Wurm (1997)

JOURNAL OF MEMORY AND LANGUAGE ARTICLE NO.
37, 438461 (1997)
ML972524
Auditory Processing of Prexed English Words Is Both Continuous and Decompositional

Lee H. Wurm
State University of New York at Stony Brook Two experiments compared continuous and discontinuous models of word recognition. Participants heard prexed words whose full-form and root uniqueness points (UPs) differed, in either a gating or lexical decision paradigm. Identication points and reaction times were analyzed using multiple regression. Full-form UPs predicted performance better than root UPs did. Fullform frequency measures had reliable facilitative relationships with performance while root frequency measures were not consistently signicant. Prex frequency had a reliable, inhibitory effect. Judged prexedness, semantic transparency, and prex likelihood were related to performance, alone or in interaction. The results provide evidence for both kinds of word recognition procedures. A model is proposed with two parallel recognition routines: a whole-word routine and a decompositional routine that considers only unbound roots that can combine with the prex in question. A preliminary rating study provides stimulus values on several dimensions and can be used as a database by other researchers. 1997 Academic Press
Spoken-word recognition is complex, and particularly little is known about cases where words are composed of more than one morpheme. There are two ways in which lexical
This research was supported by Augmentation Award for Science and Engineering Research Training Grant 93NL174 from the Air Force Ofce of Scientic Research, Grant R01 MH51663-03 (to Arthur G. Samuel) from the National Institute of Mental Health, and National Research Service Award 1 F32 MH11721-01, also from NIMH. This article is based on the authors doctoral dissertation (State University of New York at Stony Brook). A portion of this work was presented at the 37th annual meeting of the Psychonomic Society (Chicago, IL, October 31November 3, 1996). I thank Arthur Samuel, Mark Aronoff, Susan Brennan, Richard Gerrig, John Robinson, and Cynthia Connine, all of whom made comments on an earlier version of this paper. I also thank Donna Kat for her invaluable help with computer programming. Arthur Aron, David Cross, and the Statistics Lunch group at Stony Brook provided a great deal of help with data analysis, and Harald Baayen helped in the computation of morpheme frequencies. Annmarie Cano and Douglas Vakoch provided emotional support and help with many of the more tedious aspects of this project. Marcus Taft, Dennis Norris, and an anonymous reviewer made very helpful comments on an earlier version of this paper. Correspondence concerning this article should be addressed to Lee H. Wurm, Department of Psychology, State University of New York at Binghamton, Binghamton, NY 13902-6000. E-mail: lwurm@binghamton.edu. 0749-596X/97 $25.00
Copyright 1997 by Academic Press All rights of reproduction in any form reserved.
access of such morphologically complex words might occur. First, complex words might be recognized without decomposing them into their constituent morphemes. Lexical entries would correspond to whole words. Alternatively, complex words might be decomposed, with individual morphemes represented lexically. Each of these access strategies corresponds to one of the main schools of thought among psychologists. I will call one of these the discontinuous (or decomposition) approach. In models of this type, complex words must be decomposed into root and afxes prior to lexical access. Lexical access can proceed only via the root because complex words do not have lexical entries. Although there have been several discontinuous models suggested (e.g., Cutler, Hawkins, & Gilligan, 1985; Jarvella & Meijers, 1983; MacKay, 1978; Morton, 1969; 1979), this approach has been chiey associated with Taft and his colleagues (1979a; 1981; 1985; Taft & Forster, 1975; Taft, Hambly, & Kinoshita, 1986). The main nding from the work of Taft and his colleagues is that nonword decisions take longer when the nonwords carry a prex than when they do not, and they take longer still
438
AUDITORY PROCESSING OF PREFIXED WORDS
439
when the stems of the prexed nonwords are genuine stems in English. Whether or not the stems of such nonwords are genuine English stems makes no difference when the nonwords do not begin with genuine prexes. These stems are not even recognizable as stems because there is no prex to strip off. Bergman, Hudson, and Eling (1988) also found support for the afx-stripping model (Taft, 1981; 1985; Taft & Forster, 1975). Taft (1994) addressed some of the more persistent criticisms of his earlier work, and the results completely supported the earlier conclusions in favor of an afx-stripping approach. Still, Taft (1994) modied his position somewhat, writing that an interactive-activation model provided the best explanation of morphological processing. In Tafts new model, prexes are independent activation units, separate from roots (i.e., prexed words are stored in decomposed form), but there is no prex-stripping per se. In this type of scheme the equivalent of a prex-stripping procedure is a part of the access process. The opposing general approach is to access morphologically complex words on a strictly continuous, or left-to-right, basis. I will also use the term full-listing to refer to models of this type, because on this type of account, all words are listed in the lexicon. Words are accessed as complete units, whether or not they contain afxes. The lexicon in this case would contain separate entries for the words cover, uncover, covering, and so on, resulting in considerable redundancy (e.g., verbs in Finnish and Georgian can take thousands of distinct surface forms that are essentially the same vocabulary item although they differ by inection [Anderson, 1988]). This redundancy is the reason people rst proposed a decompositional approach; the storage required by a morphemic lexicon is much smaller than that required if all items are to be listed. However, Bybee (1988; see also Sandra, 1994) believes that the emphasis on storage efciency is misguided, given the huge capacity of the human brain. She also notes that linguists are relaxing their insistence on maximally efcient storage princi-
ples, as the extent of the idiosyncracies of all languages is coming to light. The Cohort model (Marslen-Wilson & Welsh, 1978; Marslen-Wilson, 1984), although not specically designed to address issues relating to morphological complexity, is a continuous processing model. Network models, too, have most often aligned themselves with this view (e.g., McClelland & Elman, 1986; Norris, 1994). Other arguments for a continuous processing model have been made by Henderson, Wallis, and Knight (1984), Rubin, Becker, and Freeman (1979), and Tyler, Marslen-Wilson, Rentoul, and Hanney (1988). An important third class of model is driven by stress (Cutler, 1976; Cutler & Norris, 1988; Grosjean & Gee, 1987). According to the metrical segmentation strategy (MSS) of Cutler and Norris (1988), a strong syllable initiates a lexical search. The unstressed syllables immediately surrounding the strong syllable are checked by an acoustic pattern-matching routine, to see if they can combine with the stressed syllable. While not involving afxstripping per se, a stress-driven model is an example of a discontinuous model; words are not processed in a strict left-to-right manner. Furthermore, stress-driven and afx-stripping models make similar predictions in many cases. Afxes will play a peripheral role in lexical access because they are typically unstressed. In fact, stressed prexes could cause processing difculties by triggering a lexical look-up on the prex (or the prex plus the initial phoneme[s] of the root) rather than on the root itself. There are also mixed models that incorporate both a continuous and a discontinuous approach. For example, Bergman et al. (1988) suggested a possibility based on the race model of Cutler and Norris (1979): the speech processor might attempt to access morphologically complex words both as full-forms and as analyzed separate parts. Data will be presented below that are suggestive of this kind of model; other variations on this idea have also been proposed (e.g., Anshen & Aronoff, 1981; Caramazza, Laudanna, & Romani,
440
LEE H. WURM
1988; Frauenfelder & Schreuder, 1992; Schriefers, Friederici, & Graetz, 1992). In recent years, researchers have begun to examine the possible effects of linguistic and computational variables on complex word processing. Using the crossmodal priming technique, Marslen-Wilson, Tyler, Waksler, and Older (1994) found that derivationally sufxed words primed and were primed by their roots, but only if they were semantically transparent (i.e., only if there was an obvious semantic relationship between the two, as in government and govern). Marslen-Wilson et al. (1994) concluded that semantics determines which words are morphologically related to each other (as well as which words are complex). This was the rst study to show the importance of semantic transparency, a variable that has interested linguists for some time. Laudanna, Burani, and Cermele (1994) found that a pair of little-examined variables affect lexical decision performance for visually presented Italian nonwords: the number of word-types beginning with a given prex and the success rate of prex-stripping for that prex (i.e., the proportion of encountered words beginning with that letter string that are in fact truly prexeda quantity I will refer to as the prex likelihood). Although the two variables are highly correlated, Laudanna et al. (1994) found that prex likelihood was the more important one. They concluded that a word beginning with a prex that has a high prex likelihood is likely to be stored and accessed in decomposed form. The movement of some researchers toward the theoretical middle ground in recent years is a major advance. Researchers are recognizing the possibility that some words may be decomposed while others may not. In accord with this, researchers have realized that in order to specify the conditions under which various lexical access strategies may be used, they must be concerned with concepts like semantic transparency. These were largely ignored in early research efforts. The present study used prexed English words in an attempt to clarify recent ndings
concerning prex likelihood and semantic transparency. The predicted identication points (IPs) for continuous and discontinuous processing models were directly tested. The effects of prex likelihood, degree of prexedness, and semantic transparency were examined for their individual contributions to the recognition process, and importantly, for their joint or interactive contributions. This has not previously been done and, as we will see, there are interesting and theoretically important interactions between these variables (and others). Furthermore, in the present study these independent variables were treated in ways that are more natural and powerful than in previous research. The main goal of the preliminary rating study was to provide information to be used in interpreting experimental data from the subsequent experiments. For each stimulus, values were obtained on several different dimensions of interest, including prex likelihood, judged prexedness, semantic transparency, root morpheme frequency, prex frequency, and neighborhood size. This study also provides a database for other researchers to use; these values have not previously been available, and some of the computations are timeconsuming. PRELIMINARY RATING STUDY This study consisted of two parts, the general procedure of which was the same; participants were presented with auditory stimuli and gave a rating of some kind. In Part I, participants rated the stimuli on semantic transparency, and in Part II, they rated the stimuli on prexedness. The semantic transparency of an item might inuence the perceptual strategies used to process it. Words like rebuild have an obvious compositional meaning (re plus build) that is lacking from words like prepare. However, it is only within the very recent past that researchers have recognized this, and the few researchers to study this variable have articially dichotomized it. One of the purposes of this rating study was to get continuous semantic-transparency ratings for stimu-
441
lus items, since no published norms exist. These ratings were used as predictors of performance in the two main experiments of this study. A second purpose was to get ratings of the prexedness of the items. Which words are prexed? is just as much a psychological question as a linguistic one. Noncircular denitions in this area are uncommon, and linguists have not been able to reach a consensus about exactly what is or is not an afx. However, even if linguists were to reach such a consensus, it is not clear that their denition would have psychological relevance. Method Participants. Twenty-one students from the Department of Psychology subject pool at the State University of New York at Stony Brook provided prexedness ratings. Thirty students from the same subject pool provided semantic-transparency ratings. All participants were native speakers of English who received course credit for their participation. Materials. Seventy-two prexed and pseudoprexed words were selected from a 150,000-word computerized phonetic dictionary (Moby Pronunciator 1.01, 1989) that met the following criteria: (1) carried a root that is a free-standing monomorphemic English word or homophonous with one; (2) began with a string of phonemes that constitutes an English prex; (3) varied (as a set) as much as possible along a continuum from high to low semantic transparency; and (4) had fullform and root uniqueness points (UPs) that differed by at least one phoneme (the UP is the point in the acoustic signal where the word in question diverges from all other words in a languagesee Marslen-Wilson, 1984; Marslen-Wilson & Welsh, 1978). In determining UPs, I searched the same computerized phonetic dictionary (Moby Pronunciator 1.01, 1989) used for initial stimulus selection. I excluded any sufxed forms that were related (by inection or derivation) to the word in question (see, for example, MarslenWilson, 1984; Tyler et al., 1988). Thus, the
UP of distaste is the /a V /, in spite of the existence of the related word distasteful. Full-forms were digitized at a sampling rate of 10 KHz, low-pass ltered at 4.8 KHz, and stored in disk les. Stimuli were spoken by a female native English speaker who was not familiar with the purpose of the study. Roots were digitally spliced away from the fullforms using visual and auditory inspection, and the resulting roots sounded quite natural. For the prexedness ratings, two stimulus lists were created. A randomly-determined half of the full-forms were assigned to list A and half were assigned to list B. The root corresponding to a given full-form was assigned to the opposite list. For example, if rebuild was in list A, build was in list B. For the semantic transparency ratings, the same stimuli were used, but they were arranged in pairs that consisted of a full-form and its root (e.g., rebuild and build). Procedure. Participants were tested in groups of one to four. Stimuli were presented over headphones in a sound-attenuating chamber. For the prexedness ratings, participants were randomly assigned to hear either list A or B. After each word, participants had 2500 ms to give a rating. A Likert scale with anchor points Not at all prexed (1) and Very prexed (8) was used. Participants were free to decide for themselves what prexed means; no example or further instructions were given. For each list, words were presented in a different random order. For the semantic-transparency ratings, on each trial, participants heard a full-form and its corresponding root and were asked to rate . . . how related in meaning the two words are on an 8-point Likert scale. Anchor points were labeled Not at all related (1) and Very related (8). On half of the trials, the full-form was presented rst, and on half the root was presented rst. Participants had 4000 ms to make their rating. Item pairs were presented in a random order. Results and Discussion Median judged prexedness values are listed for each stimulus in Appendix A. The
442
LEE H. WURM
ratings distribution had moderate negative skew, so the recommended reect and square-root transformation was performed (Tabachnick & Fidell, 1989). Thus, items that participants rated highly prexed had lower transformed prexedness scores. This transformed measure is the one that will be used in all analyses reported in this study. Median semantic-transparency values are also listed for each stimulus in Appendix A. Because the distribution of transparency scores was positively skewed, the square-root of the transparency scores was used in all analyses reported in this study. Calculation of Other Regressor Variables Several other variables were used to predict participants performance in Experiments 1 and 2. Prex likelihoods for each prex were calculated in the following way (I used pronunciation rather than spelling in selecting words). I rst identied for each prex all of the words from Websters Third New International Dictionary of the English Language (1993) that were, in fact, truly prexed. A word was considered truly prexed if the prex contributes to the meaning and syntax of the full-form. This denition implies three criteria: First, when the prex is removed, what remains must be a freestanding word. Second, the meaning of the full-form must be reasonably transparent. Finally, the semantic relationship should be fairly constant across combinations; for example, truly prexed words beginning with re- will mean roughly to do (whatever the stem means) again (M. Aronoff claried these issues for me in a personal communication, November 30, 1994). Note that words carrying bound roots are excluded from the above denition, even though such words can subjectively appear to be prexed (e.g., supersede, which is not counted as truly prexed by the above criteria, had a high prexedness rating [see Appendix A]). In some cases, the classication of a root as either bound or unbound is somewhat ambiguous, but the exclusion of words with bound roots from the calculation of prex likelihood is justied along several lines. First,
Schreuder and Baayen (1994) explained the requirement of reasonably high semantictransparency, which is almost always missing from words carrying bound roots, this way: The only way in which the language learner can discover that a certain string actually is a stem (free or bound) is when that string occurs in at least one semantically fully transparent [italics added] combination (p. 360). Even in the case of very common bound roots that combine with several different prexes (e.g., -mit or -vent), any reliable covariance between form and meaning (the importance of which is also discussed by van Orden, Pennington, & Stone, 1990) is likely to be noticed only by experts in linguistics. In addition, Henderson (1985) asserted that the morphological complexity of words with bound roots is synchronically meaningless; their afxes are fossils and not psychologically relevant at all. In short, although they appear to be prexed, they are not (the data of MarslenWilson et al. [1994], which I have already discussed, agree with this conclusion). I computed prex likelihood, then, as a ratio: the numerator was the summed frequency era, 1982) of the truly (from Francis and Kuc prexed words beginning with a given phonetic string, and the denominator was the summed frequency of all words beginning with that string in which removal of the putative prex leaves a pronounceable syllable or syllables. The decision to consider syllabication is also consistent with the criteria of Schreuder and Baayen (1994). For example, although coat begins with co-, this word was not considered a prex-stripping failure because the remainder of the word (simply the phoneme /t/ in this case) does not constitute a syllable. The notion behind prex likelihood is that decomposing words may make lexical access less efcient, if the majority of words beginning with a given prex are not in fact prexed words. Values for the prexes used in this study are listed in Appendix B and ranged from the theoretical minimum of .000 (hyper-) to the theoretical maximum of 1.000 (twi-). To illustrate what these num-
443
bers mean, consider the prex dis-, which has a prex likelihood of .092. This means that 9.2% of the time a reader or listener encounters a word beginning with dis, that word is prexed. At the upper extreme, all instances of words beginning with twi- (phonetically, /twai/) are truly prexed (words such as twice do not count as failures because removing the prex leaves the nonsyllabiable phonetic segment /s/). With few exceptions, these prex-likelihood values are quite low (M .19). This value is quite comparable to the .17 reported by Schreuder and Baayen (1994) for the seven English prexes they examined, using the CELEX database (Baayen et al., 1993; Burnage, 1990). On the face of it, prex likelihoods like these would suggest that a decomposition strategy has such a low payoff that it would make lexical access less efcient, which is what those authors concluded. As will become clear, their conclusion may have been premature. The research of Laudanna et al. (1994), in which the notion of prex likelihood (what they called the success rate) was initially explored, was discussed earlier. Although they concluded that prex likelihood was more important, they also found that the number of words beginning with a given prex was related to performance. This is essentially an unweighted measure of the density of each words lexical neighborhood. Lexical neighborhood is a metaphor for understanding how words are organized or stored in memory. For example, each of the four words in English that begin with the phonemic sequence /twai/ have only the other three as neighbors in lexical space. They reside in a sparsely populated neighborhood. Some neighborhoods have hundreds or thousands of inhabitants (e.g., over 500 words begin with the phonemic sequence pro-). The denominator of the prex-likelihood ratio is a frequency-weighted measure of neighborhood density. Regardless of the particular value of the prex likelihood, a large denominator means that a word resides in a dense lexical neighborhood, and a small de-
nominator means a sparse lexical neighborhood. Recognition is expected to be easier if the surrounding lexical neighborhood is sparsely populated, because items would have relatively little competition from neighbors (e.g., Goldinger, Luce, & Pisoni, 1989; Luce, 1987; Luce, Pisoni, & Goldinger, 1990). Neighborhood sizes for each prex are listed in Appendix B. The neighborhood size of pro-, for example, is the sum of the frequencies of the more than 500 words beginning with that sequence. Several different frequency measures were also needed. Individual full-form and root token frequencies were taken directly from era (1982). The correspondFrancis and Kuc ing family frequencies were calculated by era summing values from Francis and Kuc (1982) for each word and its morphological relatives (family frequencies were included because a consensus as to which frequency measure is more important, token or family, has not been reached). A continuous model predicts that root frequency will have no role in word recognition, and a decompositional model predicts that root frequency will explain a large proportion of the variance (Taft [1979b] found that both full-form and root frequencies inuence performance). Morpheme frequencies for each root and each prex had to be calculated in a different way. The rst step in calculating root frequencies was a non-position-specic string search for each root in the Birmingham/Cobuild corpus (18 million tokens) of the CELEX database (Baayen et al., 1993; Burnage, 1990). Frequencies were then summed across all cases where that root was a morpheme (e.g., lead is a morpheme in the word mislead, but not in the word plead). This measure is similar to the family frequency measure already calculated for each root, the key difference being the importance of position. The root-morpheme frequency measure included cases where the morpheme in question was not the rst morpheme of a complex word, whereas the family frequency measure for each root did not. For example, the family frequency measure for lead included words
444
TABLE 1 SUMMARY STATISTICS Frequency measure Full-form measures Tokena Familya Morphemic measures Token (root)a Family (root)a Morpheme (root)b Prexb
FOR
LEE H. WURM
EXPERIMENT 1 This experiment tested the predictions of continuous and discontinuous models with respect to the time at which word identication should occur. A strictly continuous model predicts that identication should occur at the full-form UP, while a strict afx-stripping model predicts that identication cannot occur before the root UP. The inclusion of several different frequency measures, some of which are predicted to be irrelevant by a continuous model, was intended to shed additional light on the issue. Finally, this experiment tested whether or not variables such as prex likelihood, semantic transparency, judged prexedness, and stress inuence the identication of morphologically complex words. Method Participants. Thirty-eight students from the Department of Psychology subject pool at the State University of New York at Stony Brook participated. All were native speakers of English with normal hearing. Participants received course credit or cash for their participation. Materials. The 72 prexed and pseudoprexed words from the preliminary rating study, along with their associated roots, were used. These items are listed in Appendix A. Procedure. The full-forms were randomly divided into two sets of 36. The full-forms from one set, along with the roots corresponding to the unpresented full-forms, were presented to one group of participants. A different group of participants heard the other half of the stimuli. Participants in groups of one to four listened to stimuli over headphones in a soundattenuating chamber. For roots, the rst gate consisted of the rst 50 ms of the word, the second gate consisted of the rst 100 ms of the word, and so on, until the entire stimulus was presented. For full-forms, the rst gate was between 50 and 99 ms long, depending on the duration of the words prex (I wanted to have the acoustic onset of one of the gates of the full-form coincide exactly with the
FREQUENCY MEASURES M (SD) Range
8 (26) 22 (66) 121 227 235 1572 (188) (354) (409) (4358)
0195 0529 0807 01786 02379 121963
Note. All values are per million tokens. a From Francis and Kuc era (1982). b From the CELEX database (Baayen, Piepenbrock, & van Rijn, 1993; Burnage, 1990).
such as lead, leading, and leader (among others). The root frequency measure included these words plus words such as mislead. Root-morpheme frequency values are shown in Appendix A. Prex frequencies were computed in essentially the same way. Counts of all words in the Birmingham/Cobuild corpus beginning with each prex string were obtained. From these, the frequencies for those cases that were, in fact, instances of prexation were summed (e.g., preview counts but preen does not). Prex frequencies are shown in Appendix B. Summary statistics for all of the frequency measures are shown in Table 1. Distributions of prex-likelihood values and all six frequency measures had severe positive skew. For all of them, the logarithm of each value was used in the analyses. This rating study provided values on several variables that were chosen as predictors of word recognition performance. The focus of Experiment 1 shifted away from the stimuli per se and toward the recognition performance of participants. Experiment 1 provided the rst direct tests of the competing models: Would root or full-form UPs be more closely related to participants IPs? Would morphemic variables like prex likelihood, prexedness, or semantic transparency be related to IPs? Which frequency measures would matter?
445
acoustic onset of corresponding root). For example, because the full-form belong had a duration of 723 ms and the root long had a duration of 645 ms (giving a prex duration of 78 ms), the rst gate of belong was 78 ms. Gate #2 then began at the acoustic onset of the root. After each gate, participants were given 7000 ms to write down what they thought the word was. Then the next gate, consisting of the already-heard portion plus 50 ms more, was heard, and so on. Stimuli were presented in a random order. Results and Discussion One item (ultrasound) was discarded from all analyses because of digitization problems. A response was considered an error if the participant never did correctly identify the presented word. Two participants data were excluded because of very high error rates (22% and 25%all other participants had error rates less than 5%). The IP, the point at which the participant correctly identied the presented word without subsequently changing his or her mind, was found for each trial. Trials on which the participant never did arrive at the correct word comprised 3.9% of the data. These trials were discarded. IPs were analyzed using multiple regression. The independence of observations assumed by linear regression models did not hold in the current experiment (i.e., each participant provided more than one observation). In repeated-measures regression analyses, this is controlled by the inclusion of N 0 1 dummy variables (35 in this case) that represent the participants. In addition, item analyses cannot be conducted in a repeated-measures regression design. The regressor variables (e.g., prexedness, semantic transparency, and so on) are intrinsic, nonvarying aspects of the items themselveseach item is essentially its own condition. This is reected in the large df value in the denominator, which equals the number of participants times the number of critical stimuli, minus the number of incorrect critical trials and the number of previous factors in the model. The relevant question be-
comes: Were the items selected appropriately (randomly, representatively, or whatever else appropriately might mean)? The interested reader can refer to Cohen and Cohen (1983). The analysis of IPs proceeded at three different levels, the rst of which was a direct comparison of the two competing classes of models. The models make specic predictions that can be tested directlyI simply determined which of the two UPs (full-form vs root) better predicted performance. The second level of analysis involved what I will call prior effects (prex stress and the six frequency measures). I use the term prior effects because of the importance of accounting for these variables before making any strong claims about factors such as prex likelihood and semantic transparency. The prioreffects variables provided another test of the two classes of modelsdecompositional models predict that root measures should be at least as important as full-form measures. Conversely, continuous models predict that only full-form frequency measures should matter. Stress, too, is irrelevant from the standpoint of a strictly continuous model. The nal level of analysis looked at what I will call decomposition variables. These include prex likelihood, prexedness, and semantic transparency, along with their interactions (with each other, with prex and root frequency measures, and with stress). Any signicant effects from level 2 (the prior effects) were partialed from this analysis because those variables had moderate to strong correlations with the variables in this level. The analysis will proceed in exactly this way for Experiment 2, as well. Since this level of the analysis contained a fairly large number of statistical tests (19), p-values were adjusted using a Bonferroni correction. UP analyses. The mean IP for the fullforms in this experiment was 441 ms (SE 4.12 ms), which is much closer to the mean full-form UP (480 ms) than to the mean root UP (714 ms, measured within the full-form tokens). One would expect actual performance in a gating experiment to be slightly earlier than the theoretically critical point because of
446
LEE H. WURM TABLE 2 RESULTS

OF
guessing; on a certain proportion of the trials, participants will stumble onto the correct response by chance before they have heard input sufcient to isolate a unique word candidate. On average, then, IPs were fairly close to the values predicted by a full-listing model and preceded the decomposition values by an average of 273 ms. This result is problematic for discontinuous models. Regression analyses supported this conclusion. Two analyses were performed. After entering the N-1 dummy variables, the second step in each analysis was to enter the relevant UP and see whether it produced a signicant increase in the total explained variance. Both UPs were signicant predictors of performance (F(1,1213) 131.68 and F(1,1213) 44.00 for full-form UPs and root UPs, respectivelyboth ps .0001). Both full-form and root UPs can be signicant because they were correlated with each other (r[70] .57, p .001). When the shared variance between them was accounted for, however, only the full-form UPs were signicant. The partial correlation between fullform UPs and IPs, controlling for participants and for root UPs, was .26 (p .001), while the correlation between root UPs and IPs disappeared when participants and full-form UPs were partialed (r .01, p .66). As noted above, this result supports continuous models over discontinuous ones. Therefore, in subsequent analyses, IPs were measured not from word onset but from the full-form UP of each stimulus. This was necessary because full-form UPs were correlated with other variables under investigation. Any conclusion about the importance of word frequency (for example) that did not account for the fact that each word has its own UP would be meaningless. Prior effects. The next level of the analysis assessed the role of prex stress and the frequency measures. After the dummy variables were entered, prex stress was also found to be a signicant predictor of IPs (F(1,1213) 4.77, p .05). Higher prex stress (this was an ordinal variable taking values of 0 [unstressed], 1 [secondary stress], or 2 [primary
FREQUENCY ANALYSES (EXPERIMENT 1) Sign F(1, 1212)
Regressor Full-form measures Log token frequency (full-form) Log family frequency (full-form) Morpheme measures Log token frequency (root) Log family frequency (root) Log morpheme frequency (root) Log prex frequency * p .01. ** p .001. *** p .0001.
0 0 / / / /
10.42* 12.08**
1 1 1.37 25.81***
stress]) was associated with later IPs. This could be easily explained if IPs were measured from word onset: stressed prexes have longer durations. However, IPs were measured from the full-form UP. The MSS (Cutler & Norris, 1988), while not addressing such variables as semantic transparency, predicts that lexical access should be root-driven when items have unstressed prexes. This would mean later IPs for those items because root UPs were always later than full-form UPs. The observed relationship runs in the opposite direction. However, the MSS makes a second prediction that is consistent with the observed relationship: an effect like this could appear if participants mistakenly treated stressed prexes as the rst portion of roots. An attempt to access a root morpheme, using the prex as the assumed rst syllable (because it is stressed), will fail. For example, participants might hear the initial portion of the word copilot and mistakenly access words like cope. Items like these might have late IPs, because participants are likely to perseverate on a mistaken hunch in a task like gating. Which frequency measures were related to IPs? Six regression analyses were performed. In each one, the dummy variables were entered rst, along with prex stress, and then one of the frequency measures described above. As can be seen in Table 2, both of the full-form frequency measures had signicant
447
facilitative relationships with IPs. This is the type of relationship generally expected: higher frequency values were associated with earlier IPs. Of the four morpheme-relevant measures, only prex frequency was a signicant predictor of performance, and it had an inhibitory relationship with the dependent variable: more common prexes led to slower performance. This effect has a natural explanation in terms of neighborhood densityprex frequency was signicantly associated with the density measure described earlier (r[70] .66, p .001). This suggests that earlier IPs should be found when prex frequency is low because these items reside in relatively sparse lexical neighborhoods. Decomposition variables. This group consisted of prex likelihood, judged prexedness, and semantic transparency, as well as interactions involving them. Continuousprocessing models predict that none of these analyses will be signicant. As will be seen, that prediction does not hold. Prex likelihood turned out to play a major role in mediating the effects of other variables. A separate analysis was conducted for each regressor. As the rst step, the dummy variables, prex stress, prex frequency, and fullform family frequency (the stronger of the two signicant full-form measures) were entered. Then the variable being considered was entered. Higher values of semantic transparency were associated with earlier IPs (F(1,1210) 51.98, p .001). This nding supports decompositional models because the semantic transparency of a morphological combination is irrelevant from a continuous perspective. There was also a trend toward the same kind of IP advantage for higher levels of prexedness (F(1,1210) 8.97, p .10). By itself, prex likelihood had virtually no effect on IPs (F(1,1210) 1). Four analyses were conducted to assess the interactions between these three variables (three possible two-way interactions and the three-way interaction). In all cases, the dummy variables, prex stress, prex frequency, and family frequency of the full-form
FIG. 1. Mean identication point (IP) as a function of semantic transparency and prex likelihood, in milliseconds (ms). An IP of 0 corresponds to the full-form uniqueness point of each stimulus.
were partialed. In addition, the main effects of the variables making up the interaction were partialed. For the three-way interaction, the two-way interactions were also partialed. These steps are all necessary to satisfy assumptions of the general linear model (Cohen & Cohen, 1983). For ease of interpretation, readers should note three things about the gures that will be shown. First, untransformed prexedness ratings are shown so that values can be interpreted more naturally (transformed values were used in the statistical tests). Second, negative IPs reect the fact that, on average, identication of the items took place some 39 ms before the full-form UP. Third, although the Y-axes show mean IPs based on median splits, readers are reminded that none of the predictor variables were dichotomized in the analyses; this is simply the best way to illustrate the nature of each interaction. The interaction between prex likelihood and semantic transparency was signicant and is shown in Fig. 1 (F(1,1208) 22.84, p
448
LEE H. WURM
.001). As can be seen in the gure, higher semantic transparency was associated with earlier IPs but only for items with high-likelihood prexes. As with the signicant main effect of semantic transparency, this interaction supports a decompositional processing account. If we consider high values on decomposition variables to be cues associated with successful morphemic access, then the earliest IPs on the graph indicate that multiple cues result in enhanced performance. The latest IPs on the graph seem to reect a cost incurred when the processing system deals with items that, according to one variable (e.g., prex likelihood), should be decomposed but according to another (e.g., semantic transparency) should not. The three-way interaction was also signicant (F(1,1204) 13.32, p .01) due to the fact that the two-way interaction between semantic transparency and prex likelihood shown in Fig. 1 was more pronounced the lower judged prexedness was. This may be due to the moderately high correlation between prexedness and semantic transparency (r[70] .58, p .01). Twelve additional regression analyses were conducted, each of which looked at the interaction between two of the following variables: prex stress, prex frequency, root morpheme frequency, prex likelihood, prexedness, and semantic transparency. As a rst step, the dummy variables, prex stress, prex frequency, and family frequency for the full-form were partialed, along with the two variables whose interaction was being assessed. The interaction term was entered next. It might be that morphemic variables that do not account for much variance overall (i.e., their main effects are not signicant) are nonetheless important for certain kinds of words (i.e., they interact). One of these interactions was signicant, providing further evidence of decompositional processing. This was the root frequency 1 prexedness interaction, shown in Fig. 2 (F(1,1208) 9.11, p .05). One interesting possibility concerning the positive slope of the
line for highly prexed items is competition: a high-frequency root may compete for activation with the full-form that contains it, delaying recognition of the full-form. Schreuder and Baayen (1994) assert that this type of competition is possible, and further evidence for it will be presented below. EXPERIMENT 2 The gating paradigm has some distinct advantages. First of all, researchers can be precise about how much acoustic information is available at each response point. In addition, gating responses provide a picture of how the group of word candidates forms and is narrowed down. In spite of this, gating has been criticized for various reasons, the most important being that listeners hear multiple repetitions of each stimulus and that responses are not speeded. These criticisms have been convincingly addressed by several studies (e.g., Cotton & Grosjean, 1984; Marslen-Wilson, 1984; Salasoo & Pisoni, 1985; Tyler, 1984; Tyler & Marslen-Wilson, 1986; Tyler & Wes-
FIG. 2. Mean identication point (IP) as a function of root frequency and judged prexedness, in milliseconds (ms). An IP of 0 corresponds to the full-form uniqueness point of each stimulus.
449
sels, 1983; 1985), but it is always wise to ensure that a given result is not due to the peculiarities of an experimental paradigm. Experiment 2 tests the same factors as Experiment 1, with a different methodology (auditory lexical decision). Participants heard each stimulus uninterrupted and intact, and their word/nonword responses were made as quickly as possible. Method Participants. Participants were 110 students from the Department of Psychology subject pool at the State University of New York at Stony Brook. All were native speakers of English with normal hearing. Participants received course credit or cash for their participation. Materials. The words used in this experiment are listed in Appendix C. The full-forms from Experiment 1 that carried a two-syllable prex were dropped in the interest of uniformity. Three hundred and sixty stimuli were divided into two lists of 180. Each list contained: 30 of the full-forms from Experiment 1; the roots of the other 30 full-forms from Experiment 1; 30 ller words; 30 ller nonwords; and 60 morphological nonwords. These last 60 nonwords were of four different types: no obvious morphemic structure (e.g., pangort); genuine English prex with a nonword root (e.g., prezelp); nonprex with a real word root (e.g., grevent); and genuine English prex with a real word root (e.g., precorrect). Fifteen of each type were contained in each list. In order to use the data on all 60 of the fullforms, it was necessary to assess the effect of List 1 presentation on List 2 performance. Certain stimulus items were included with this in mind. First, the ller items (30 words and 30 nonwords) were identical in each list and were included to assess the effect of exact repetition. Second, a subset of the participants (44/110) listened to the stimuli described above plus an additional 40 words, 20 in each of the two lists. These words were included to assess practice effects. Filler words were chosen from the same
distributions as the other word stimuli, in terms of word frequency, number of phonemes and syllables, and stress. To discourage strategic responding, these words were of a variety of different morphological types (monomorphemic, sufxed, and prexed, including both bound and unbound roots). Stress was approximately balanced for words and nonwords. A practice list of similar composition was used prior to the main experiment. This list consisted of 24 stimuli. Procedure. Participants in groups of one to four listened to stimuli over headphones in a sound-attenuating chamber, and stimulus presentation was randomized for each group of participants. Half of the participants heard list A and then list B; the order was reversed for the other half. On each trial, a participant heard an auditory stimulus and made a speeded lexical decision by pressing buttons on a response board with his/her dominant hand. Participants pushed one button for words and another button for nonwords. Results and Discussion Participants were excluded if they had an error rate greater than 15% or a mean RT greater than 1000 ms on either of the two lists they heard. Twelve participants were excluded by these criteria. Analyses were conducted on responses of the remaining 98 participants. The lexical decision task is essentially a kind of familiarity judgment. Participants have some familiarity threshold above which they respond Word and below which they respond Nonword. Table 3, which shows mean RTs for each type of word as a function of whether it was in the rst list a participant heard or in the second, can be understood if one assumes that because of the overall higher familiarity after some experience with the various kinds of stimuli (particularly the quite novel nonwords), participants familiarity thresholds were higher in List 2. The familiarity interpretation explains the 50-ms facilitation for ller words, which was signicant in an ANOVA by subjects (F1(1,97) 54.84, p .001) and by items (F2(1,29) 39.38, p .001). These words were exactly repeated and
450
TABLE 3
LEE H. WURM
MEAN RTs (ms) AS A FUNCTION OF WORD TYPE LIST ORDER (EXPERIMENT 2) List heard Word type Fillersa Full-formsb Rootsb Practice-effect wordsc
a b
AND
First 637 603 489 656
Second 587 607 486 684
Difference
050** /4 03 /28
These items were exactly repeated in both lists. Any effect here would be essentially a repetition priming effect, caused by prior presentation of a shared root morpheme. For example, if a given participant heard rebuild in the rst list, build would be in the second list (and vice versa). c This is strictly a practice effect, as none of the items in this group was repeated. These items were heard by only a subset of the participants (44 out of 110). p .10 for subjects, p .05 for items. ** p .001 for subjects and items.
thus highly familiar. The familiarity-threshold account can also explain the 28-ms cost for practice-effect words, which were new and thus unfamiliar (signicant by items, marginal by subjects F1(1,97) 3.70, p .10; F2(1,39) 7.04, p .05). Full-forms and roots more or less broke even, due to the raised familiarity thresholds canceling the moderate increase in familiarity (all four F-ratios 1). The corresponding effects for the nonwords averaged 031 ms, with the negative sign reecting the importance of increased familiarity for nonwords. Data from the rst and second list each participant heard were pooled, and rst or second list heard was used as a regressor in the analyses. However, this variable will not be discussed further; it did not account for a signicant proportion of the variance nor did it interact with any of the variables under study. Analysis of the word data parallels the analyses for Experiment 1. The same regressors were used and the underlying logic was identical, but the dependent variable was RTs instead of IPs. UP analyses. Measuring RTs from word
onset, the importance of the two UPs in predicting performance was assessed as before. As in the gating data, both were signicant (F(1,5332) 120.85, p .0001, and F(1,5332) 55.16, p .0001, for the fullform and root UPs, respectively). The partial correlation coefcients lead to the same interpretation as before: the correlation between full-form UPs and RTs, partialing root UPs and participants, was .12 (p .001), and the analogous partial correlation for root UP, though signicant, was much smaller (r .03, p .05). Thus, as in the gating analyses, RTs for the subsequent levels of the analysis were measured from the full-form UP of each stimulus. Prior effects. Prex stress was marginally related to RTs (F(1,5332) 3.04, p .10); items with stressed prexes tended to have somewhat slower RTs (consistent with the gating result). Table 4 shows the results of the frequency measure analyses. The four morphemic measures had inhibitory relationships with RT, while the two full-form measures had the expected facilitative relationships. As discussed in Experiment 1, the effect of prex frequency could be interpreted in light of that variables correlation with neighborhood density. The root frequency effects cannot be explained in this way but may indicate competition between roots and the full-forms that contain them (as discussed by Schreuder and Baayen [1994] and in connection with Fig. 2
TABLE 4 RESULTS
OF
FREQUENCY ANALYSES (EXPERIMENT 2) Sign F(1, 5332)
Regressor Full-form measures Log token frequency (full-form) Log family frequency (full-form) Morpheme measures Log token frequency (root) Log family frequency (root) Log morpheme frequency (root) Log prex frequency * p .01. *** p .0001.
0 0 / / / /
9.72* 36.06*** 48.43*** 26.40*** 30.96*** 30.59***
451
FIG. 3. Mean reaction time (RT) as a function of semantic transparency and prex likelihood, in milliseconds (ms).
of the present study). Such root effects support decompositional models, while the UP results and the full-form frequency effects t the predictions of a continuous processing model (as was true in Experiment 1). From the analyses conducted to this point, then, it was determined that prex frequency, token frequency of the root, and family frequency of the full-form would be partialed from the further analyses. Again, the following p-values have been adjusted using a Bonferroni correction. Decomposition variables. The results for the decomposition variables mirrored those from the gating experiment and once again suggest decompositional processing. Higher levels of prexedness (F(1,5329) 35.46, p .001) and semantic transparency (F(1,5329) 66.60, p .001) were associated with faster RTs, and once again prex likelihood was not signicant. As in the gating experiment, the prex likelihood 1 semantic transparency interaction was signicant (F(1,5327) 10.75, p .05). Figure 3 shows this interaction and resembles
Fig. 1, at least as far as the slopes of the two lines are concerned. In this gure, however, performance on low prex-likelihood items relative to high prex-likelihood items was somewhat poorer than it was in the gating experiment. This interaction supports decompositional models. Five of the next set of 12 interactions were signicant, as well. All of these signicant interactions involved either prex likelihood or prex stress (or both). Figures 4 and 5 show the interactions between prex stress and prexedness and between prex stress and semantic transparency (F(1,5327) 10.57, p .05, and F(1,5327) 57.87, p .001, respectively). Figure 4 shows that when prexedness was low, the stress value of the prex was largely irrelevant. Theoretically, for low prexedness values, listeners would not process the prex and root separately. For highly prexed items, one would expect to observe a RT advantage (recalling the main effect of prexedness); however, this advantage was erased for those highly prexed items that also happened to have stressed prexes. This suggests that stressed prexes fool the lexical
FIG. 4. Mean reaction time (RT) as a function of prex stress and judged prexedness, in milliseconds (ms).
452
LEE H. WURM
FIG. 5. Mean reaction time (RT) as a function of prex stress and semantic transparency, in milliseconds (ms).
access system, drawing processing away from the root, a possibility that was discussed earlier. Figure 5 shows the interaction between semantic transparency and prex stress. Given the strong correlation between prexedness and semantic transparency, the expectation was that this gure would resemble Fig. 4. It did, except for items low on semantic transparency with fully stressed prexes. As predicted by the MSS (Cutler & Norris, 1988), strongweak words that are functionally monomorphemic (low on transparency) enjoyed a RT advantage. Figures 6 and 7 show the prex stress interactions with root frequency (F(1,5327) 10.82, p .05) and prex frequency (F(1,5328) 12.92, p .01). The two show the same general pattern. For items with lowfrequency roots or prexes, prex stress had a weak facilitative association with RTs, while for items with high-frequency roots or prexes, prex stress had a stronger, inhibitory association. These effects t the speculation made earlier about the disruptiveness of words with strong prexes; the disruptiveness of
high-frequency roots (Fig. 6) or prexes (Fig. 7) was heightened by a potentially-disruptive (i.e., fully-stressed) prex. Figure 8 shows the prex frequency 1 prex likelihood interaction (F(1,5328) 11.22, p .05). The inhibitory effect of prex frequency was strong for items with low prex likelihoods, while there was only a weak effect of prex frequency for items with high prex likelihoods. One interpretation of this pattern is that the perceptual system strips off prexes when the prex likelihood is above some threshold value. Prex frequency cannot have much of an effect if the prex has been stripped away, which would be the case for items very high on prex likelihood. The delayed word recognition for high-frequency prexes is consistent with the observation made previously about high-density lexical neighborhoods. As I noted in Experiment 1, none of these variables is predicted to have any relevance by continuous processing models (and obviously, no interactions between them are predicted). Therefore, all of the signicant interactions
FIG. 6. Mean reaction time (RT) as a function of prex stress and root frequency, in milliseconds (ms).
453
Discontinuous processing is implied by the results involving the decomposition variables and root frequency. These variables are completely irrelevant from the standpoint of the continuous, left-to-right processing model implicated by the UP and full-form frequency results, yet all of these variables were signicant either as main effects or in interactions. In particular, it is clear that prex likelihood affects word-recognition performance, in combination with other variables. Such results cannot be accommodated within a strictly continuous framework. Conditional Root UPs (CRUPs) The original prediction concerning the decomposition variables was that they would be positively related to IPs and RTs. Items high on prexedness or semantic transparency, according to the prediction, would be more likely to be decomposed into morphemes and accessed via their roots. They would therefore have later IPs and slower RTs, because root UPs were always later in the stimuli than fullform UPs.
FIG. 7. Mean reaction time (RT) as a function of prex stress and prex frequency, in milliseconds (ms).
constitute evidence for decompositional processing. GENERAL DISCUSSION The results of this study indicate that the auditory processing of prexed words has both continuous and decompositional characteristics. The results of the two main experiments with respect to the full-form vs root UPs, as well as the outcomes of the full-form frequency analyses, support left-to-right processing models. Although the reliable inhibitory relationship between prex frequency and performance, observed in both experiments, was not specically predicted by either class of model, it is best viewed as a by-product of continuous processing. Decompositional models would predict that if prex frequency were going to be used for word recognition, higher frequency would be associated with better performance, not worse (higher frequency, as I have noted, is generally associated with faster processing). Therefore, this result also favors continuous processing models, though only mildly.
FIG. 8. Mean reaction time (RT) as a function of prex frequency and prex likelihood, in milliseconds (ms).
454
LEE H. WURM
The fact that the observed effects were in the opposite direction (as well as the overall earliness of IPs in Experiment 1, which occurred on average 39 ms before the full-form UP) may be explained by a model that has two parallel but nonindependent processes, a whole-word process and a morphemic one, but only if the morphemic process is selective about the candidates it considers. At word onset, both processes in such a model would begin and run in parallel. The whole-word process simply checks the accumulating input against the lexicon, in exactly the same way as the class of continuous models I have described throughout this paper. This process will achieve a match at the fullform UP. I will describe the operation of the morphemic process in the following paragraphs. The morphemic process, running in parallel, strips any prex and then attempts to make a lexical match using the portion of the signal beginning just after the prex. The relevant UP for this process depends on the root rather than the full-form, but it is not the root UP I have been discussing throughout this paper. The process I am describing is selective in that, in attempting to match the input to the lexicon, it considers only unbound roots that attach to this prex. The morphemic process will achieve a match at what I will call the conditional root UP, or CRUP. A words CRUP is the root UP given this particular prex. Usually, the CRUP of a prexed word will be the same point as the full-form UP. However, there are many exceptions to this general rule. For example, consider the word discredit. The full-form UP of this word is the second /d/, because of words such as discretion. The root UP of discredit is the /t/, because credible is still a competitor prior to that point. The CRUP of discredit, though, is the /r/: the only words still consistent with discr- are discretion, discrepant, and their morphological relatives, but because the morphemic process in this model only considers unbound roots, those
words do not count (-cretion and -crepant are not unbound roots). Such a morphemic process could be a historical by-product of the way language is used. Speakers coin new words as they need them, essentially continuously (Henderson, 1985). Writers do this, too: Baayen (1994) found that approximately 15 new words ending in -ity appear each month in The New York Times. Although they are perfectly intelligible, they are nonwords in that they have never been seen before by readers and are not in any dictionary. Evidence comes from more anecdotal sources, as well: two examples I have recently heard (one in the hallway and one on the evening news) are They were interrupted *midsong by a power outage, and Violence threatens to *re-Balkanize the region. These nonwords were quite easy to interpret, and the speakers presumably knew that they would be. Aronoff (1976) has noted that new word forms tend to be quite highly transparent while older ones may not be because of semantic drift. I have argued above that unbound roots are required for high semantic transparency. To the extent that this argument is on the mark, a decompositional process that only considers unbound roots would prove extremely useful in the recognition of newly coined forms, which we encounter extremely often (Henderson, 1985). In order to prevent the perceptual system from committing prematurely to the wrong word, the perceptual system would have to wait for verication from the whole-word process before ultimately committing to a decision. Otherwise, the perceptual system might decide at the CRUP that the word being heard is discredit, when in fact it is discretion (for example). However, even though the perceptual system would not be allowed to fully commit to discredit until discretion has been ruled out, facilitated recognition performance would reect an activation boost received by the root credit at the CRUP (a boost that was not received by bound roots like -cretion). Furthermore, it is possible
455
that other kinds of computations can begin once a CRUPs-based hypothesis has been generated, while the system is waiting for the whole-word response. These might include accelerated access of meaning, integration of the lexical hypothesis with a sentence representation, and/or access of a grammatical category. As a post hoc check of this account, all of the words were analyzed to see if their CRUPs differed from their full-form UPs. Of the 71 full-forms in Experiment 1, nine (12.7%) had CRUPs that were earlier than their full-form UPs, as in the above example of discredit. For the remaining 62 words, the CRUP was the same point as the fullform UP. This suggests that even if premature commitment to an incorrect lexical hypothesis cannot be avoided in the way I just described (i.e., even if the morphemic process does not wait for the whole-word process), the occurrence of such errors would be fairly low; perhaps the computational advantages outweigh the occasional cost. The mean CRUP across all 71 words was 456 ms, a value that corresponds to the mean IP in Experiment 1 (441 ms) even better than the mean full-form UP (480 ms) did. However, there was conicting evidence regarding CRUPs. In a regression analysis for CRUPs just like that for full-form UPs and root UPs, although CRUPs did account for a signicant increase in the explained variance (F(1,1213) 81.52, p .0001, for the gating experiment), the effect was not as large as that for full-form UPs. Still, this notion is deserving of some direct investigation with stimuli explicitly chosen to test it, because of the close overall correspondence between CRUPs and IPs and because several of the results in this study are most easily interpretable from the perspective of a selective parallel model. One could argue that the CRUPs result is artifactual and due to the combination of the gating task (where prexes might be argued to be processed completely and independently of what follows them) and the stimulus set (there were no stimuli with bound roots or nonroots). If this were the case, CRUPs should
not predict RTs in Experiment 2, because such strategic responding was unlikely: that experiment not only used a different methodology, but also contained a variety of different types of words and nonwords (sufxed, prexed [with bound as well as unbound roots], and monomorphemic). The CRUPs result for Experiment 2 was just like that for Experiment 1: CRUPs signicantly predicted RTs (F(1,5332) 54.11, p .0001), doing a better job than root UPs but a somewhat worse job than full-form UPs. There are two ndings in the literature that may relate to CRUPs. Schriefers, Zwitserlood, and Roelofs (1991) found that prexed words were identied earlier than unprexed words with identical predicted recognition points (in Dutch). For example, both the prexed word opstaan and the root staan have the nal /n/ as their UP, but in a gating experiment participants identied the prexed words an average of 37 ms earlier (compared to the 39 ms advantage found in Experiment 1 of this study). This general prexation advantage was replicated in two subsequent experiments and could not be explained by either class of model. It is not possible to say whether CRUPs would explain this effect, but the possibility is intriguing. Taft (1988) also found what appears to be a general prexation advantage in lexical decision times, although he provided relatively little methodological detail. To explain his nonword data, he suggested an activate and check model that is similar to the smart morphemic half of the dual-route model suggested in this paper, but one that considers bound rather than unbound roots. The model did not apply, however, to his word data. CRUPs may also help explain the prex stress interactions from Experiment 2 of the current study. Specically, if the basic stressdriven model is modied so that it is sensitive to the decomposition variables and so that it operates via CRUPs rather than root UPs, there seem to be two possible predicted outcomes. One of these is illustrated fairly well by Fig. 4: prex stress should not matter for items low on prexedness (for example), but
456
LEE H. WURM
should begin to affect RTs as prexedness increases. The RTs that should stand out as faster (because of the facilitative potential of CRUPs) are those for words high on prexedness with unstressed prexes. CONCLUSION Most existing models of word recognition have too rigidly insisted on either continuous processing with total disregard for decompositional variables or on strict decompositional processing in every instance. The current study demonstrates that both theoretical positions are wrong; there is evidence for both full-form processing and at least some decomposition, the latter perhaps involving CRUPs. In hindsight, the insistence of some researchers on one or the other kind of mechanism seems difcult to understand. The ease with which people coin and understand new words, the idea of semantic drift, and the exibility of the perceptual system demonstrated in various tasks should have suggested earlier that there might be both types of processing mechanism. In addition, as pointed out in the Introduction, while nonword data from studies such as Taft et al. (1986) suggest an important role for roots, this does not imply that the system always uses root access for all words. As I have noted throughout this paper, the Cohort model (Marslen-Wilson, 1984; Marslen-Wilson & Welsh, 1978) cannot accommodate the results having to do with the decomposition variables. In addition, the suggested competition between high-frequency roots and the full-forms that carry them ies directly in the face of the basic cohort principle. The current results suggesting an important role for semantic transparency in decompositional processing is, however, consistent with the conclusion of Marslen-Wilson et al. (1994). Their model may be able to accom-
modate the current results, if modied to reect the importance of prex likelihood, for example. Cutler et al. (1985) argue in favor of a decompositional serial autonomous model, based on process considerations rather than storage and efciency considerations. They believe that a strict prex-stripping account is probably wrong, but that decomposition of prex and stem is a strategy that seems routinely available to the language processor. However, one aspect of their framework that would seem not to t the current results is their insistence that listeners compute the meaning of stems before afxes. While this makes sense in many situations, particularly those involving sufxing (e.g., sad / -ness), the current results indicate that prexes are dealt with very early in the recognition process. Prex likelihood should not play a role in the early stages of recognition if stem processing has to be completed rst. Network models (e.g., McClelland & Elman, 1986; Norris, 1994) might be able to accommodate the current results without having two separate processes. Such models are often claimed to have the ability to extract structural (e.g., that re- is a separable unit) or distributional (e.g., relating to prex likelihoods) information based on covariance of form and meaning. Van Orden et al. (1990) provide a useful discussion of this issue. A nal interesting point concerns the fact that the structure of the input itself determines the specics of perceptual processing. It may well be protable to apply the techniques developed here crosslinguistically, choosing languages with known differences in morphology and afx structure. Such a test could illustrate how the perceptual process gets instantiated differentially in response to different language environments.
457
APPENDIX A Judged Prexedness, Semantic Transparency, and Root Morpheme Frequency Full-form (root) ablaze (blaze) abreast (breast) abridge (bridge) acquire (choir) aloft (loft) antibody (body)c antisocial (social)c archbishop (bishop) ascend (send) aspire (spire) atingle (tingle) beget (get) behold (hold) belong (long) beside (side) circumscribe (scribe)c concave (cave) condense (dense) copilot (pilot) counterplot (plot)c default (fault) defrost (frost) discredit (credit) disgure (gure) disgust (gust) distaste (taste) embattle (battle) embody (body) engage (gauge) engulf (gulf) entitle (title) forecast (cast) foreshadow (shadow) forewarn (warn) hyperspace (space)c increase (crease) infringe (fringe) insecure (secure) intercourse (course)c interstate (state)c midyear (year) mist (t) mislead (lead) Judged prexednessa 6 6 4 1 5 8 8 6 6 3 6 6 5 3 5 5 6 5 8 7 7 8 8 8 4 8 6 6 5 7 7 6 7 7 7 3 7 8 7 7 7 7 8 Semantic transparencya 6 2 2 1 2 2 8 7 2 2 3.5 2 4 1 5 2 2 4.5 8 7 3 7 7 4 1 7 4 3 2 1.5 3 2 2 7.5 5.5 1 3 8 2 6.5 7 3 7.5 Root morpheme frequencyb 17 83 71 8 11 705 537 40 72 20 3 2379 527 1508 1236 2 42 33 25 32 55 16 58 229 5 93 92 705 0 22 46 127 72 50 154 8 20 184 749 497 1420 145 557
458
LEE H. WURM
APPENDIX A Continued Full-form (root) mistrial (trial) monorail (rail)c perfume (fume) persuade (suede) preamble (amble) prex (x) preheat (heat) proclaim (claim) rebuild (build) recite (cite) rephrase (phrase) reprint (print) reword (word) subsample (sample) supersede (seed)c surcharge (charge) surname (name) transcend (send) twilight (light) twinight (night) ultrasound (sound)c unbend (bend) unborn (born) unbuckle (buckle) unbutton (button) underscore (score)c unplug (plug) unscramble (scramble) unusual (usual)
a b
Judged prexednessa 8 7 1 5 8 8 8 6 8 7 8 8 8 8 7 7 6 5 3 5.5 7 8 8 8 8 7 8 8 8
Semantic transparencya 7 6 4 1 3 1 7 7 7 2 7 7 6.5 7 2 7 7 3.5 6 5 6 8 8 8 8 6 7.5 8 8
Root morpheme frequencyb 66 89 13 3 12 52 143 177 277 26 50 95 502 21 3 162 421 72 530 625 295 77 113 7 34 57 17 19 311
Values can range from 18. Per million tokens. c These items were used only in Experiment 1.
459
APPENDIX B Neighborhood Size, Prex Likelihood, and Prex Frequency

Prex aantiarchbecircumconcocounterdedisemenforehyperinintermidmismonoperpreproresubsupersurtranstwiultraunundera b
Neighborhood size 16,803 72 18 6562 109 4828 1690 22 5733 2416 644 2017 372 7 9444 1101 163 243 48 2153 2214 3832 8561 608 109 556 346 4 9 3271 543
Prex likelihood .090 .028 .722 .304 .018 .012 .010 .545 .008 .092 .003 .092 .212 .000b .094 .212 .264 .181 .083 .005 .013 .001 .067 .061 .284 .009 .092 1.000 .889 .283 .232
Prex frequencya 2454 28 20 204 1 228 18 26 104 766 78 371 129 1 21963 256 60 79 22 484 56 152 1881 649 49 654 47 9 2 1072 234
dle, eclipse, frontier, gratify, grocer, hindrance, hockey, image, inert, infant, inspiring, legal, linguist, minus, mission, mystique, operate, referee, subsequent, teeter, torpedo, velour. Practice-effect words (these items were heard by 44 of the 110 participants): acumen, audition, brutal, catapult, cognizant, compel, cutlets, despair, detachment, dinnerware, eyeball, uttering, frankly, heavenly, hesitance, instinct, iron, mazurka, mermaid, mistletoe, murmur, nourish, partial, privilege, pseudonym, ranking, reinforce, shrubbery, slumber, spaghetti, splatter, stainless, subpoena, sweetheart, thinning, towel, twitter, vernacular, worthy, wristwatch. REFERENCES
Anderson, S. R. (1988). Morphology as a parsing problem. Linguistics, 26, 521544. Anshen, F., & Aronoff, M. (1981). Morphological productivity and phonological transparency. Canadian Journal of Linguistics, 26, 6372. Aronoff, M. (1976). Word formation in generative grammar. Cambridge, MA: MIT Press. Baayen, R. H. (1994). Productivity in language production. Language and Cognitive Processes, 9, 447 469. Baayen, R. H., Piepenbrock, R., & van Rijn, H. (1993). The CELEX lexical database. [CD-ROM] Bergman, M. W., Hudson, P. T. W., & Eling, P. A. T. M. (1988). How simple complex words can be: Morphological processing and word representations. Quarterly Journal of Experimental Psychology, 40A, 41 72. Burnage, G. (1990). CELEX: A guide for users. Nijmegen, Netherlands: CELEX. Bybee, J. L. (1988). Morphology as lexical organization. In M. Hammond & M. Noonan (Eds.), Theoretical morphology: Approaches in modern linguistics (pp. 119141). New York: Academic Press. Caramazza, A., Laudanna, A., & Romani, C. (1988). Lexical access and inectional morphology. Cognition, 28, 297332. Cohen, J., & Cohen, P. (1983). Applied multiple regression/correlation analysis for the behavioral sciences (2nd ed.). Hillsdale, NJ: Erlbaum. Cotton, S., & Grosjean, F. (1984). The gating paradigm: A comparison of successive and individual presentation formats. Perception & Psychophysics, 35, 4148. Cutler, A. (1976). Phoneme monitoring reaction time as a function of preceding intonation contour. Perception & Psychophysics, 20, 5560. Cutler, A., Hawkins, J. A., & Gilligan, G. (1985). The
Per million tokens. The value of .000 for hyper- seems odd, given words such as hyperactive. In fact, there were 24 words beginning with hyper- in Websters Third New International Dictionary of the English Language (1993) that qualied as truly prexed, but none of them was listed in the frequency count of Francis and Kuc era (1982).
APPENDIX C Words Used in Experiment 2 Critical items: Items from Appendix A not denoted by a superscript c. Fillers: alphabet, anthem, bedazzle, bridal, bugle, curbside, deduction, dependent, dwin-
460
LEE H. WURM Marslen-Wilson, W. D. (1984). Function and process in spoken word recognition. In H. Bouma & D. G. Bouwhuis (Eds.), Attention and Performance: Vol. X. Control of Language Processes (pp. 125150). Hillsdale, NJ: Erlbaum. Marslen-Wilson, W. D., Tyler, L. K., Waksler, R., & Older, L. (1994). Morphology and meaning in the English mental lexicon. Psychological Review, 101, 333. Marslen-Wilson, W. & Welsh, A. (1978). Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology, 10, 29 63. McClelland, J. L., & Elman, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18, 186. Moby Pronunciator (Version 1.01) (1989). Monterey, CA: Illumind. [Computer software] Morton, J. (1969). The interaction of information in word recognition. Psychological Review, 76, 165178. Morton, J. (1979). Facilitation in word recognition: Experiments causing change in the logogen model. In P. A. Kolers, M. E. Wrolstad, & H. Bouma (Eds.), Processing of visible language (Vol. 1, pp. 259 268). New York: Plenum. Norris, D. (1994). Shortlist: A connectionist model of continuous speech recognition. Cognition, 52, 189 234. Rubin, G. S., Becker, C. A., & Freeman, R. H. (1979). Morphological structure and its effects on visual word recognition. Journal of Verbal Learning and Verbal Behavior, 18, 399412. Salasoo, A., & Pisoni, D. B. (1985). Interaction of knowledge sources in spoken word recognition. Journal of Memory and Language, 24, 210231. Sandra, D. (1994). The morphology of the mental lexicon: Internal word structure viewed from a psycholinguistic perspective. Language and Cognitive Processes, 9, 227269. Schreuder, R., & Baayen, R. H. (1994). Prex stripping re-revisited. Journal of Memory and Language, 33, 357375. Schriefers, H., Friederici, A., & Graetz, P. (1992). Inectional and derivational morphology in the mental lexicon: Symmetries and asymmetries in repetition priming. Quarterly Journal of Experimental Psychology, 44A, 373390. Schriefers, H., Zwitserlood, P., & Roelofs, A. (1991). The identication of morphologically complex spoken words: Continuous processing or decomposition? Journal of Memory and Language, 30, 2647. Tabachnick, B. G., & Fidell, L. S. (1989). Using multivariate statistics (2nd ed.). New York: Harper Collins. Taft, M. (1979a). Lexical access via an orthographic code: The Basic Orthographic Syllable Structure (BOSS). Journal of Verbal Learning and Verbal Behavior, 18, 2139. Taft, M. (1979b). Recognition of afxed words and the
sufxing preference: A processing explanation. Linguistics, 23, 723758. Cutler, A., & Norris, D. G. (1979). Monitoring sentence comprehension. In W. E. Cooper & E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett (pp. 113134). Hillsdale, NJ: Erlbaum. Cutler, A., & Norris, D. (1988). The role of strong syllables in segmentation for lexical access. Journal of Experimental Psychology: Human Perception and Performance, 14, 113121. Emmorey, K. D. (1989). Auditory morphological priming in the lexicon. Language and Cognitive Processes, 4, 7392. Francis, W. N., & Kuc era, H. (1982). Frequency analysis of English usage: Lexicon and grammar. Boston: Houghton Mifin. Frauenfelder, U. H., & Schreuder, R. (1992). Constraining psycholinguistics models of morphological processing and representation: The role of productivity. In G. E. Booij & J. Van Marle (Eds.), Yearbook of Morphology 1991 (pp. 165183). Dordrecht, Netherlands: Kluwer. Goldinger, S. D., Luce, P. A., & Pisoni, D. B. (1989). Priming lexical neighbors of spoken words: Effects of competition and inhibition. Journal of Memory and Language, 28, 501518. Grosjean, F., & Gee, J. (1987). Prosodic structure and spoken word recognition. Cognition, 25, 135155. Henderson, L. (1985). Toward a psychology of morphemes. In A. W. Ellis (Ed.), Progress in the psychology of language (Vol. 1, pp. 1572). Hillsdale, NJ: Erlbaum. Henderson, L., Wallis, J., & Knight, K. (1984). Morphemic structure and lexical access. In H. Bouma & D. G. Bouwhuis (Eds.), Attention and performance: Vol. X. Control of language processes (pp. 211 226). Hillsdale, NJ: Erlbaum. Jarvella, R. J., & Meijers, G. (1983). Recognizing morphemes in spoken words: Some evidence for a stemorganised mental lexicon. In G. B. Flores dArcais & R. J. Jarvella (Eds.), The process of language understanding (pp. 81112). London: Wiley. Laudanna, A., Burani, C., & Cermele, A. (1994). Prexes as processing units. Language and Cognitive Processes, 9, 295316. Luce, P. A. (1987). Neighborhoods of words in the mental lexicon. Dissertation Abstracts International, 47, 5078. Luce, P. A., Pisoni, D. B., & Goldinger, S. D. (1990). Similarity neighborhoods of spoken words. In G. T. M. Altmann (Ed.), Cognitive models of speech processing: Psycholinguistic and computational perspectives (pp. 122147). Cambridge, MA: MIT Press. MacKay, D. G. (1978). Derivational rules and the internal lexicon. Journal of Verbal Learning and Verbal Behavior, 17, 6171.
AUDITORY PROCESSING OF PREFIXED WORDS word frequency effect. Memory and Cognition, 7, 263272. Taft, M. (1981). Prex stripping revisited. Journal of Verbal Learning and Verbal Behavior, 20, 289297. Taft, M. (1985). The decoding of words in lexical access: A review of the morphographic approach. In D. Besner, T. G. Waller, & G. E. Mackinnon (Eds.), Reading research: Vol. 5. Advances in theory and practice (pp. 83123). New York: Academic Press. Taft, M. (1988). A morphologicaldecompositional model of lexical representation. Linguistics, 26, 657 667. Taft, M. (1994). Interactive-activation as a framework for understanding morphological processing. Language and Cognitive Processes, 9, 271294. Taft, M., & Forster, K. I. (1975). Lexical storage and retrieval of prexed words. Journal of Verbal Learning and Verbal Behavior, 14, 638647. Taft, M., Hambly, G., & Kinoshita, S. (1986). Visual and auditory recognition of prexed words. Quarterly Journal of Experimental Psychology, 38A, 351366. Tyler, L. K. (1984). The structure of the initial cohort: Evidence from gating. Perception & Psychophysics, 36, 417427.
461
Tyler, L. K., & Marslen-Wilson, W. D. (1986). The effects of context on the recognition of polymorphemic words. Journal of Memory and Language, 25, 741 752. Tyler, L. K., Marslen-Wilson, W. D., Rentoul, J., & Hanney, P. (1988). Continuous and discontinuous access in spoken word-recognition: The role of derivational prexes. Journal of Memory and Language, 27, 368 381. Tyler, L. K., & Wessels, J. (1983). Quantifying contextual contributions to word-recognition processes. Perception & Psychophysics, 34, 409420. Tyler, L. K., & Wessels, J. (1985). Is gating an on-line task? Evidence from naming latency data. Perception & Psychophysics, 38, 217222. van Orden, G. C., Pennington, B. F., & Stone, G. O. (1990). Word identication in reading and the promise of subsymbolic psycholinguistics. Psychological Review, 97, 488522. Websters third new international dictionary of the English language, unabridged. (1993). Springeld, MA: Merriam-Webster. (Received January 13, 1997) (Revision received May 15, 1997)

Wurm (1997)

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wurm (1997)

Uploaded by

Copyright:

Available Formats

JOURNAL OF MEMORY AND LANGUAGE ARTICLE NO.

37, 438461 (1997)

Auditory Processing of Prexed English Words Is Both Continuous and Decompositional

AUDITORY PROCESSING OF PREFIXED WORDS

AUDITORY PROCESSING OF PREFIXED WORDS

AUDITORY PROCESSING OF PREFIXED WORDS

FREQUENCY MEASURES M (SD) Range

0195 0529 0807 01786 02379 121963

AUDITORY PROCESSING OF PREFIXED WORDS

LEE H. WURM TABLE 2 RESULTS

FREQUENCY ANALYSES (EXPERIMENT 1) Sign F(1, 1212)

AUDITORY PROCESSING OF PREFIXED WORDS

AUDITORY PROCESSING OF PREFIXED WORDS

First 637 603 489 656

Second 587 607 486 684

FREQUENCY ANALYSES (EXPERIMENT 2) Sign F(1, 5332)

9.72* 36.06*** 48.43*** 26.40*** 30.96*** 30.59***

AUDITORY PROCESSING OF PREFIXED WORDS

AUDITORY PROCESSING OF PREFIXED WORDS

AUDITORY PROCESSING OF PREFIXED WORDS

AUDITORY PROCESSING OF PREFIXED WORDS

Judged prexednessa 8 7 1 5 8 8 8 6 8 7 8 8 8 8 7 7 6 5 3 5.5 7 8 8 8 8 7 8 8 8

Semantic transparencya 7 6 4 1 3 1 7 7 7 2 7 7 6.5 7 2 7 7 3.5 6 5 6 8 8 8 8 6 7.5 8 8

AUDITORY PROCESSING OF PREFIXED WORDS

APPENDIX B Neighborhood Size, Prex Likelihood, and Prex Frequency

You might also like

9.72* 36.06* 48.43* 26.40* 30.96* 30.59***