words is that they are divided into their constituent syllables (Ferrand et al., 1997). In addition, V1 C2 consistency was a significant predictor in some of the analyses on naming latencies, and so its contribution to naming is unclear. To provide further evidence concerning the contribution of V1 C2 and BOB consistency to naming, Experiment 2 contrasted naming performance on words at the extremes of both segmentsÕ consistency distributions. Method Participants Thirty undergraduate students at the University of Western Ontario participated in the experiment. None of these participants had completed Experiment 1. Participants volunteered for the study as one method of earning credit in an introductory psychology class. All participants were native English speakers and had normal or corrected to normal vision. Materials One hundred disyllabic words chosen from the CELEX lexical database were included in this experi-
270
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
ment. There were four lists of words. Two lists were matched for the consistency of the V1 C2 segment, and the consistency of the BOB segment was varied (high, low). In the other two lists, the words were matched for the consistency of the BOB segment, and the consistency of the V1 C2 segment was varied (high, low). Pairs of lists were also matched for number of intervocalic consonants, V1 consistency, V2 consistency, the number of phonemes, mean printed word frequency (using the CELEX norms), and initial phoneme. For the BOB lists, the constraint on matching V2 consistency was relaxed slightly in order to get a sufficient number of stimuli. The high-consistency BOB items had a lower average V2 consistency ðM ¼ :611Þ than the low-consistency BOB items ðM ¼ :726Þ, although not significantly so, tð48Þ ¼ 1:54, p ¼ :13. Importantly, this difference in V2 consistency can only work against the hypothesis that the spelling–sound consistency of the BOB affects phonological processing, because high-consistency V2 segments led to faster naming latencies in Experiment 1. Following the suggestion of Kawamoto, Kello, Jones, and Bame (1998) and Kawamoto, Kello, Higareda, and Vu (1999), the initial phonemes of the words were also matched for simple and complex onsets. In addition, Kessler et al. (2002) reported that the vowel following the initial consonants in words also has an impact on voice response time measurements. For this reason, whenever possible, the following vowel was also matched for the high- and low-consistency conditions (e.g., BEACON–BEAKER, SPIRAL–SPIDER). All words were monomorphemic, six letters in length, and were stressed on the first syllable. The characteristics of these words can be found in Table 4. The full list of words can be found in Appendix A. Procedure Participants in the experiment were tested individually, and the testing session lasted approximately 10 min. The procedure was the same as for Experiment 1, except that there were only 100 experimental trials and there were no breaks during the experiment. Each participant saw a different random ordering of the stimuli. Table 4 Characteristics of the stimuli in Experiment 2 Variable
High BOB
Low BOB
High V1 C2
Low V1 C2
Frequency per million Median frequency BOB consistency V1 C2 consistency V1 consistency V2 consistency No. of phonemes
9.2 2.7 .829 .518 .453 .611 5.44
8.9 3.1 .390 .517 .453 .726 5.52
9.8 1.6 .796 .749 .535 .742 5.40
8.8 5.2 .783 .230 .535 .751 5.48
No pairwise comparison on any of the control consistency measures was significant (all ps > :15).
Results For this experiment, naming latencies that exceeded 1200 ms were not included in the analyses. Latencies exceeding this value were rare and represented a substantial deviation from every participantÕs mean naming latency. This affected approximately 1.6% of the data. For the analysis treating participants as a random factor ðF1 Þ, consistency (high, low) and orthographic segment (BOB, V1 C2 ) were analyzed as within-participants factors. For the analysis treating items as a random factor ðF2 Þ, the same variables were analyzed as between-items factors. The mean naming latencies and error rates by participants are presented in Fig. 2. Of particular interest for this experiment was the interaction between consistency and segment, which was significant by participants, F1 ð1; 29Þ ¼ 21:83, p < :001, MSe ¼ 167:61, but not by items, F2 ð1; 96Þ ¼ 2:16, p ¼ :14, MSe ¼ 1071:23. To further examine this interaction, tests of the simple main effects of consistency were conducted. The effect of BOB consistency was significant in the analysis by participants, F1 ð1; 29Þ ¼ 24:28, p < :001, MSe ¼ 172:91, and approached significance in the analysis by items, F2 ð1; 96Þ ¼ 3:12, p ¼ :08, MSe ¼ 167:61. As expected, high-BOB-consistency words were named faster than low-BOB-consistency words. The effect of consistency was not significant for V1 C2 words, F1 ð1; 29Þ ¼ 2:63, ns, MSe ¼ 163:53, F2 < 1. The main effect of consistency was significant by participants, F1 ð1; 29Þ ¼ 5:77, p < :05, MSe ¼ 168:83, but not by items, F2 ð1; 96Þ ¼ 1:06, ns, MSe ¼ 1071:23. In the analysis of error rates, there were no significant effects. In particular, there was no indication of an interaction between consistency and item type, F s < 1. Overall, participants in this experiment made few errors. The low-consistency BOB condition (2.9%) produced a slightly higher error rate than the high-consistency BOB condition (1.9%), while the two V1 C2 conditions produced the same number of errors (2.1% for both high and low consistency). Discussion The results of this experiment support the interpretation of Experiment 1 by providing further evidence that the spelling–sound consistency of the BOB segment influences readersÕ ability to name disyllabic words. Low-consistency BOB words took significantly longer to name than high-consistency BOB words. In contrast, this experiment provided little evidence that the spelling–sound consistency of the phonological-syllable defined V1 C2 segment affects naming latencies or errors. The results from this factorial experiment also suggest that readers do not compute pronunciations syllable by syllable. If they had, then there should have been an effect of V1 C2 consistency. Instead, the results suggest
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
271
Fig. 2. Mean naming latencies and error rates (in parentheses) for conditions in Experiment 2.
that consonants beyond the syllable boundary provide information that readers use to help determine the pronunciation of the initial part of the word. One possible criticism is that the difference in V1 consistency between the BOB and V1 C2 conditions may have biased the results towards the BOB condition. The numbers in Table 4 do not support this view however. The difference in V1 consistency for the two conditions is very small (.08), while the difference in the naming latency results is striking. Additional analyses on the data with V1 consistency as a covariate produced an identical pattern of results.4
Experiment 3 Initial research on spelling–sound regularity and consistency effects in monosyllabic words appeared to indicate that spelling–sound consistency does not affect naming latencies for high-frequency words (e.g., Seidenberg et al., 1984). Furthermore, Jared and Seidenberg (1990) found a significant frequency by regularity interaction in their study of polysyllabic word naming. Naming latencies for high-frequency disyllabic words, therefore, may not be affected by the spelling–sound consistency of the BOB. On the other hand, more recent evidence suggests that spelling–sound consistency does 4
An items analysis of the naming latencies from Experiment 1 for the words that also appeared in Experiment 2 were included in that experiment ðN ¼ 94Þ revealed the same pattern. The interaction between the segment and consistency was marginally significant, F ð1; 89Þ ¼ 3:06, p ¼ :08, MSe ¼ 1461:48. The effect of consistency for BOB words (628 ms versus 600 ms) was significant, F ð1; 89Þ ¼ 4:41, p < :05, MSe ¼ 1570:81, but was not for V1 C2 words (599 ms versus 601 ms), F < 1. There were no significant effects for error rates.
affect the processing of high-frequency monosyllabic words. First, Treiman et al. (1995) found that frequencyby-consistency interaction measures did not predict naming latencies or error rates in their analyses of two mega-studies. Second, Jared (1997, 2002) found significant consistency effects for high-frequency words when the frequencies of friends and enemies were the same as for low-frequency words that produce consistency effects. Interactions of frequency and consistency were weak, and primarily due to a few low-frequency words with particularly unusual pronunciations (e.g., PLAID; Jared, 2002). This more recent evidence suggests that the spelling–sound consistency of the BOB should affect the phonological processing of high-frequency disyllabic words as well as low-frequency disyllabic words. However, because frequency-by-consistency interaction measures were not included in the analyses on the data from Experiment 1, the question has not been addressed here. A factorial design experiment was conducted to answer this question. Method Participants Thirty-six undergraduate students who had not participated in the previous experiments completed this experiment. Participants volunteered for the study as one method of earning credit in an introductory psychology class. All participants were native English speakers and had normal or corrected to normal vision. Materials and procedure Sixty-four words were chosen from the CELEX lexical database. Thirty-two of these were low-frequency words and 32 were high-frequency words. Half of the words in each frequency group had a low BOB consistency, and half had a high BOB consistency. All words
272
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
Table 5 Characteristics of the stimuli in Experiment 3 Variable
High frequency High BOB
Frequency per million Median frequency BOB consistency V1 C2 consistency V1 consistency V2 consistency No. of phonemes
Low frequency
Low BOB
99.4 93.1 .881 .536 .501 .732 5.56
101.9 61.5 .384 .516 .430 .593 5.44
High BOB 4.2 2.2 .849 .488 .473 .641 5.80
Low BOB 6.3 2.7 .386 .480 .497 .731 5.67
No pairwise comparison on any of the control consistency measures was significant (all ps > :20).
were monomorphemic, six letters in length, and stressed on the first syllable. The four groups of words were matched for V1 C2 consistency, V1 consistency, V2 consistency, number of intervocalic consonants, initial phoneme, and number of phonemes. BOB consistency was matched for the two groups of high-BOB-consistency words and the two groups of low-BOB-consistency words. Mean printed word frequency was matched using the CELEX norms for the two groups of high-frequency words and the two groups of low-frequency words. The characteristics of these words are presented in Table 5. The full list of words is presented in Appendix B. The procedure was the same as in the previous experiments, except that the participants were presented with only 64 words, and this experiment was run on a Power MacIntosh 6100 rather than the LC III. The same type of monitor was used to present stimuli. Each participant received a different random ordering of the stimuli. Results Naming latencies that exceeded 1200 ms were not included in the analyses. This affected approximately 0.7% of the data. For the analyses treating participants as a random factor ðF1 Þ, BOB consistency (high, low) and frequency (high, low) were analyzed as within-participants factors. For the analyses treating items as a random factor ðF2 Þ, the same variables were analyzed as between-items factors. The mean naming latencies and error rates by participants are presented in Fig. 3. In the analysis of naming latencies, the main effect of consistency was significant by participants, F1 ð1; 35Þ ¼ 30:43, p < :0001, MSe ¼ 280:32, and approached significance by items, F2 ð1; 60Þ ¼ 2:92, p ¼ :09, MSe ¼ 1386:56. High-consistency words were named faster ðM ¼ 524; SE ¼ 9:9Þ than low-consistency words ðM ¼ 540; SE ¼ 11:0Þ. As expected, the main effect of frequency was also significant, with high-frequency words named faster ðM ¼ 510; SE ¼ 8:1Þ than low-frequency words ðM ¼ 554; SE ¼ 11:4Þ, F1 ð1; 35Þ ¼ 91:52, p < :0001, MSe ¼ 728:80, F2 ð1; 60Þ ¼ 22:12, p < :001, MSe ¼ 1386:56. Notably, the
interaction between consistency and frequency was not significant, F s < 1.5 In the analysis of error rates, high-consistency words were observed to produce significantly fewer errors than low-consistency words, F1 ð1; 35Þ ¼ 8:77, p < :01, MSe ¼ :018, F2 ð1; 60Þ ¼ 8:09, p < :01, MSe ¼ :001. The effect of word frequency was also significant, with highfrequency words producing fewer errors than low-frequency words, F1 ð1; 35Þ ¼ 12:30, p < :001, MSe ¼ :014, F2 ð1; 60Þ ¼ 12:46, p < :01, MSe ¼ :001. Again, there was no indication of an interaction between consistency and frequency, F s < 1. Discussion The spelling–sound consistency of the BOB segment significantly influenced readersÕ ability to name disyllabic words regardless of their frequency. This finding is in agreement with Treiman et al.Õs (1995) and JaredÕs (1997, 2002) results for monosyllabic words. The previous reports of an interaction between frequency and consistency or regularity did not match high- and low-frequency words for the degree of consistency, or the frequencies of friends and enemies (e.g., Jared & Seidenberg, 1990; Kay & Bishop, 1987; Seidenberg et al., 1984), whereas JaredÕs (1997, 2002) studies and this study did. The spelling– sound consistency of the BOB, therefore, clearly has a ubiquitous effect on the dynamics of phonological processing. As with monosyllabic words, the consonants that follow the initial vowel in disyllabic words appear to help constrain the pronunciation of the vowel, and this happens regardless of the frequency of the word. 5 An items analysis of the naming latencies from Experiment 1 for the words that also appeared in Experiment 3 were included ðN ¼ 58Þ again revealed the same pattern. The interaction between the frequency and consistency was not significant, F < 1. The main effect of consistency was marginally significant, F ð1; 53Þ ¼ 3:65, p ¼ :062, MSe ¼ 1238:70, and the main effect of frequency was significant, F ð1; 53Þ ¼ 8:45, p < :01, MSe ¼ 1238:70. There were no significant effects for error rate.
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
273
Fig. 3. Mean naming latencies and error rates (in parentheses) for conditions in Experiment 3.
Experiment 4 In the two factorial experiments above, the spelling– sound consistency of the BOB was varied. This segment is relevant for the pronunciation of the initial vowel in these words and also the medial consonants. In the introduction, however, it was suggested that vowel inconsistency may be the cause for the effect of the number of syllables found by Jared and Seidenberg (1990, Expt. 3). The results of Experiment 1 supported the hypothesis, finding that V1 consistency and V2 consistency affected naming latencies and error rates in many of the analyses. In particular, V2 consistency appeared to play a very strong role in the dynamics of phonological processing. Replicating the effect of V2 consistency in a factorial experiment is important, as this would support the claim that the spelling– sound consistency of vowels is particularly influential in naming words. It would also provide further evidence that spelling–sound consistency in the later part of a disyllabic word can influence naming latencies. Method Participants Twenty-five undergraduate students at the University of Western Ontario, who had not participated in the previous experiments, participated in this experiment. Participants volunteered for the study as one method of earning credit in an introductory psychology class. All participants were native English speakers and had normal or corrected to normal vision. Materials and procedure Fifty disyllabic words chosen from the CELEX lexical database were included in this experiment. Twenty-
five of these words had a low V2 consistency and the other 25 words had a high V2 consistency. These two lists of words were matched for initial phoneme, number of phonemes, mean BOB consistency, V1 consistency, V1 C2 consistency, number of intervocalic consonants, and mean CELEX frequency. These word lists were also equated on the number of words containing complex vowel segments, and the length (in number of letters) of the V2 C4 segment. The words were equated on these characteristics because it is possible that complex vowel segments and longer V2 C4 segments are typically associated with second-syllable stress, whereas all of the target words are actually stressed on the first syllable. As in all of the factorial experiments presented, all words were monomorphemic, six letters in length, and were stressed on the first syllable. The characteristics of these words can be found in Table 6. The full list of words can be found in Appendix C. This experiment was run on a Power MacIntosh 6100/60. The procedure was the same as for the previous experiments, except that the participants were presented with only 50 words. Each participant received a different random ordering of the stimuli. Table 6 Characteristics of the stimuli in Experiment 4 Variable Frequency per million Median frequency BOB consistency V1 C2 consistency V1 consistency V2 consistency No. of phonemes
High V2
Low V2
10.3 2.8 .623 .494 .410 .844 5.56
8.9 3.4 .671 .540 .459 .283 5.60
No pairwise comparison on any of the control consistency measures was significant (all ps > :15).
274
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
Results Naming latencies that exceeded 1200 ms were not included in the analyses. This affected approximately 1.6% of the naming latencies. For the analysis treating participants as a random factor ðt1 Þ, V2 consistency (high, low) was analyzed as a within-participants factor. For the analysis treating items as a random factor ðt2 Þ, V2 consistency was analyzed as a between-items factor. In the analysis of naming latencies, the effect of spelling–sound consistency was significant by participants, t1 ð25Þ ¼ 4:18, p < :001, and approached significance by items in a one-tailed test, t2 ð48Þ ¼ 1:61, p ¼ :055. Highconsistency words were named faster ðM ¼ 561; SE ¼ 16:2Þ than low-consistency words ðM ¼ 579; SE ¼ 17:5Þ. In the analyses of errors, the effect of consistency was not significant, t1 ð25Þ ¼ 0:34, ns, t2 ð48Þ ¼ 0:56, ns. Highconsistency words produced 4.9% errors ðSE ¼ 1:22Þ and low-consistency words produced 3.7% errors ðSE ¼ :88Þ.6 Discussion The results of this study support those found in Experiment 1, where V2 consistency was found to contribute significantly to the prediction of naming latencies and errors in many of the regression analyses. In this experiment, however, V2 consistency significantly affected naming latencies but did not affect error rates. The size of the effect on naming latencies was numerically similar to the effect produced by the BOB manipulation in Experiment 2. The results suggest that the phonological representation of the second syllable of a disyllabic word is often computed before articulation of the word is initiated. If word naming typically begins when only the onset phonology is available to the reader (Kawamoto et al., 1998, 1999), then one would not expect the consistency of an orthographic segment in the latter part of a disyllabic word to have such a strong effect on time to initiate a pronunciation. This finding suggests that, at least for six-letter words, the entire orthographic representation of a word may activate the entire phonological representation of the word in parallel before a reader begins to name the word. General discussion The aim of this study was to extend our understanding of spelling–sound consistency effects to disyl6
An items analysis of the naming latencies from Experiment 1 for the words that also appeared in Experiment 4 were included ðN ¼ 46Þ also produced the same pattern. Because the four missing words were all from one group, we covaried out word frequency. The effect of consistency was marginally significant, F ð1; 43Þ ¼ 2:88, p ¼ :097. The effect of consistency was not significant for error rate.
labic words. The study examined whether the spelling– sound consistency of any simple or higher-order orthographic segment influences the naming of disyllabic words. Experiment 1 found that across a number of regression analyses, the consistency of vowel segments and the BOB segment were most often significant predictors of naming latencies and errors. Greater consistency led to faster naming latencies and fewer errors. Presumably, this effect is due to more efficient, or less noisy, phonological processing for more consistent stimuli. Experiment 2 established in a factorial experiment that BOB consistency has a significant impact on naming latencies and errors, whereas the consistency of the syllable defined V1 C2 does not. The finding that BOB consistency was a significant predictor of naming performance in hierarchical regression analyses after the consistency of simple segments had been included suggests that the BOB provides information about pronunciations beyond that provided by the simple segments of which it is composed, and that readers have learned to use this information. We suggested that the utility of the BOB in converting print to sound may be due to it constraining the possible pronunciations of the initial vowel segment, much as the word-body constrains the possible pronunciations of the vowel in monosyllabic words (Treiman et al., 1995). This hypothesis could be tested by calculating the mean consistency (H 2 statistic) of various segments of disyllabic words across the entire set of disyllabic words, as Treiman et al. (1995) did for monosyllabic words. Such calculations are beyond the scope of the current project, however. To calculate the mean consistency of the BOB segment, for example, would require first calculating the consistency of all BOB segments appearing in English disyllabic words, which probably number in the thousands. Here, we calculated the consistency of only the approximately 350 different BOB segments appearing in the 1000 experimental words. The mean consistency of the other segments would have to be calculated for comparison purposes, and this would also require that the entire set of each segment be identified. We have left this for a future research project. However, we can easily calculate the mean consistency of segments across just the words that we used in Experiment 1. For the V1 segment this was .42, for the V1 C2 this was .51 and for the BOB this was .60. These figures provide evidence that the BOB constrains the pronunciation of the first vowel and that it is more constraining than the V1 C2 . We decided to include the BOB segment in our consistency measures because it is a segment that is not restricted by a syllable boundary and there was some evidence from previous research that the BOB is relevant to naming (Taft, 1992). We were somewhat surprised by how well the consistency of this segment predicted naming performance. It now needs to be determined whether there is anything special about the BOB
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
segment, which must have an orthographically legal ending, or whether the consistency of a segment that includes the first vowel and all subsequent consonants before the next vowel is a better predictor of naming. Taft (1992) suggested that there are BOB nodes in the word recognition system. In this view BOB consistency should be the best predictor of disyllabic word naming. However, if we are correct in our explanation of why BOB consistency is predictive of disyllabic word naming, and if the additional consonants before the second vowel further constrain the pronunciation of the first vowel, then the consistency of the segment which includes the first vowel and all intervocalic consonants might be an even better predictor of naming performance. It would be more computationally efficient for a system not to have to parse the letter string based on orthotactic constraints as is necessary to identify BOB segments. Experiments 3 and 4 in the present study examined two other important issues regarding spelling–sound consistency effects in disyllabic word naming. Experiment 3 established the effect of BOB consistency amongst high-frequency disyllabic words, giving greater support to recent evidence indicating that spelling–sound consistency effects are not limited to low-frequency words (Jared, 1997, 2002; Treiman et al., 1995). Experiment 4 indicated that manipulations of the spelling–sound consistency of orthographic segments in the second syllable of disyllabic words (i.e., V2 ) also affect naming latencies. The influence of the consistency of the V2 segment, rather than the V2 C4 segment, may be a consequence of the consonant(s) following the V2 segment not often providing useful information to constrain the pronunciation of the vowel. This hypothesis could be tested by comparing the mean consistency of V2 and V2 C4 segments across the entire set of disyllabic words. Again, we have left this to future research. We did calculate the mean consistency of V2 and V2 C4 across the words used in Experiment 1. These were .63 and .74, respectively, suggesting that the C4 provides only weak constraint for the V2 . Interestingly, the mean consistency of the V2 was considerably higher than that of the V1 (.42). In Experiments 2–4, the results of the item analyses were not as strong as we would have liked. In the latency data, item analyses always approached significance ðp < :10Þ, but typically did not reach conventional levels of significance. However, because the factorial experiments produced results that were generally consistent with the regression analyses (which were conducted on item means), we are fairly confident that the results are reliable. There are several reasons why weak results in the item analyses may have been obtained. One possibility is that the effects are small and require larger stimulus sets than were used in the factorial experiments to detect statistically significant differences. Another possibility is that consistency effects for some words may have been attenuated because of activation of phono-
275
logical representations from semantics. Strain, Patterson, and Seidenberg (1995) observed a much smaller exception effect for low-frequency words that were high in imageability than for those that were low in imageability, suggesting that the naming of low-frequency words with atypical spelling–sound correspondences is facilitated by activation from semantics. In the present studies, the consistency effects for high-imageability words may have been attenuated, which in turn may have weakened the results of the item analyses. A third possibility is that the item analyses may be overly conservative. Raaijmakers, Schrijnemakers, and Gremmen (1999) claimed that when items are matched in studies (that is, they are not simply randomly chosen from the population of all possible items), the assumption of random sampling for the items ANOVA is violated. They argue that in such cases, there is a bias against finding a significant result. Relation to previous findings with disyllabic words Jared and Seidenberg (1990) These findings can now be compared with those in the published literature, and the consistency measures calculated for this study can be used to attempt to account for past data. The closest approximation to the type of investigation conducted here is the study by Jared and Seidenberg (1990). However, they used a rough measure of consistency that, in essence, compared words that had only one pronunciation for their constituent syllables (i.e., regular) with words that had more than one pronunciation for one of their syllables (i.e., regular-inconsistent words and exception words). The degree of consistency was not measured. The first experiment from Jared and Seidenberg (1990) is an ideal test case for the generality of BOB consistency effects. There were eight groups of 20 words each in the experiment. Four groups corresponded to the manipulations in the first syllable position and another four to the second-syllable position. For each syllable, there were regular-inconsistent words, exception words, and regular-consistent control words for both of these groups of words. We calculated the BOB consistency and V2 consistency for each of the words in these lists. The mean naming latencies for each group of words are presented in Table 7, along with the mean BOB consistency and V2 consistency. The BOB consistency statistics can be used to accurately predict the naming latencies across the eight conditions, r ¼ :80, tð7Þ ¼ 3:30, p < :01 (one-tailed). In addition, the V2 consistency statistics are also correlated with naming latencies, r ¼ :59, tð7Þ ¼ 1:81, p ¼ :055 (one-tailed). The words that took the longest to name, the second syllable exception words, had the lowest BOB consistency. This group of words also had the lowest V2 consistency. The words with the fastest naming latencies,
276
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
Table 7 Mean naming latencies (ms) and BOB consistency measures for conditions in Jared and Seidenberg (1990, Expt. 1) Condition
Latency
BOB
V2
First syllable Exception Regular inconsistent Exception control Regular inconsistent control
617 595 589 581
.348 .478 .663 .755
.723 .685 .686 .777
Second syllable Exception Regular inconsistent Exception control Regular inconsistent control
638 619 606 585
.321 .566 .436 .618
.363 .609 .518 .526
the regular-consistent controls for first syllable regularinconsistent words, had the highest BOB consistency and the highest V2 consistency. The only group of words whose mean naming latency is not well predicted is the second-syllable regular-inconsistent group, which does not have particularly low BOB consistency or V2 consistency, but took relatively long to name. The BOB consistency measures developed here are also able to account for the interaction of frequency and regularity in Jared and SeidenbergÕs (1990) Experiment 2. Although these BOB consistency measures could only be used for the disyllabic words in their study (about 2/3 of the stimuli), an analysis showed that the mean BOB consistency for high-frequency regular and exception words was similar, but that the mean BOB consistency for low-frequency regular words was substantially higher than that for low-frequency exception words. Therefore their frequency by regularity interaction could have arisen because of a confound with BOB consistency. These retrospective analyses highlight just how relevant the degree of consistency is for phonological processing of disyllabic words. Stress effects: Rastle and Coltheart (2000) There is one other recent study that has compared the naming latencies of different groups of disyllabic words. Rastle and Coltheart (2000, Expt. 1) compared naming latencies for high- and low-frequency words which were divided into ÔregularÕ words that were stressed on the first syllable, and ÔirregularÕ words that were stressed on the second syllable. This was intended as an evaluation of a simple rule that disyllabic words should be stressed on the first syllable. Using the CELEX database, we calculated that 83% of English disyllabic words have this stress pattern. The results, however, showed no significant stress effect in naming latencies for either high- or lowfrequency words. For both frequency groups, the difference between ÔregularÕ and ÔirregularÕ stressed words was only 3 ms. The regression analyses in Experiment 1 of this paper also indicate that stress pattern may have little
impact for correct naming latencies on disyllabic words, although stress does appear to influence naming errors. Rastle and Coltheart interpreted their finding as indicating that the simple stress rule was not correct and they went on to develop a more complex series of stress assignment rules. We examined whether the BOB consistency statistics developed for the present study could account for their finding. In fact, the BOB consistency of the two stress groups was quite similar. For low-frequency words the BOB consistencies are .611 and .516 for regular and irregular stressed words, respectively, tð43Þ ¼ :92, ns. For high-frequency words, the means were .670 and .510, respectively, tð40Þ ¼ 1:61, ns. No difference in naming latencies for irregular or regular stress groups would, therefore, be expected given such small and non-significant differences in BOB consistency. Rastle and Coltheart (2000) did, however, find a difference between regularly and irregularly stressed words in their Experiment 3 where stress regularity was determined by a complex set of rules. When the regularity of a wordÕs stress pattern was based on this new algorithm, an interaction between frequency and stress regularity was found. For low-frequency words, regularstress words ðM ¼ 515Þ were named faster than those with irregular stress ðM ¼ 543Þ. For high-frequency words, there was no difference between words with regular ðM ¼ 479Þ and irregular stress ðM ¼ 480Þ. We again examined whether the BOB consistency statistics developed for the present study could account for this interaction. For low-frequency words the irregularly stressed words had a significantly lower mean BOB consistency ðM ¼ :247Þ than the regularly stressed words ðM ¼ :493Þ, tð44Þ ¼ 2:78, p < :01. For high-frequency words the mean BOB consistencies were not significantly different (M ¼ :550 and M ¼ :680), tð10Þ ¼ :98, ns. The statistics developed for this study, therefore, can account for the interaction reported by Rastle and Coltheart (2000). Thus, an alternative of explanation of Rastle and ColtheartÕs results is that they reflect BOB consistency rather than stress regularity. Implications for theories of spelling to sound processing Dual route accounts Dual-route accounts of phonological processing suggest that there are two routes for producing the phonology of a word from its orthography. The assembly route uses a series of grapheme–phoneme conversion (GPC) rules in which the translation of a letter or letters to the corresponding phonology is based on the most frequent translation for the grapheme, regardless of the context of letters in which the grapheme appears. The GPC rules, therefore, always provide the ÔregularÕ pronunciation. The lexical route directly activates a wordÕs phonological form from its orthography. A single node corresponding to the wordÕs orthographic
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
input activates a node for the wordÕs phonological output. Although this model can simulate the basic finding that exception words take longer to name than regular words, it cannot simulate more fine-grained spelling– sound effects. In particular, contrary to human performance, the model produces an exception effect for words that have many friends (e.g., TALK) and does not produce a consistency effect for regular words with many enemies (Jared, 2002). It is difficult to imagine how the dual-route model could account for the present results as well. The effect of BOB consistency, like word-body consistency, suggests that readers learn spelling–sound relationships for orthographic segments larger than graphemes. Furthermore, the GPC rules are based on monosyllabic words and therefore provide stressed pronunciations. How, then, could the model adequately produce the correct pronunciation for unstressed syllables, of which there will always be one in a disyllabic word? This would mean, in essence, that every word will have at least one irregular spelling–sound mapping. To overcome this problem, Rastle and Coltheart (2000) suggested that there is an additional rule system that will determine where stress should be placed and which vowel should be Ôreduced.Õ In this process there is a considerable amount of checking and double-checking to determine if a word contains certain letter patterns. If a word contains a prefix letter pattern, for instance, an additional procedure is needed to check if the remainder of the word is, in fact, orthographically legal. If it is, then the pronunciation of the prefix is retrieved from the lexicon, and the GPC rules are applied to the remainder of the word, which is stressed. If there is no prefix letter pattern, then the system looks for a suffix letter pattern. If there is one, the same checking and double-checking must occur. However, the application of this algorithm would change some of the fundamental assumptions of the model. Previous accounts of the assembly route have described it as operating serially, one grapheme at a time, sending activation to the phoneme output system (Coltheart et al., 1993; Coltheart & Rastle, 1994). In this revised view, monomorphemic disyllabic words must be analyzed holistically by the assembly route once to look for letter patterns, and then again to apply the GPC rules. The initial process to identify certain letter patterns appears to weaken the serial nature of processing in this route.Furthermore, this extra process added to the non-lexical route would mean that most words would be processed by the lexical route, because that route is assumed to be faster than the assembly route. As a consequence, there should rarely be a regularity or consistency effect when naming disyllabic words, particularly for high-frequency words. The present study indicates that such effects do occur, even with high-frequency words. The serial nature of the modelÕs assembly route would also make it difficult to produce any effect
277
of V2 consistency, since the lexical route would often have activated the whole wordÕs phonology well before the assembly route would send activation from the second vowel grapheme. How the model could overcome these challenges and maintain the serial nature of the non-lexical route is not easy to imagine. Connectionist accounts Connectionist theories, in principle, may have an easier time accounting for the present data. These models learn the statistical relationships between input and output mappings (Plaut et al., 1996; Seidenberg & McClelland, 1989; Zorzi et al., 1998), and are therefore well suited to account for consistency effects. Knowledge of spelling– sound correspondences is encoded by the weights on connections between orthographic and hidden units and between these hidden units and phonological units. The same set of units and connections is used for all words. Weights on connections are modified slightly with each exposure to a word. Words sharing an orthographic segment (e.g., word body) that maps to the same pronunciation (MUST, DUST, JUST) will have a similar effect on the weights; therefore exposure to one word improves performance on the other. Words that are orthographically similar but phonologically dissimilar push the weights in competing directions, so that exposure to one of the words (e.g., TREAT) has a negative impact on performance on the other (e.g., GREAT). Over time, then, the net effect of training on a large variety of words is poorer performance on inconsistent words compared to entirely consistent words. Phonological representations will stabilize particularly slowly for inconsistent words that have a letter pattern that is typically pronounced another way than is correct for that word. The Plaut et al. (1996) model simulates consistency effects from studies with monosyllabic words quite well (Jared, 1997, 2002), as does the model described by Zorzi et al. (1998). The principles of this type of model can be applied to the case of disyllabic word naming. First, the hidden units mediating between orthographic and phonological units allow the system to detect orthographic subpatterns that are predictive of pronunciations. In the case of monosyllabic words, the model is sensitive to consistency at the level of the word body because the coda constrains the pronunciation of the vowel. The effect of BOB consistency for disyllabic words may also occur because the consonants following the first vowel reliably constrain its pronunciation, and if so, one might expect that effects of the consistency of the BOB would be produced by an extended model. Conversely, such a model would not be expected to be sensitive to the consistency of higher-order segments that do not provide information beyond the simple segments of which they are composed, or in other words, that do not reliably constrain pronunciations. Furthermore, although the model should be sensitive to the consistency
278
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
of all simple segments, because the consistency of consonants is typically high, whereas the consistency of vowels is much more variable, differences in performance on words would be expected to be due more to differences in settling times for vowel pronunciations than for consonant pronunciations. This model should also be able to accommodate findings that inconsistencies in the second syllable (e.g., V2 consistency) have strong effects. Because processing in the connectionist models is parallel, an inconsistency anywhere in a word will affect the degree of error in the phonological representation, or the time it takes to clean up the phonological representation, and delay articulation. More importantly, because of parallel processing, this type of model does not ÔparseÕ words, or unavoidably process parts of words independently from other parts. This means two things. First, large orthographic segments can affect processing regardless of syllable structure or their position in a word. Second, this type of model should also be able to produce multiple effects from the same segment, as seen with the BOB and V1 segment in Experiment 1. The non-linear nature of these models allows for activation from the correspondences of a single segment (e.g., V1 ) and higher-order correspondences that depend on activation from larger complex segments (e.g., BOB). In principle, then, connectionist models are well suited to account for the data collected in this study. The main difficulty for this view, however, is determining the orthographic and phonological representations for disyllabic words. Plaut et al. (1996) employed a scheme that identified segments as being the onset, vowel, or coda of a monosyllabic word. It is doubtful that this approach can be directly extended to represent disyllabic words, however, because in some cases the same BOB segment (e.g., AND) can include both the coda of the initial syllable and the onset of the second syllable (e.g., TANDEM), or just the coda of the first syllable (e.g., BANDSTAND). Where or how these segments would be encoded in the orthographic representation in a connectionist system, without predisposing the system to identify the segments, remains a question. Other approaches could be employed in which the medial letters are treated as a segment on their own (i.e., onset/ vowel/ middle consonants/ vowel/ coda). This would eliminate the built-in syllabic parse that would result from a simple extension of the Plaut et al. (1996) representations. Fixed-slot representation would also overcome the parsing problem, but in that case what is learned about the letter A when it appears in the second position in a word may not extend to the letter A when it appears in the third position in a word. Therefore, although connectionist models are well suited to account for the data because of their intrinsic learning of the statistical relationship between spelling and sound, further computational work must be done to produce a working model that will generate the effects described in the present study.
Further questions These experiments have provided some answers to several key questions regarding disyllabic word naming, but there are still many unanswered questions that will require further research. For instance, is BOB consistency still a good predictor of naming latency when six-letter words with more than two syllables, such as BANANA, are included? These words have fewer consonants between the vowels to help constrain the vowel pronunciations. Does the consistency of both V2 and V3 segments in these words affect naming latencies? In the present study, the impact of the consistency of the syllable body was examined ðV1 C2 Þ but perhaps the consistency of entire syllables influences word naming. Another possibility, mentioned above, is that the consistency of a segment containing all consonants following the first vowel is a better predictor than BOB consistency. In the present study, all words were short and likely processed in one fixation. Further work needs to examine spelling–sound consistency effects in longer words. A question is whether the consistency of vowels towards the ends of longer words (i.e., words requiring several fixations) influences naming times. Perhaps they have less influence than V2 consistency here because processing of the initial part of the word largely determines the wordÕs pronunciation. Another issue concerns how syllabic stress is represented and processed. We have only briefly touched on this issue here. As well, the role of morphology in naming and how it affects the role of spelling–sound consistency remain to be examined. Thus, although the present study has begun to shed light on how printed polysyllabic words are converted to their phonological form, many questions remain.
Conclusions The present study has provided evidence for spelling– sound consistency effects in disyllabic words. The consistencies of the BOB and V2 orthographic segments influenced the naming of disyllabic words in Experiment 1, and this was replicated in factorial experiments. The present study also indicates that spelling–sound consistency affects the computation of phonology for high-frequency words. These results suggest that spelling–sound consistency is a fundamental characteristic of phonological processing, affecting the naming of all types of words. The effect of the consistency of larger segments, such as the BOB, appears to arise because the consonant context helps to determine the pronunciation of vowel. In addition, these experiments provide some evidence to suggest that readers do not compute pronunciations syllable by syllable. There instead appears to be simultaneous mapping to phonology of smaller orthographic segments and larger orthographic segments that contain the smaller
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
ones. These findings are very important for the development of word recognition theories and computational models of phonological processing, indicating that these models must be sensitive to the statistical nature of the particular languageÕs spelling to sound mappings for a variety of word segments. Acknowledgments The research reported here was based on a doctoral dissertation completed by D. Chateau under the supervision of D. Jared. Portions of the research were presented at the 41st Annual meeting of the Psychonomic Society, New Orleans, November, 2000. The research was supported by a grant from the Natural Sciences and Engineering Research Council of Canada to D. Jared (OGP01 53380). We thank Rebecca Treiman, Yasushi Hino, and Marco Zorzi for helpful comments on this research.
Appendix B. (continued) High-frequency words High-BOB consistency Dinner Doctor Finger Former Happen Market Matter Middle Pardon Sector Seldom Simple Terror Winter
High-BOB consistency
Low-BOB3 consistency
High-V1 C2 consistency
Low-V1 C2 consistency
High-V2 consistency
Beaker Picnic Common Dangle Fiscal Foster Gargle Gather Hazard Hither Hostel Corset Lather Mangle Margin Mildew Motion Parcel Censor Ransom Rescue Roster Spiral Staple Tether
Beacon Pillar Cosmos Dazzle Filter Format Garble Gadget Hamper Hiccup Hockey Corpus Ladder Manage Marvel Mingle Motive Pardon Cellar Random Reckon Rocket Spider Stable Tendon
Bamboo Bumble Burlap Cashew Cattle Column Damsel Doctor Fidget Fumble Goblin Huddle Jabber Juggle Kettle Mussel Pebble Gambit Ramble Ribbon Talent Tender Ticket Trifle Tumble
Banner Bunker Burrow Canyon Cannon Combat Damage Donkey Fiddle Funnel Gothic Hunger Jagged Jungle Kennel Musket Pelvis Gasket Rankle Riddle Tangle Tennis Tinsel Tribal Tunnel
Bovine Burden Cancel Cellar Cherry Corset Decade Forest Funnel Fabric Hermit Hostel Liquor Magnet Motion Muster Pamper Potent Rustle Tandem Tartan Vermin Volley Wallet Zenith
Appendix B. Stimuli used in Experiment 3
High-BOB consistency Coffee Corner
Low-frequency words
Low-BOB consistency
High-BOB consistency
Low-BOB consistency
Danger Damage Figure Fellow Heaven Modest Master Modern Palace Season Second Social Target Window
Dazzle Dimple Filter Format Hamper Marvel Mingle Motive Pillar Spider Stable Sandal Tendon Wombat
Dangle Dither Fiscal Foster Hazard Margin Mildew Mangle Picnic Spiral Staple Savage Tether Wobble
Appendix C. Stimuli used in Experiment 4
Appendix A. Stimuli used in Experiment 2
High-frequency words
279
Low-V2 consistency Bishop Burlap Cannon Census Chilli Cortex Damage Famine Fathom Fungus Herpes Horror Liquid Mascot Menace Mutton Pastor Proton Rumpus Talcum Tarmac Vertex Volume Walrus Zircon
References
Low-frequency words
Low-BOB consistency
High-BOB consistency
Low-BOB consistency
Common Couple
Cosmos Corpus
Column Corset
Andrews, S. (1982). Phonological recoding: Is the regularity effect consistent? Memory & Cognition, 10, 565–575. Baayen, R. H., Piepenbrock, R., & Gulikers, L. (1995). The CELEX Lexical Database (Release 2) [CD-ROM]. Philadelphia, PA: Linguistic Data Consortium, University of Pennsylvania [Distributor].
280
D. Chateau, D. Jared / Journal of Memory and Language 48 (2003) 255–280
Brown, P., Lupker, S. J., & Colombo, L. (1994). Interacting sources of information in word naming: A study of individual differences. Journal of Experimental Psychology: Human Perception and Performance, 18, 987–1003. Carreiras, M., Alvarez, C. J., & de Vega, M. (1993). Syllable frequency and visual word recognition in Spanish. Journal of Memory and Language, 32, 766–780. Cohen, J. D., MacWhinney, B., Flatt, M., & Provost, J. (1993). PsyScope: An interactive graphic system for designing and controlling experiments in the psychology laboratory using Macintosh computers. Behavior Research Methods, Instruments, and Computers, 25, 257–271. Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993). Models of reading aloud: Dual-route and parallel-distributed-processing approaches. Psychological Review, 100, 589–608. Coltheart, M., & Rastle, K. (1994). Serial processing in reading aloud: Evidence for dual-route models of reading. Journal of Experimental Psychology: Human Perception and Performance, 20, 1197–1211. Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J. (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108, 204–256. Ferrand, L., Segui, J., & Humphreys, G. W. (1997). The syllableÕs role in word naming. Memory & Cognition, 25, 458–470. Glushko, R. J. (1979). The organization and activation of orthographic knowledge in reading aloud. Journal of Experimental Psychology, 5, 674–691. Jared, D. (1997). Spelling–sound consistency affects the naming of high-frequency words. Journal of Memory and Language, 36, 505–529. Jared, D. (2002). Spelling–sound consistency and regularity effects in word naming. Journal of Memory and Language, 46, 723–750. Jared, D., Levy, B. A., & Rayner, K. (1999). The role of phonology in the activation of word meanings during reading: Evidence from proofreading and eye movements. Journal of Experimental Psychology: General, 128, 219–264. Jared, D., McRae, K., & Seidenberg, M. S. (1990). The basis of consistency effects in word naming. Journal of Memory and Language, 29, 687–715. Jared, D., & Seidenberg, M. S. (1990). Naming multisyllabic words. Journal of Experimental Psychology: Human Perception and Performance, 16, 92–105. Kawamoto, A. H., Kello, C. T., Higareda, I., & Vu, J. V. Q. (1999). Parallel processing and initial phoneme criterion in naming words: Evidence from frequency effects on onset and rime duration. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25, 362–381. Kawamoto, A. H., Kello, C. T., Jones, R., & Bame, K. (1998). Initial phoneme versus whole-word criterion to initiate pronunciation: Evidence based on response latency and initial phoneme duration. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24, 862–885. Kay, J., & Bishop, D. (1987). Anatomical differences between nose, palm, and foot, or, the body in question: Further dissection of the processes of sub-lexical spelling–sound translation. In M. Coltheart (Ed.), Attention and Performance XII: The psychology of reading (pp. 449–469). Hillsdale, NJ: Erlbaum.
Kessler, B., Treiman, R., & Mullennix, J. (2002). Phonetic biases in voice key response time measurements. Journal of Memory and Language, 47, 145–171. Levelt, W. J., & Wheeldon, L. (1994). Do speakers have access to a mental syllabary? Cognition, 50, 239–269. Lorch, R. F., Jr., & Myers, J. L. (1990). Regression analyses of repeated measures data in cognitive research. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 149–157. Perea, M., & Carreiras, M. (1998). Effects of syllable frequency and syllable neighborhood frequency in visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 24, 134–144. Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56–115. Raaijmakers, J. G. W., Schrijnemakers, J. M. C., & Gremmen, F. (1999). How to deal with ‘‘the language-as-fixed-effect fallacy’’: Common misconceptions and alternative solutions. Journal of Memory and Language, 41, 416–426. Rastle, K., & Coltheart, M. (1999). Lexical and nonlexical phonological priming in reading aloud. Journal of Experimental Psychology: Human Perception and Performance, 25, 461–481. Rastle, K., & Coltheart, M. (2000). Lexical and nonlexical print-to-sound translation of disyllabic words and nonwords. Journal of Memory and Language, 42, 342–364. Schiller, N. O. (2000). Single word production in English: The role of subsyllabic units during phonological encoding. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 512–528. Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523–568. Seidenberg, M. S., Waters, G. S., Barnes, M. A., & Tanenhaus, M. K. (1984). When does irregular spelling or pronunciation influence word recognition? Journal of Verbal Learning and Verbal Behavior, 23, 383–404. Stanhope, N., & Parkin, A. J. (1987). Further explorations of the consistency effect in word and nonword pronunciation. Memory & Cognition, 15, 169–179. Strain, E., Patterson, K., & Seidenberg, M. S. (1995). Semantic effects in single word naming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1140– 1154. Taft, M. (1979). Lexical access via an orthographic code: The basic orthographic syllabic structure (BOSS). Journal of Verbal Learning and Verbal Behavior, 18, 21–39. Taft, M. (1992). The body of the BOSS: Subsyllabic units in the lexical processing of polysyllabic words. Journal of Experimental Psychology: Human Perception and Performance, 18, 1004–1014. Treiman, R., Mullenix, J., Bijeljac-Babic, R., & RichmondWelty, E. D. (1995). The special role of rimes in the description, use, and acquisition of English orthography. Journal of Experimental Psychology: General, 124, 107–136. Zorzi, M., Houghton, G., & Butterworth, B. (1998). Two routes or one in reading aloud? A connectionist dual-process model? Journal of Experimental Psychology: Human Perception and Performance, 24, 1131–1161.