Effect of onset cluster complexity in speeded naming: A test of rule-based approaches

Journal of Experimental Psychology:Human Perception and Performance1999, Vol. 25, No. 2,361-375

Copyright 1999 by the American Psychological Association, Inc.0096-1523/99/S3.00

Effect of Onset Cluster Complexity in Speeded Naming:A Test of Rule-Based Approaches

Alan H. KawamotoUniversity of California, Santa Cruz

Christopher T. KelloCarnegie Mellon University

Undergraduates participated in 3 speeded naming experiments investigating the effect of onsetcluster complexity on response latency. Words with complex onsets (e.g., spin) had shorterresponse latencies than words with simple onsets (e.g., sin), despite the fact that words withcomplex onsets had more letters and phonemes but fewer neighbors, properties previouslyfound to increase naming latency. Moreover, the magnitude of the effect depended on theparticular complex onset. These onset complexity effects can be explained by the constraintimposed by the 2nd letter on the 1st letter and 1st phoneme for words with an onset. Thisconstraint ultimately arises because phonemes increase in sonority from the beginning of thesyllable to the nucleus. Dual-route models cannot account for these results, but analogy andparallel distributed models can, if the criterion to initiate articulation is based on the initialphoneme.

For a wide range of stimuli, studies have shown thatcomplex stimuli require more processing than simple stimuli.For example, sentences that are syntactically complexrequire more time to comprehend than sentences that aresimple (Ferreira, Henderson, Anes, Weeks, & MacFarlane,1996; Forster, 1970), and visual stimuli that are morecomplex take more time to rotate mentally than simplestimuli (Bethell-Fox & Shepard, 1988; Folk & Luce, 1987).

In this research, we determined whether the complexity ofthe onset of a monosyllabic word would affect how quickly aword is responded to in speeded naming tasks by comparingthe response latencies of words such as spin that havecomplex onsets (i.e., onsets comprised of consonant clus-ters) with words such as sin that have simple onsets (i.e.,onsets comprised of singleton consonants). We were inter-ested in cluster complexity, not grapheme complexity. Thus,the stimuli included words such as spin, in which the onsetcomprises two simple graphemes, s and p, but not wordssuch as shin, in which the onset comprises a single complexgrapheme sh. We made this decision to avoid the addedcomplication that is involved in correctly segmenting acomplex grapheme (Coltheart, 1985; Henderson, 1982).

Alan H. Kawamoto, Department of Psychology, University ofCalifornia, Santa Cruz; Christopher T. Kello, Department ofPsychology, Carnegie Mellon University.

This research was supported in part by grants from the SocialSciences Division and the Committee on Research of the Univer-sity of California, Santa Cruz. Portions of this article werepresented at the meeting of the Psychonomic Society, Philadelphia,November 1997.

We thank Kimberly Lloyd, Kimberly Stienke, and Dov Andrewsfor assistance in organizing the participants in the experiments. Wealso thank Max Coltheart and Guy Van Orden for their commentson earlier versions of this article.

Correspondence concerning this article should be addressed toAlan H. Kawamoto, Department of Psychology, University ofCalifornia, Santa Cruz, California 95064. Electronic mail may besent to [email protected].

We examined how onset complexity effects might ariseduring processing, particularly during those stages involvedin how the pronunciation of a word is determined. To focusmore clearly on the issues, we examined rule-based schemesof dual-route models of word naming in the section below.We chose to focus on dual-route models, not because weendorse this class of models but because the predictions of aprevious study of onset complexity by Frederiksen and Kroll(1976) assumed a dual-route model. We then discuss articu-latory effects of the initial phoneme, especially as they affecthow response latency is measured.

Predictions Made by Rule-Based Schemes

All dual-route models assume two routes to pronuncia-tion, a rule route by which the pronunciation of a word isassembled using spelling-to-sound rules and a lexical routeby which a pronunciation is simply accessed from apreviously stored database (Baron & Strawson, 1976; Brown,1987; Coltheart, 1978; Forster & Chambers, 1973; Frederik-sen & Kroll, 1976; Kay & Bishop, 1987; Norris & Brown,1985; Patterson & Morton, 1985; Shallice, Warrington, &McCarthy, 1983). Within the class of dual-route models,specific models might logically differ along the followingdimensions. First, the units used as the basis of the rules candiffer in size (e.g., grapheme, cluster, body). In some cases,these units may overlap spatially with each other such thatsome letters might be part of more than one rule (e.g., bothgrapheme and cluster for words with complex onsets).Second, the different parts of a word may be processedsequentially from left to right or in parallel. Third, the rulesmay be retrieved and applied without any particular order-ing, probabilistically, or in a particular order. This orderingmight be determined by frequency, by complexity of theinput, or by the size of the input unit. Fourth, rules may ormay not be graded in strength. Fifth, the lexical route cancorrespond to a serial search or a process of analogy. Sixth,the outputs of the rule and lexical routes may be independent

361

362 KAWAMOTO AND KELLO

(e.g., race models) or may be combined. Seventh, thecriterion to initiate a pronunciation can be based on differentunits: the initial phoneme, the onset, the syllable, or thewhole word.

Given the number of dimensions in which dual-routemodels can differ, it is somewhat daunting to examinepredictions made by all possible variants. Instead, wenarrowed our discussion by examining only some of thedifferences arising from the rule route. In particular, weexamined the size of the input (grapheme vs. cluster), thetime it takes to apply a rule (no difference or increasing timewith increasing complexity), and the order of rule applica-tion (sequentially from left to right or in parallel). Inaddition, we examined assumptions about the criterion toinitiate articulation (initial phoneme, onset, syllable orword). These particular dimensions were singled out be-cause specific assumptions along these dimensions weremade by Frederiksen and Kroll (1976) in a study of onsetcomplexity.

Before continuing, we note that the specific assumptionmade by Frederiksen and Kroll (1976) regarding the crite-rion to initiate pronunciation is based on the onset + voweland is thus not explicitly represented here. In particular,Frederiksen and Kroll (1976) stated that "the articulation ofan initial consonant cannot be undertaken without transla-tion of the vocalic element to follow" (p. 369). We did notinclude predictions made by that criterion here because thepredictions made are identical to those of the simpler onsetcriterion. However, there is an important distinction betweenthe onset (as well as initial phoneme) criterion and theonset + vowel (as well as whole word) criterion; the onset +

vowel criterion allows coarticulatory constraints of thevowel to affect the beginning of the onset, whereas the onsetcriterion requires coarticulatory constraints only at the endof the onset. There is, in fact, a great deal of evidence ofcoarticulatory effects of the vowel on articulation of thepreceding consonants in normal speech (Amerman, Da-niloff, & Moll, 1970; Daniloff & Moll, 1968). However, ithas been argued elsewhere that articulation in the speedednaming task may be different from normal speech produc-tion (Kawamoto, Kello, Jones, & Bame, 1998). In particular,if participants interpret the instructions to respond as quicklyas possible to mean as soon as possible after the stimulusappears, then coarticulatory effects of the vowel may notnecessarily extend to the beginning of a preceding conso-nant, although coarticulatory effects should, of course, befound at the consonant-vowel boundary.

We now turn to specific predictions of the rule routeprocessing times for each possible combination of assump-tions for four types of words comprised of a consonant (C)or consonants (CC) that precede and follow a vowel (V):CVC, CVCC, CCVC, and CCVCC (see Table 1). Note thatthese predictions are specific to the rule route in the wordnaming process and thus do not reflect possible differencesin processing before that stage (e.g., letter recognition) orafter that stage (e.g., integration with the lexical route).

In the first set of predictions, the rules are based ongraphemes. Because complex graphemes are being ex-cluded, there is no difference in the length of time to retrieveand apply a particular rule. Moreover, there is no differencein the time to retrieve and apply the same rule for simpleonsets and complex onsets. That is, it takes the same amount

Table 1Response Latency Predictions

Input Processing time

Grapheme

Clusters Simple = complex

Simple < complex

Criter-ion

IP

ON

SY

IP

ON

SY

IP

ON

SY

Rule ap-plication

SerialParallel

SerialParallel

SerialParallel

SerialParallel

SerialParallel

SerialParallel

SerialParallel

SerialParallel

SerialParallel

Predictions

n.d.n.d.

CVC = CVCC < CCVC = CCVCCCVC = CVCC < CCVC = CCVCC

CVC < CVCC = CCVC < CCVCCCVC < CVCC = CCVC < CCVCC

n.d.n.d.

n.d.n.d.

n.d.n.d.



CVC < CVCC = CCVC < CCVCCCVC < CVCC = CCVC < CCVCC

Note. £ is = if processing is not variable, < if processing is variable (determined by last to finish),n.d. indicates no difference (i.e., CVC = CVCC = CVCC = CCVCC). IP, ON, and SY correspondto initial phoneme, onset, and syllable, respectively. C = consonant; V = vowel.

ONSET CLUSTER COMPLEXITY 363

of time to retrieve and apply the rule for 5 in sin and spin.Thus, if the criterion to initiate articulation is based on theinitial phoneme, the processing times based on the rule routewould be identical for all words. This prediction holds forboth sequential and parallel application of rules because theinitial phoneme alone determines when articulation is initi-ated. Next, the predictions assume that the criterion toinitiate articulation is based on the onset. If rules are appliedsequentially, words with complex onsets have longer re-sponse latencies than words with simple onsets becausemore rules need to be applied for complex onsets thansimple onsets. However, if rules are applied in parallel, thenthe prediction depends on whether the rule retrieval andapplication times are variable. In particular, if rule retrievaland application are variable (i.e., there is a distribution ofretrieval and application times for a particular rule), thencomplex onsets would have longer response latencies thansimple onsets because it is the last rule to finish thatdetermines when articulation is initiated. Of course, if ruleretrieval and application are not variable (i.e., it always takesexactly the same amount of time to retrieve and apply allrules), then there would be no effect of onset complexity.Finally, the last set of predictions assumes that the criterionto initiate articulation is based on the syllable. If rules areapplied sequentially, then words with more graphemesrequire more rules and thus have longer response latencies.If rules are applied in parallel, then the predictions againdepend on whether rule retrieval and application tunes arevariable. If the times are variable, then the more graphemes aword has the longer the response latency because it is the lastrule to be retrieved and applied that determines responselatency. If rules always take exactly the same time, then nodifferences would be found.

In the next set of predictions, the rules are based onclusters. Because all the stimuli have an onset, a nucleus,and a coda, all stimuli have three clusters. If complexclusters are processed as quickly as simple clusters, thepredictions are simple: No differences across conditionswould be found for any of the criteria and for either serial orparallel processing of the rules because all four word typeshave the same number of clusters and thus all require thesame number of rules to be applied. However, if rulesinvolving simple clusters are retrieved and applied morequickly than rules involving complex clusters, differentpredictions arise. For the initial phoneme and the onsetcriteria, words with simple onsets are responded to morequickly than words with complex onsets under serial orparallel rule application. For the syllable criterion, however,serial and parallel rule application may have slightly differ-ent predictions. Under serial rule application, words with nocomplex clusters are the fastest, words with a singlecomplex cluster are slower, and words with two complexclusters are the slowest. Under parallel application, predic-tions depend on whether the rule retrieval and applicationtimes are variable. If times are variable, predictions areidentical to the serial rule application assumptions, but iftimes are not variable, then rule application times for wordsthat have only simple clusters are the fastest, with the timesfor all other words being identical.

As can be seen in Table 1, for the 18 possible combina-tions of assumptions, there are only three distinct predic-tions: (a) no difference, (b) words with simple onsets beingfaster than words with complex onsets, and (c) the morecomplex clusters a word has the longer the response timewill be. Admittedly, more than one set of assumptions givesrise to each of the three predictions. In fact, for onedimension, sequential versus parallel order in rule applica-tion, there are no differential predictions if rule retrieval andapplication times are variable, despite the fact that Frederik-sen and Kroll (1976) assumed that serial processing wasnecessary to account for their results. Thus, one cannot hopefor empirical tests to narrow the choices to a single set ofassumptions, although some of the possibilities could beeliminated.

Given these clear alternatives, one might ask whetherthere is already evidence that can be used to address theissue of onset complexity on response latency. To ourknowledge, there has been only one previous study that ispotentially relevant to the issue of onset complexity. In thatstudy, Frederiksen and Kroll (1976) reported that words andpseudowords with fewer letters in the onset had shorternaming latencies than words and pseudowords with moreletters in the onset when matched for overall length. Theeffects were large, on the order of 25-60 ms when onsetsdiffered in size by one letter and on the order of 115-170 mswhen the onsets differed in size by two letters. Many of thedifferences were statistically significant by items despite fewitems. Although these differences were large, the questionregarding the effect of onset complexity on response timeswas nonetheless left completely open because (a) thedifferences may have been due to words with complexgraphemes as the onset (e.g., shin) rather than words withcomplex onsets (e.g., spin) and (b) the naming latency dataprobably conflated response latency and initial segmentduration. In the following section, we argue that specificarticulatory and acoustic properties of initial consonants ofcomplex onsets render response latencies obtained by voicekeys invalid for some of the consonants and may renderthem invalid for the remaining consonants.

Problems With Naming Latency Data

One problem that arises in studying the effect of onsetcomplexity on naming times is the partial confounding ofonset complexity with acoustic and articulatory characteris-tics of the initial phoneme. In particular, word-initial simpleonsets of English words can correspond to any consonant inthe English language (with the exception of/rj/ and /3/), butcomplex onsets must begin with either a plosive (/p/, /t/, /k/,/b/, /d/, and /g/) or a voiceless fricative (/s/, /JV, /f/, /h/, and/0/). (We excluded words such as pew and mute because thesemivowel /j/ following the initial consonant arises only inone particular vowel context, and thus it is not clear whetherthe /j/ should be included with the onset or with the nucleus.)

This confounding of onset complexity and initial pho-neme is problematic because naming latencies obtained byvoice keys may not be valid measures of response latenciesfor exactly those consonants that begin complex onsets: the


plosives and voiceless fricatives (Kawamoto et al., 1998). Inparticular, the initial part of a word-initial plosive consonantis silent when a word beginning with a plosive is uttered inisolation as it is in the standard naming task. In fact, theonset of acoustic energy that could potentially trigger a voicekey corresponds to the end of the plosive and the beginningof the subsequent phoneme (a liquid, a semivowel, or avowel), especially for voiceless plosives. The situation forvoiceless fricatives is potentially better in that acousticenergy is generated soon after the response has beeninitiated. However, as many investigators have been awareof and as Sakuma, Fushimi, and Tatsumi (1997) havespecifically shown for Japanese CV syllables, voice keystend to miss the low-intensity, high-frequency energy ofvoiceless fricatives and are triggered instead by the follow-ing vowel. Thus, for both plosives and voiceless fricatives,naming latency could conflate response latency and initialsegment duration.

Most investigators, aware that acoustic characteristicsaffect when a voice key is triggered, match words indifferent conditions in terms of their initial phoneme whenusing the naming task (Glushko, 1979). Indeed, Frederiksenand Kroll (1976) also acknowledged the potential effect ofvarying initial phonemes on naming latencies measured byvoice keys. After first comparing all stimuli, they subse-quently compared only those stimuli that matched on theinitial phoneme. However, the matching did not affect theirconclusions because they again found longer naming laten-cies for words with more consonants preceding the vowel.

Unfortunately, matching initial phonemes does not pro-vide a complete solution because the second consonant itselfcan be a plosive for complex onsets beginning with s. Theproblem that arises is illustrated graphically in Figure 1, in

sit

slit

spit

split

Figure 1. Acoustic waveforms of the words sit, slit, spit, andsplit.

which the acoustic waveforms corresponding to sit, slit, spit,and split are depicted. For words with a simple onset such assit or a complex onset in which the second consonant is Isuch as in slit, the voice key might very well be triggered bythe beginning of the vowel or the liquid, respectively (i.e., atthe end rather than the beginning of the initial segment). Bycontrast, for words with a complex onset that begins with sand is followed by a voiceless plosive such as spit and split,the voice key might again be triggered by the vowel and theliquid, respectively (i.e., at the end of the second phoneme).Thus, a word such as spit might have a longer naminglatency than sit, not because it had a complex onset butbecause the voice key was not triggered until the vowel. Aparticularly compelling example of this point is illustratedby Sakuma et al. (1997, Figure 3), who showed that thevoice key is triggered by the vowel for an utterance thatbegins with [j"t], 237 ms after the onset of acoustic energy.

Because of the potential conflation of response latencyand segmental duration, we argue that the results reported byFrederiksen and Kroll (1976) may not be due to onsetcomplexity but instead may be due to the proportion of thestimuli in the complex onset condition that comprise an sfollowed by a plosive. Moreover, this problem would alsoaccount for the larger naming latency difference betweencomplex onsets comprised of three consonants and simpleonsets compared with the naming latency difference be-tween complex onsets comprised of two consonants andsimple onsets; the only complex onsets in English that arethree phonemes long all begin with /s/ followed by avoiceless plosive (/spl/, /spr/, /skr/, and /str/), whereascomplex onsets beginning with /s/ comprised of two pho-nemes are followed not only by voiceless plosives but alsoby voiced nonplosives (/si/, /sm/, /sn/, and /sw/).

Given the problems of the previously reported data andthe importance of the issue in adjudicating among assump-tions regarding the units on which rules are based as well asthe criterion used to initiate articulation, it is important toreexamine the issue of onset complexity on response timesin the naming task. To ensure more accurate measurementsof the onset of acoustic energy, especially of the low-intensity, high-frequency energy generated by voicelessfricatives, we digitized the verbal responses and used analgorithm sensitive to changes in energy at high frequencies.We also used the postvocalic naming task introducedpreviously (Kawamoto et al., 1998) to obtain more validmeasures of response latency for words beginning withplosives.

Experiment 1

In Experiment 1, we examined the effect of onsetcomplexity for words beginning with plosive consonants. Asalready noted, a voice key would be triggered by thesegment following a word-initial plosive (especially voice-less plosives), either a vowel or semivowel (e.g., pure) in thesimple onset case or a liquid in the complex onset case. If avoice key is triggered by a vowel, a semivowel, or a liquid, itwould seem that a comparison of words beginning withplosives, even with a voice key, would provide a valid test of


onset complexity. Unfortunately, it does not because thelatency of acoustic energy in the standard naming responseconflates response latency and initial segment duration forplosives. In particular, the use of acoustic latencies asmeasures of response latencies for pairs of words matchedon initial phoneme assumes that the initial segment dura-tions across conditions are identical.

However, there are two reasons why one might expect thatinitial segment durations are shorter in complex onsets thansimple onsets, which in turn would lead to shorter acousticlatencies in the standard naming task for complex onsetsthan simple onsets for words beginning with plosives ifresponse latencies were equal. First, it is well-known fromthe speech literature that a single segment is longer than thatsame segment in a cluster (Haggard, 1973; Klatt, 1974;Umeda, 1977). Second, in many models, vowels take moretime to process than consonants because the mapping ofvowels is more ambiguous than the mapping of consonants(Berent & Perfetti, 1995; Kawamoto, 1988). This possibilitycan arise because of a conflict in simultaneous application ofconflicting rules or conflicts between the rule and lexicalroutes, both of which are more common for vowels thanconsonants (Berndt, Reggia, & Mitchum, 1987). If this is thecase, the release of the plosive, which is coarticulated withthe following segment, would occur later for simple onsetsthan complex onsets because the initial consonant of asimple onset is followed by a vowel, whereas the initialconsonant of a complex onset is followed by a consonant (aliquid).

Obtaining a valid measure of response latency is thuscritical for words beginning with plosives. To do so, we usedthe postvocalic naming task introduced by Kawamoto et al.(1998). In this task, participants begin to say "uuhhh"before they see the target and continue saying "uuhhh" up tothe time they begin articulating the target. Because theplosive is initially silent, the closure onset (i.e., the transitionfrom the vocalization of the "uuhhh" to the silence corre-sponding to the beginning of the plosive) is used as ouroperational definition of the beginning of the word. Thus,this vocalic offset latency corresponds to response latency.Of course, if one defines response latency as the point whenthe articulators move to the target position of the initialphoneme (evidenced by the decrease in the amplitude of theinitial vowel in the postvocalic naming task), there will stillbe a short delay between the true response latency and ouracoustically based measure of response latency.

All the targets were monosyllabic English words. Wordswith complex graphemes such as sh and kn were notincluded because these graphemes must be segmentedcorrectly to generate the correct pronunciation. Both high-and low-frequency words were included to determine theeffect of the lexical route. Because low-frequency words areaffected by the rule route more than high-frequency words,one would expect a bigger effect for low- than high-frequency words. In addition, we included only regularmonosyllabic words because many previous studies haveshown an effect of regularity on naming latency (seeSeidenberg & McClelland, 1989).

Method

Participants. Twenty-seven University of California, SantaCruz, undergraduates participated in the experiment to fulfill apsychology class requirement. All participants were native Englishspeakers who did not know the purpose of the experiment and hadnormal or corrected-to-normal vision.

Stimuli. The stimuli consisted of 30 pairs of high- and low-frequency monosyllabic words that had regular pronunciations andthat began with a plosive consonant (/p/, A/, /k/, /b/, /d/, or /g/). Thehigh- and low-frequency pairs of words were matched on the basisof the onset and nucleus, and, in most cases, the letters constitutingthe initial consonant cluster and vowels were also matched for apair, so that stimuli were also matched on grapheme-to-phonemecorrespondence rules. Half the pairs were made up of words withsimple onsets, and the other half were made up of words withcomplex onsets. The pairs of words that were selected were chosenso that printed word frequency (Kucera & Francis, 1967) wascomparable across the onset complexity condition. Furthermore,none of the stimuli had the same body (i.e., had the same vowel andfinal consonant cluster), nor did any of the stimuli rhyme withanother stimulus. However, words in the complex onset conditionwere significantly longer in terms of the number of letters,F(l, 56) = 43.56, p < .001, and in terms of the number ofphonemes, F(l, 56) = 44.64, p < .001, and had significantly fewerneighbors (based on the abridged Medical Research Council's[MRC] database), F(l, 58) = 52.54, p< .001.

There were also 27 filler trials consisting of monosyllabic words,all of which had a simple or complex onset that began with anonplosive consonant. None of the fillers had the same body as anyof the test stimuli. The first 2 trials presented to the participant werealways filler trials. There were also 15 practice stimuli similar incomposition to the test and filler trials. None of the practice stimulihad the same body as any of the experimental words.

Apparatus. The experiment was controlled by an IBM PC-compatible Pentium computer. A 2-s, 11,025-Hz digital sample ofeach naming response was recorded using a 16-bit SoundBlastersound card (Creative Labs, Milpitas, CA) and a Sennheisersupercardioid headset microphone (Sennheiser Electronic Corpora-tion, Wedeemark, Germany). The SoundBlaster started digitizingfrom the microphone at the onset of the target stimulus on themonitor, and it continued digitizing for 2 s. The audio sample wasthen stored on disk for off-line analyses, as described in the Resultssection.

Procedure. When participants arrived, the experimenter askedthem to read a set of instructions presented on the computer screenand answered any questions they had. These instructions indicatedthat a word would be presented on the screen that was to be readquickly and accurately. Participants were also instructed to say"uuhhh" when they initiated a trial and to continue saying"uuhhh" until they named the target. Each participant then put onthe headset microphone and began the practice trials, during andafter which any further questions were addressed. The experi-menter then started the experimental trials and left the room.

Each trial proceeded as follows: A "Ready? (Remember to say'Uuhhh')" prompt appeared in the center of the screen, and theparticipants pressed a button to start the trial. A centered asteriskthen replaced the Ready? prompt for 550 ms, after which the trialword appeared in lowercase letters for 2 s, replacing the asterisk.The Ready? prompt then again appeared for the next trial.

Results

Two independent judges listened to the naming responsesand marked pronunciation errors, which included failures to


say "uuhhh," mispronunciations, incorrect word pronuncia-tions, multiple responses, and stutters. The single inconsis-tency in the error coding of the two judges was resolved. Thedata from 3 participants were not included because of higherror rates or failure to follow instructions. The erroneousresponses (50 in all) of the remaining participants wereremoved and analyzed separately. Only correct responseswere used in determining response latency.

In the postvocalic naming task, response latency forplosives corresponded to the vocalic offset latency (i.e., theend of the "uuhhh") and was determined by the followingalgorithm described in detail by Kello and Kawamoto(1998). A 10-ms "analysis window" started at the beginningof the recording and calculated the mean absolute amplitudeand number of sign changes in the slope of the wave formwithin the window. If both measures fell below a predeter-mined threshold, the algorithm then recursively split andreanalyzed the window to more precisely locate the offset ofacoustic energy. Otherwise, the algorithm moved the win-dow forward through the recording and recalculated thesignal-detection measures. We determined the thresholdvalues by trial and error. We guessed at initial values for thethresholds, tested the algorithm on some sample data,examined a graphic display of the wave form, adjusted thethresholds, and repeated the process until we were satisfiedwith the estimates determined by the algorithm. Note thatthese thresholds were set high enough so that the low-intensity voicing during the closure of voiced stops thatsometimes occurred (especially for male participants) al-most always fell below the threshold.

A two-way analysis of variance (ANOVA) with frequencyand onset complexity as the two factors and vocalic offsetlatency as the dependent measure was carried out. Thefactors were within participants but between items, andanalyses with both participants and items as the randomfactors were carried out. Sixty-one responses with latenciesless than 200 ms or greater than 1,000 ms were also excludedin the latency analysis. The mean vocalic offset latencies andpercentage of errors for each condition, together with theirrespective standard errors, are shown in Table 2, and themeans for each item are shown in Appendix A. The maineffect of frequency was significant by participants and byitems, with low-frequency words 9 ms slower than high-frequency words: F(l, 23) = 8.41, p < .01, by participants;F(l, 56) = 4.92, p < .05, by items. Words with complex

Table 2Mean Vocalic Offset Latencies and Percentage of ErrorsBased on Participant Means in Experiment 1

Latency % error

Category M SE M SE

High frequencySimpleComplex

Low frequencySimpleComplex

339339

351 '346

8.39.6

9.39.6

4.21.7

5.03.1

1.20.7

1.20.7

onsets had latencies that were 3 ms less than words withsimple onsets, but this effect was not significant (Fs < 1 byparticipants and items). The effect of onset complexity wasslightly larger for low- than high-frequency words, but theinteraction was not reliable (Fs < 1 by participants anditems).

The percentages of errors and their standard errorsgrouped by experimental condition are also shown in Table2. We conducted a two-way ANOVA with frequency andonset complexity as the experimental factors and percentageof errors as the dependent measure. The main effect offrequency was not significant: F(l, 23) = 2.42, p > .1, byparticipants; F(l, 56) < 1 by items. However, there weremore errors made to words with simple onsets than to wordswith complex onsets. This main effect of onset complexitywas significant by participants, F(l, 23) = 4.6, p < .05, andmarginally significant by items, F(l, 56) = 3.80, p < .06.The interaction between frequency and onset complexitywas not significant (F < 1 by participants and items).

Discussion

The only response latency effect that was significant was amain effect of frequency. This result replicates numerousprevious results demonstrating faster response times forhigh- than low-frequency words in the standard naming task(Forster & Chambers, 1973; Frederiksen & Kroll, 1976;Taraban & McClelland, 1987) and in the postvocalic namingtask (Kawamoto, Kello, Higareda, & Vu, 1999).

The main effect of onset complexity on response latency,the effect of primary interest, was not significant. Interest-ingly, however, the response latency of words with complexonsets was slightly shorter than the response latency ofwords with simple onsets, a pattern opposite that found byFrederiksen and Kroll (1976) and opposite that predicted bymany different sets of assumptions enumerated in theintroduction. The effect was nonexistent for high-frequencywords as predicted if the rule route were primarily respon-sible for the effect, but this interaction between frequencyand onset complexity was not significant.

One possible reason for the results contradicting thepredictions laid out in the introduction is that the pronuncia-tion of words with c and g as the onset is conditioned by thefollowing vowel (compare call vs. cell and gust vs. gist).However, if the vowel information is not included in therule, the mapping would be ambiguous and a longerresponse latency would be predicted if one assumes probabi-listic application of rules or simultaneous application of allrules for words with simple onsets. Even if vowel informa-tion is included, problems can still arise (compare gist andgift). By contrast, c and g as the initial letters of a complexonset are not ambiguous. Thus, at least for rules based onclusters (see Table 1), there may very well be a specificeffect attributable to the multiple mappings of these particu-lar consonants.

In Experiment 1, all the stimuli began with plosiveconsonants. However, complex onsets can also begin withvoiceless fricatives. To ensure that the onset complexityresults found in this experiment were not due to an


idiosyncratic property of plosives or to the pronunciation ofsimple onsets corresponding to c or g, we conducted anotherexperiment using a different initial consonant, s. The use ofanother class of complex onsets is particularly importantbecause Stemberger and Treiman (1986) found differencesin the kind of speech errors that were made depending onwhether complex onsets did or did not begin with /s/. Inparticular, they found that for errors involving the loss oraddition of a segment in the onset, clusters beginning with/s/ were much more likely to involve the first consonant thanclusters not beginning with /s/.

Experiment 2

In this experiment, the onsets of stimuli with complexonsets consisted of the voiceless sibilant /s/ followed by avoiceless plosive. To ensure the accurate detection of theonset of acoustic energy, we digitized and stored the verbalresponses and used an algorithm sensitive to low-intensity,high-frequency acoustic energy.

Method

Participants. Thirty-six University of California, Santa Cruz,undergraduates participated in the experiment to fulfill a psychol-ogy class requirement. All participants were native English speak-ers who did not know the purpose of the experiment and hadnormal or corrected-to-normal vision.

Stimuli. The stimuli consisted of 20 pairs of high- and low-frequency monosyllabic words that began with 5 and that hadregular pronunciations. Members of each pair of high- andlow-frequency words were matched on the basis of the onset andthe nucleus. Half the pairs of words had simple onsets and the otherhalf had complex onsets (s followed by p or t). The pairs of wordswere chosen so that printed word frequency (Kucera & Francis,1967) was comparable across the onset complexity condition.Furthermore, none of the stimuli had the same body, nor did any ofthe stimuli rhyme with another stimulus. However, words in thecomplex onset condition were significantly longer in terms ofthe number of letters, F(l, 36) = 15.76,;? < .001, and in terms ofthe number of phonemes, F(l, 36) = 33.92, p < .001, and hadsignificantly fewer neighbors (based on the abridged MRC data-base), F(l, 38) = 10.53,p < .005.

There were also 44 filler trials consisting of monosyllabic words,all of which had a simple or complex onset that began with plosiveand nonplosive consonants. None of the fillers had the same bodyas any of the test stimuli. The first 2 trials presented to theparticipant were always filler trials. There were also 10 practicestimuli similar in composition to the test and filler trials. None ofthe practice stimuli had the same body as any of the experimentalwords.

Apparatus. The apparatus used in Experiment 1 was used inthis experiment.

Procedure. The procedure was similar to that in Experiment 1except that the standard naming task instead of the postvocalicnaming task was used.

Results

Two independent judges listened to the naming responsesand marked pronunciation errors, which included mispronun-ciations, incorrect word pronunciations, multiple responses,

and stutters. There were no inconsistencies in the errorcoding. The erroneous responses (19 in all) were removedand analyzed separately. Only correct responses were usedin determining response latency.

To determine the acoustic onset of each correct namingresponse, we applied an "audio signal finding" algorithm tothe digitized response (Kello & Kawamoto, 1998). Thealgorithm to detect the onset of acoustic energy in thestandard naming task was similar to the algorithm to detectthe offset of acoustic energy in the postvocalic naming task,except that it was the presence rather than the absence ofacoustic energy that was being searched for.

A two-way ANOVA with frequency and onset complexityas the two factors and acoustic onset latency as thedependent measure was carried out. The factors were withinparticipants but between items, and analyses with bothparticipants and items as the random factors were carriedout. Twenty-two responses with latencies less than 200 msor greater than 1,000 ms were also excluded in the responselatency analysis. The mean acoustic latencies and percent-age of errors for each condition, together with their respec-tive standard errors, are shown in Table 3, and the means foreach item are shown in Appendix B. Both main effects weresignificant by participants and by items, with the response tolow-frequency words 16 ms slower than high-frequencywords: F(l, 35) = 18.60, p < .001, by participants;F(l, 36) = 10.04, p < .005, by items. Responses to wordswith complex onsets were 16 ms faster than words withsimple onsets: F(l, 35) = 13.28, p < .001, by participants;F(l, 36) = 11.74, p < .005, by items. The interaction wasnot reliable (Fs < 1 by participants and by items).

The percentage of errors and their standard errors groupedby experimental condition are also shown in Table 3. Weconducted a two-way ANOVA with frequency and onsetcomplexity as the experimental factors and percentage oferrors as the dependent measure. No effect was significant(Fs < 1 by participants and by items).

Discussion

In this experiment, words with complex onsets hadresponse latencies that were significantly shorter than wordswith simple onsets. This effect was slightly larger for low-than high-frequency words, but, as in Experiment 1, theinteraction was not significant.

Table 3Mean Acoustic Latencies and Percentage of Errors Basedon Participant Means in Experiment 2

Latency > error

Category M SE M SE

High frequencySimpleComplex

Low frequencySimpleComplex

456440

475454

17.714.8

17.516.1

1.11.4

1.71.1

0.70.7

0.70.5


One possible explanation for the result in this experimentis that rules involving consonant clusters were processedfaster than rules involving single consonants, not slower orat the same rate, as assumed in the introduction. However,this choice was arbitrary and lacked any independentmotivation. Moreover, any model that ascribes the effect torule complexity cannot account for the difference in theoutcome between Experiments 1 and 2. Of course, onemight attribute the difference in the outcome of Experiments1 and 2 to the task (postvocalic naming vs. standard naming)or to the type of complex onset (plosive-liquid vs. /s/-plosive). For example, there may be differences in how aninitial plosive consonant is articulated compared with aninitial sibilant. In Experiment 3, we examined the composi-tion of complex onsets and a different account of the data.

Experiment 3

We propose that the shorter latencies of complex onsetscompared with simple onsets and the difference in themagnitude of this effect depending on the particular complexonset is due to conditional probability (i.e., the probability ofa particular initial consonant conditioned on the secondletter). This conditional probability is itself due to phonotac-tic constraints that arise because phonemes increase insonority toward the nucleus of a syllable (Jespersen, 1904)and because of the quasi-regular mapping from letters tophonemes and from phonemes to letters. For example, theconditional probability of s as the first letter of the onsetgiven p as the second letter of the onset is equal to 1.0because a p can only be preceded by s. By contrast, theconditional probability of 5 given a vowel in the secondposition (i.e., a simple onset) is much less than 1.0 because agiven vowel can be preceded by virtually any consonantexcept perhaps q.

Instead of using conditional probability as a factor,however, we used a measure that is closely related to it: thenumber of alternatives that could possibly be the first letterof a word given a particular second letter of a word. Thismeasure led to the following ranking, from the fewestpossible letters to the most possible letters:

1. m, p, t, c, k (preceded by s)2. w (preceded by d, s, t)3. n (preceded by g, k, m, p, s)4. / (preceded by b, c,f, g, p, s)5. h (preceded by c, g, p, r, s, t, w)6. r (preceded by b, c, d,f, g, p, t, w)7. vowel (preceded by almost every consonant).

The conditional probability argument as well as thenumber of different alternative first letters argument pre-sented above make two assumptions. First, the input isprocessed in such a way that the second letter from the leftcan affect the processing of the first letter. This occurs if allthe letters are processed in parallel or if the letters areprocessed serially from left to right with a temporal overlapin the processing of the first and second letters. Second, theposition of a consonant in the input relative to the vowel isknown. (This assumption was also made by Plaut, McClel-land, Seidenberg, & Patterson, 1996, in the representation

that they used.) Although the consonant position assumptionmakes the constraint imposed particularly striking when/?, t,c, k, and m are the second letters because only a single letter,s, can precede those letters, the arguments are still valid ifconsonant position is not known. In particular, if vowels arealso allowed to precede a given letter, the relative order ofthe different classes of second letters defined by the numberof possible letters that could precede it will be identical to orat least very similar to the original ranking listed abovebecause almost every vowel can precede any given letter(e.g., am, aim, each, elk, it, opt, up). Consequently, if thenumber of possible letters that could precede a class ofletters is one of the variables in a linear regression, littledifference in the magnitude of the correlation coefficientwould be expected if vowels were or were not included aspossible initial letters.

If the constraints discussed above are used when readingaloud and if one assumes that articulation is initiated as soonas the initial phoneme becomes available, then responsetimes should decrease as the number of alternative initialletters decreases because there are fewer alternatives thatneed to be considered. Thus, the rank order indicated abovecorresponds to the ranking of predicted response timesordered from fastest to slowest. On this basis, words withcomplex onsets are predicted to have shorter responselatencies than words with simple onsets. Moreover, themagnitude of the onset complexity effect should decrease asthe difference in the ranking between a particular complexonset and a simple onset decreases. In particular, a smalldifference in response latency is predicted between complexonsets comprised of an / or r preceded by a plosivecompared with simple onsets, but a large difference inresponse latency is predicted between complex onsetscomprised of a plosive preceded by an s compared withsimple onsets, exactly the pattern of results obtained inExperiments 1 and 2.

There are, in fact, additional predictions that can be testedusing the data from Experiment 1 because / and r as thesecond consonant in the onset allow different numbers ofalternative first letters. We reanalyzed the data from Experi-ment 1 with the onset factor now expanded to separateplosives followed by / from plosives followed by r. Consis-tent with predictions based on the number of alternative firstletters, words with complex onsets consisting of plosivesfollowed by / had reliably shorter latencies than words withsimple onsets: F(\, 23) = 4.99, p < .05, by participants;F(l, 34) = 4.75, p < .05, by items. Words with complexonsets consisting of plosives followed by / had reliablyshorter latencies than words with complex onsets consistingof plosives followed by r: F(l, 23) = 10.21, p < .005, byparticipants; F(l, 34) = 5.06, p < .05, by items. Inconsistentwith the number of alternatives predictions, complex onsetsconsisting of plosives followed by r and words with simpleonsets had comparable latencies (Fs < 1 by participants anditems). However, this last inconsistent result was not toosurprising because of all the complex clusters, those with ras the second letter provide the least amount of constraintand thus are the least likely to be significantly faster thanwords with simple onsets.


Although conditional probability could account for part ofthe differences in the magnitude of the onset complexityeffect found in Experiments 1 and 2, the differences mightalso be attributed to composition of the complex clusters(plosive followed by liquid vs. /s/ followed by voicelessplosive) or to the task (postvocalic naming task vs. standardnaming task). Thus, it is important to carry out an experi-ment in which the initial phoneme is the same but theconsonant that follows it is different. Moreover, the secondconsonants and the vowel should provide as different aconstraint as possible but should be as similar phoneticallyto each other as is reasonably possible. We thus chose wordsthat begin with an s followed by a vowel or the consonants /and m as stimuli. Finally, only low-frequency stimuli wereincluded because there were few high-frequency wordsbeginning with sm.

Method

Participants. Twenty-six University of California, Santa Cruz,undergraduates participated in the experiment to fulfill a psychol-ogy class requirement. All participants were native English speak-ers who did not know the purpose of the experiment and hadnormal or corrected-to-normal vision.

Stimuli. The stimuli consisted of 30 low-frequency monosyl-labic words that began with the voiceless sibilant /s/. Ten words hadsimple onsets, another 10 had complex onsets comprised of si, anda final 10 had complex onsets comprised of sm. All the words hadregular pronunciations, and none of the stimuli had the same body,nor did any of the stimuli rhyme with another stimulus. Theretended to be more letters per word in the complex onset conditions,F(l, 28) = 2.81, p < .11, and there was significantly morephonemes per word in the complex onset conditions, F(l, 28) =5.79, p < .01. For both the number of letters and number ofphonemes, the rank order from fewest to most was simple, si, andsm onsets. There was also a significant difference in the number ofneighbors, F(2, 27) = 4.95, p < .05, with the number of neighborsincreasing in the order sm, si, and simple onsets.

There were also 108 filler trials consisting of monosyllabicwords, most of which had an onset. The onset of the fillers beganwith plosive and nonplosive consonants and were simple as well ascomplex. None of the fillers had the same body as any of the teststimuli. The first 2 trials presented to the participant were alwaysfiller trials. There were also 15 practice stimuli that were similar incomposition to the test and filler trials. None of the practice stimulihad the same body as any of the experimental words.

Apparatus. The apparatus used in Experiment 1 was used inthis experiment.

Procedure. The procedure was identical to that used inExperiment 2.

Results

Two independent judges listened to the naming responsesand marked pronunciation errors, which included mispronun-ciations, incorrect word pronunciations, multiple responses,and stutters. There were no inconsistencies in the errorcoding. Two participants' data were removed because ofhigh error rates or problems with the recording. Theerroneous responses (eight in all) of the remaining partici-pants were removed and analyzed separately. Only correctresponses were used in determining response latency. The

acoustic latencies were determined using the same algorithmused in Experiment 2.

A one-way ANOVA with onset type as the factor andacoustic latency as the dependent measure was carried out.This factor was within participants but between items, andparticipants and items were random factors. Five responseswith acoustic latencies less than 200 ms or greater than1,000 ms were also excluded from the acoustic latencyanalysis. The mean acoustic latencies and percentage oferrors for each condition, together with their respectivestandard errors, are shown in Table 4, and the means for eachitem are shown in Appendix C. The effect of onset type wassignificant by participants, F(2, 46) = 3.58, p < .05, butonly marginally significant by items, F(2, 27) = 2.93, p <.08. Planned comparisons showed that (a) words with sm asthe onset had shorter acoustic latencies than words withsimple onsets, F(l, 23) = 5.28, p < .05, by participants,F(l, 18) = 5.06,p < .05, by items; (b) words beginning withsm had slightly shorter acoustic latencies than words begin-ning with si, but this difference was only marginallysignificant by participants, F(l, 23) = 3.34, p < .09,F(l, 18) = 1.08, p > .3, by items; and (c) words with si asthe onset also had shorter acoustic latencies than words withsimple onsets, but this difference was not reliable, F(l, 23) <2, by participants, F(l, 18) = 2.12, p > .1, by items.

One reason why pairwise differences predicted by theconditional probability argument may not have been foundin all cases is that the differences are small. Because thenumber of possible items is also small, there may not havebeen sufficient power. Indeed, the only case in which thepredicted differences were found was for the case thatpredicted the largest difference (sm vs. simple onset). Giventhat the conditional probability argument predicts decreas-ing response latency with decreasing number of possibleinitial letters, a correlational analysis may in fact be moreappropriate. We therefore correlated acoustic latency of eachof the 30 items with the number of possible initial lettersgiven the second letter, assuming that words having m, I, or avowel as second letters had 1, 6, or 20 possible initial letters,respectively. There was a positive correlation (r = .42) thatwas significant, F(l, 28) = 5.87, p < .05. Note that theidentical correlation would have been obtained if we as-sumed that each letter (m, I, or a vowel) could also bepreceded by every vowel as well as the consonants indicatedearlier instead of just the consonants.

The percentage of errors and their standard errors groupedby experimental condition are also indicated in Table 4. Weconducted a one-way ANOVA with onset type as the

Table 4Mean Acoustic Latencies and Percentage of Errors Basedon Participant Means in Experiment 3

Latency error

Onset type M SE M SE

SimpleComplex: s/Complex: sm

469462448

24.020.918.3

3.62.10.0

1.61.50.0


experimental factor and percentage of errors as the depen-dent measure. There was an effect of onset type that wassignificant by participants, F(2,46) = 3.50, p < .05, but notby items, F(2,27) = 1.25, p > .3.

Discussion

In this experiment, the larger the conditional probabilityof the first letter given the second letter of a word (i.e., thefewer the number of possible first letters), the shorter theresponse latency. This argument accounts for the effect ofonset complexity as well as the difference in the magnitudeof the onset complexity effect for different complex onsetsfound in this experiment and the preceding experiments.

General Discussion

Our results indicate that monosyllabic words with com-plex onsets have shorter response latencies than words withsimple onsets and that the magnitude of the effect dependson the particular complex onset involved. The shorterresponse latency for words with complex onsets comparedwith simple onsets was found despite the fact that (a) wordswith complex onsets had more letters and more phonemesthan words with simple onsets, and naming latency gener-ally increases with word length (Butler & Mains, 1979;Cosky, 1976; Eriksen, Pollack, & Montague, 1970; Forster& Chambers, 1973; Frederiksen & Kroll, 1976; Jared &Seidenberg, 1990; Richardson, 1976; Treiman, Mullennix,Bijeljac-Babic, & Richmond-Welty, 1995; Weekes, 1997),and (b) words with complex onsets had fewer neighbors thanwords with simple onsets, and response latency generallyincreases as number of neighbors decreases (Andrews, 1989,1992; Peereman & Content, 1995).

We believe that the difference between simple andcomplex onsets as well as the difference in magnitude of theeffect depending on the particular complex onset can all beexplained by the conditional probability of the first lettergiven the second letter. That is, the second consonant in acomplex onset provides a great deal of information aboutwhat possible letters can occur in the first position, whereasa vowel does not. In fact, in some cases, there is only a singleconsonant that can occur. This constraint is presumed to bedue to sequencing constraints that ultimately arise becausesonority increases from the beginning and ending of asyllable toward the nucleus (Jespersen, 1904) and because ofthe quasi-regular mapping from phonemes to letters andfrom letters to phonemes.

In the discussion below, we discuss a number of addi-tional rule-based schemes, none of which can account for thedata. We then show how analogy models and paralleldistributed processing models can account for the data.

Alternative Rule-Based Accounts

In this section, we discuss alternative rule-based accounts.As in Table 1, we also include the effects of the criterion toinitiate pronunciation and the order of rule application onfour types of words: CVC, CVCC, CCVC, and CCVCC. The

first set of possibilities that we discuss is that the rules areordered by frequency rather than complexity, and thus morefrequent rules take less processing time because they areretrieved and applied more quickly. Note that the frequencythat we are considering is not the conventional use offrequency in discussions of rules in that frequency does notcorrespond to the relative frequency of a mapping of aparticular input to two or more outputs, but the frequency ofa rule relative to all other rules. We first examine rules thatare based on graphemes. Unlike in Table 1, where noordering of grapheme-based rules was made, frequencywould differentially order the rules. However, as we discussin detail below, the predictions that are made for frequency-based rules are virtually identical to those in which norule-ordering scheme is assumed.

Assuming an initial phoneme criterion, response latencydepends only on the time required for processing of theinitial letter to be completed. Thus, both serial and parallelrule application would predict no difference in processingtime because frequency of the initial letter is all that matters,and initial letter is matched across the four types of words.Assuming an onset criterion, response latency depends onlyon the time for processing of the second consonant to becompleted relative to the first consonant. Serial rule applica-tion predicts that simple onsets would be processed morequickly than complex onsets because the second consonantis processed after the first consonant. However, the predic-tion for parallel rule application depends on whether process-ing completion times are variable. In particular, if comple-tion times are not variable, completion times depend on therelative frequency of the first and second letters; if thesecond letter is as frequent or more frequent than the firstletter, simple and complex onsets will be processed equallyquickly, but if the second letter is less frequent than the firstletter, complex onsets will take more time than simpleonsets. By contrast, if completion times are variable, simpleonsets will always be processed more quickly than complexonsets no matter what the relative frequency of the first andsecond letters are because there will always be someoccasions on which the second consonant will reach thresh-old after the initial consonant.

Assuming a whole-word criterion, response latency de-pends only on the time for processing of the slowest letter tobe completed. Serial rule application would predict that theprocessing time increases as the number of letters increases.However, as in the onset criterion case, the predictions forparallel rule application depend on whether processingcompletion times are variable. In particular, if completiontimes are not variable, then completion times depend on therelative frequency of all the letters; if both letters of acomplex onset or coda are as frequent or more frequent thanthe single letter of a simple onset or coda, respectively, thensimple and complex onsets as well as simple and complexcodas will be processed equally quickly; otherwise, complexonsets and codas will take more time than simple onsets andcodas, respectively. (The preceding assumes that vowels aremore frequent than consonants; if the opposite is assumed,then all words would take the same time to be processedbecause it is the vowel that is the bottleneck.) By contrast, if


completion times are variable, then simple onsets andcodas will always be processed more quickly than complexonsets and codas because there will always be someoccasions on which the second consonant will reach thresh-old after the initial consonant. In summary, frequency ofgrapheme-based rules cannot account for the shorter re-sponse latencies of complex onsets compared to simpleonsets.

We next discuss rules that are ordered by frequency of thesimple and complex clusters. It turns out that the predictionsmade under this assumption are identical to the predictionsin Table 1 because simple onsets are more frequent thancomplex onsets (see Table 5).

Up to this point, we have assumed that rules do notoverlap spatially. That is, a given letter is used in either agrapheme-based rule or a cluster-based rule. (Note thatgraphemes and simple clusters were identical in our experi-ments.) However, if this restriction is relaxed, then addi-tional possibilities arise. One possibility is that rules basedon graphemes and clusters are used simultaneously. Thisassumption is simply a combination of the above twopossibilities, so clearly an advantage for complex onsetscannot be predicted. Another possibility is to assume thatcomplex clusters but not simple clusters participate in rules.Now, the initial phoneme of a simple onset is influencedonly by grapheme-based rules, whereas the initial phonemeof a complex onset is influenced by both grapheme-basedrules and onset-based rules. If the influence of grapheme-based rules and complex onset-based rules supports thesame phoneme as is the case for all the stimuli in this study,then the initial phoneme of complex onsets would beactivated more quickly than would the initial phoneme ofsimple onsets. Note that this account does not require anordering of the rules, only a combined influence of graph-eme and complex cluster rules. However, such an accountstill cannot explain the difference between different complexonsets. By again assuming some sort of frequency-basedordering of complex onset rules, one could in principlepredict differential response times. However, this assump-tion still cannot account for the pattern of differentialresponse times. In particular, words beginning with sm werefaster than words beginning with si in Experiment 3, but the

Table 5Frequencies (Based on Type Counts) of Mappings of Onsetsfor Monosyllabic Words in the Kucera andFrancis (1967) Database

Second letter

First letter Vowel I m p t

Plosivesb->/b/ 153 35 52c — /k/ 92 42 47d->IAl 102 28g->/8f 73 21 49p-*/p/ 123 29 33f-»/t/ 102 40

Nonplosives-»/s/ 114 38 14 40 63 14 20

number of words beginning with sm was less than thenumber of words beginning with si.

The last rule-ordering scheme is based on the order ofacquisition. However, the order of acquisition cannot ac-count for the advantage that complex onsets have becausechildren are taught rules based on single graphemes (i.e.,simple onsets) before rules based on multiple graphemes(i.e., blends) in phonics-based reading instruction. Thus,simple onsets would again be predicted to be processedmore quickly than complex onsets.

Before ending this section, we consider the possibilitythat articulation is initiated only after the onset and thevowel are known. One reason for considering this possibilityis that there is a great deal of evidence of coarticulatoryeffects of the vowel on articulation of preceding consonantsin normal speech production (Amerman et al., 1970; Da-niloff & Moll, 1968). Moreover, as mentioned in theintroduction, the onset plus vowel was the criterion assumedby Frederiksen and Kroll (1976) in their study. However,this criterion cannot account for the results either. If ruleswere applied serially, words with complex onsets shouldhave longer latencies than words with simple onsets becausethe vowel in words with a complex onset is the thirdphoneme of the word, whereas the vowel in words with asimple onset is the second phoneme. If the rules wereapplied in parallel, words with complex onsets should haveeither the same or longer response latencies than words withsimple onsets.

Thus, a number of different rule-based schemes cannotaccount for the results. There may very well be a set ofassumptions that will, but we have not been able to find sucha set of assumptions. Rather than continuing to discussadditional rule-based schemes, we instead turn our attentionto the two other classes of word naming models: paralleldistributed processing models and analogy models.

Parallel Distributed Processing and Analogy Models

In parallel distributed processing models of lexicalmemory, sublexical units representing the spelling, pronun-ciation, and meaning form connections to each other, eitherdirectly or mediated by hidden units (Kawamoto, 1988,1993; Plaut et al., 1996; Seidenberg & McClelland, 1989;Van Orden, Pennington, & Stone, 1990). The strengths of theconnections between units are modified to reflect the statisti-cal correlation in the activation of the units. In particular, theconstraints that a consonant in the second position has on theletter (and phoneme) in the first position that exists in thespelling and pronunciation of words can be learned. Thus,these connections reflect the constraint that an m in thesecond position predicts an s in the first-letter position andan /s/ in the first phoneme position, assuming that aconsonant is in the first position. (As noted earlier, therepresentation used by Plaut et al., 1996, allows thisunambiguous mapping.) By contrast, an / in the second-letter position is associated with a number of differentconsonants, one of which is s. Because the mapping with an Iin the second position is more variable than an m in thesecond position, there is a smaller constraint with / than m,


but the constraint with / is still greater than a vowel. Giventhese constraints, one would thus expect that activation ofthe initial phoneme in a complex onset should reachthreshold more quickly than that same phoneme in a simpleonset. Moreover, the difference between different complexonsets could also be accounted for because the biggestadvantage should arise in situations in which the mapping ofthe second letter to the first phoneme is the least ambiguous.Such a pattern would arise in fully recurrent networks wheretime can be indexed as processing cycles as well as instrictly feed-forward networks if processing is carried out in"cascade" (e.g., McClelland & Rumelhart, 1988). Thus, ifone assumes that articulation is initiated as soon as the initialphoneme becomes available (an assumption not necessarilymade in all parallel distributed processing models), then theentire pattern of results found in these experiments can beaccounted for.

Analogy-based naming models such as the conspiracymodel (Taraban & McClelland, 1987) and the phonologicalcompetition model (O'Seaghdha, Dell, Peterson, & Juliano,1992) should also be able to account for our results as longas one again assumes that articulation is initiated as soon asthe initial phoneme becomes available. For example, sparhas neighbors that all begin with s (e.g., star, spur, and span)because there are no other four-letter monosyllabic wordsthat end in par. Thus, all words that differ by one letter wouldactivate /s/ in the initial phoneme position. The one excep-tion to this generalization arises if the target has a neighborbeginning with sh (e.g., skip has ship as a neighbor). Thus,there can be at most one competitor to the /s/ in the initialposition. By contrast, a word such as slip having si as anonset would have neighbors such as flip and clip along withneighbors such as snip and slap. The response times forwords beginning with si would thus be expected to be slowerthan words beginning with an s followed by a plosivebecause there would be more competitors at the initialphoneme position. Finally, words with a simple onset suchas sand would have the neighbors band, hand, land, andwand. With a large number of competitors at the initialphoneme position, the response times for words with asimple onset would be expected to be slower than evenwords beginning with si.

Note that for Taraban and McClelland's (1987) con-spiracy model and the O'Seaghdha et al. (1992) phonologi-cal competition model, the representation of the pronuncia-tion is a flat representation without any units interveningbetween the word units and the phoneme units. However,there is a speech production model proposed by MacKay(1987) that is closely related to the word naming modelsdiscussed above in which a hierarchical representation isused. In particular, there are units representing syllables andsubsyllabic units such as the onset. It is important to notethat with respect to the issue considered in this research,there is an additional layer of units for complex onsetscompared with simple onsets. Because activation from aword node spreads along the leftmost branch to the firstphoneme of the word, it takes longer for the first phoneme of

a complex onset to be activated than the single phoneme of asimple onset to be activated in MacKay's model. If a hierarchi-cal representation such as that considered by MacKay hadbeen assumed by Taraban and McClelland and O'Seaghdhaet al., they would have predicted that words with complexonsets would have longer response latencies than those withsimple onsets. Our results are thus evidence for a flatrepresentation and against a hierarchical representation.

In the parallel distributed processing and analogy-basedaccounts of the data presented above, we assumed thatarticulation was initiated as soon as the initial phonemebecame available. This assumption is critical; if either anonset criterion or a syllable criterion is assumed, then eventhe parallel distributed processing and analogy-based ac-counts cannot account for the results. That is, all thephonemes in the complex onset must reach threshold beforearticulation can be initiated, whereas only the single pho-neme in the simple onset must reach threshold beforearticulation can be initiated. As was the case with the onsetand syllable criteria for rule-based approaches, words withcomplex onsets will have longer response latencies thanwords with simple onsets if there is variability in when aparticular phoneme is activated.

Other Stages in the Processing

Up to this point, we have assumed that any differences inresponse latency arise at a particular stage in the processing—the stage at which the pronunciation is determined. Ofcourse, effects could arise earlier, at the letter recognitionstage, or later, at the articulatory stage. As we argue, neitherof these possibilities can account for the data because effectsarising at these stages would probably lead to longer ratherthan shorter response latencies for complex than simpleonsets.

The first possibility is that the onset complexity effectsarise before the determination of the pronunciation (i.e., atthe letter recognition stage). We argue that this possibility isunlikely; the complex onset condition had more letters andthere was thus a greater chance of masking, interference, orlower attention for words with complex onsets than wordswith simple onsets.

The second possibility is that these effects arise after thestage when the pronunciation is generated. For example,differential response times might arise from differentialdifficulty in articulating simple versus complex onsets. Hereagain, however, we argue that this possibility would predictthat simple onsets should be faster than complex onsets. Asone index of articulatory difficulty, young children correctlyarticulate simple onsets before they correctly articulatecomplex onsets. In fact, complex onsets are initially pro-nounced as just a single consonant (Smith, 1973). Moreover,when multiple consonants are articulated, one consonant isoften incorrect (/bweoV for bread). If this difficulty observeddevelopmentally is manifested later as on-line productiondifficulty in adults, then words with complex onsets should


have longer response latencies than words with simpleonsets, not shorter latencies, as we found.

The third possibility is that these effects are due to somesort of response preparation after the entire pronunciationhas become available. However, this possibility cannotaccount for the data because responses that have morephonemes tend to take more time to initiate (Sternberg,Monsell, Knoll, & Wright, 1978), not less time, as we found.

Final Remarks

Complexity is almost always defined in terms of quantity.For example, the more letters, syllables, or morphemes aword has, the more complex the word is thought to be. Ifonly quantity is important, then it is easy to see why moreprocessing may be required. However, if the additionalcomplexity serves as a context, then there are cases in whicha more complex stimulus is beneficial. Perhaps the mostconvincing evidence of this possibility is the word superior-ity effect, in which identification of a letter is better in thecontext of a word than in isolation (Reicher, 1969). Thechallenge is to understand when added complexity is abenefit and not a liability and how to account for the effectwhen it arises.

References

Amerman, J. D., Daniloff, R., & Moll, K. L. (1970). Lip and jawcoarticulation for the phoneme /ae/. Journal of Speech andHearing Research, 13, 147-161.

Andrews, S. (1989). Frequency and neighborhood effects onlexical access: Activation or search? Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 15, 802-814.

Andrews, S. (1992). Frequency and neighborhood effects onlexical access: Lexical similarity or orthographic redundancy?Journal of Experimental Psychology: Learning, Memory, andCognition, 18, 234-254.

Baron, J., & Strawson, C. (1976). Use of orthographic andword-specific knowledge in reading words aloud. Journal ofExperimental Psychology: Human Perception and Performance,2, 386-393.

Berent, I., & Perfetti, C. A. (1995). A rose is a REEZ: The twocycles model of phonology assembly in English visual wordrecognition. Psychological Review, 102, 146-184.

Berndt, R. S., Reggia, J. A., & Mitchum, C. C. (1987). Empiricallyderived probabilities for grapheme-to-phoneme correspon-dences in English. Behavior Research Methods, Instruments,and Computers, 19, 1-9.

Bethell-Fox, C. E., & Shepard, R. N. (1988). Mental rotation:Effects of stimulus complexity and familiarity. Journal ofExperimental Psychology: Human Perception and Performance,14, 12-23.

Brown, G. D. (1987). Resolving inconsistency: A computationalmodel of word naming. Journal of Memory and Language, 26,1-23.

Butler, B., & Hains, S. (1979). Individual differences in wordrecognition latency. Memory & Cognition, 7, 68-76.

Coltheart, M. (1978). Lexical access in simple reading tasks. In G.Underwood (Ed.), Strategies of information processing (pp.151-216). San Diego, CA: Academic Press.

Coltheart, M. (1985). Cognitive neuropsychology and the study ofreading. In M. I. Posner & O. S. M. Marin (Eds.), Attention andperformance XI (pp. 3-37). Hillsdale, NJ: Erlbaum.

Cosky, M. J. (1976). The role of letter recognition in wordrecognition. Memory & Cognition, 4, 207-214.

Daniloff, R. G., & Moll, K. L. (1968). Coarticulation of Uprounding. Journal of Speech and Hearing Research, 2, 707-721.

Eriksen, C. W., Pollack, M. D., & Montague, W. E. (1970). Implicitspeech: Mechanism in perceptual encoding. Journal of Experi-mental Psychology, 84, 502-507.

Ferreira, E, Henderson, J. M., Anes, M. D., Weeks, P. A., &MacFarlane, D. K. (1996). Effects of lexical frequency andsyntactic complexity in spoken-language comprehension: Evi-dence from the auditory moving-window technique. Journal ofExperimental Psychology: Learning, Memory, and Cognition,22, 324-335.

Folk, M. D., & Luce, R. D. (1987). Effects of stimulus complexityon mental rotation rate of polygons. Journal of ExperimentalPsychology: Human Perception and Performance, 13, 395^404.

Forster, K. I. (1970). Visual perception of rapidly presented wordsequences of varying complexity. Perception & Psychophysics,8, 215-221.

Forster, K. I., & Chambers, S. (1973). Lexical access and namingtime. Journal of Verbal Learning and Verbal Behavior, 12,627-635.

Frederiksen, J. R., & Kroll, J. F. (1976). Spelling and sound:Approaches to the internal lexicon. Journal of ExperimentalPsychology: Human Perception and Performance, 2, 361-379.

Glushko, R. J. (1979). The organization and activation of ortho-graphic knowledge in reading aloud. Journal of Experi-mental Psychology: Human Perception and Performance, 5,674-691.

Haggard, M. (1973). Abbreviation of consonants in English pre-and post-vocalic clusters. Journal of Phonetics, 1, 9-24.

Henderson, L. (1982). Orthography and word recognition inreading. San Diego, CA: Academic Press.

Jared, D., & Seidenberg, M. S. (1990). Naming multisyllabicwords. Journal of Experimental Psychology: Human Perceptionand Performance, 16, 92-105.

Jespersen, O. (1904). Lehrbuch derPhonetik. Berlin: Teubner.Kawamoto, A. H. (1988). Distributed representations of ambiguous

words and their resolution in a connectionist network. In S. I.Small, G. W. Cottrell, & M. K. Tanenhaus (Eds.), Lexicalambiguity resolution (pp. 195-228). San Mateo, CA: Kaufman.

Kawamoto, A. H. (1993). Nonlinear dynamics in the resolution oflexical ambiguity: A parallel distributed processing account.Journal of Memory and Language, 32, 474-516.

Kawamoto, A. H., Kello, C. T, Higareda, I., & Vu, J. (1999).Parallel processing and initial phoneme criterion in namingwords: Evidence from frequency effects on onset and rimeduration. Journal of Experimental Psychology: Learning,Memory, and Cognition, 25, 362-381.

Kawamoto, A. H., Kello, C. T., Jones, R. M., & Bame, K. A.(1998). Initial phoneme versus whole word criterion to initiatepronunciation: Evidence based on response latency and initialphoneme duration. Journal of Experimental Psychology: Learn-ing, Memory, and Cognition, 24, 862-885.

Kay, J., & Bishop, D. (1987). Anatomical differences betweennose, palm, and foot, or, the body in questions: Further dissec-.tion of the processes of sub-lexical spelling-sound translation. InM. Coltheart (Ed.), Attention and performance XII: The psychol-ogy of reading (pp. 449-461). Hillsdale, NJ: Erlbaum.


Kello, C. T., & Kawamoto, A. H. (1998). Runword: An IBM-PCsoftware package for the collection and acoustic analysis ofspeeded naming responses. Behavior Research, Methods, Instru-ments, and Computers, 30, 371-383.

Klatt, D. (1974). The duration of [s] in English words. Journal ofSpeech and Hearing Research, 17, 51-63.

Kucera, H., & Francis, W. (1967). Computational analysis ofpresent-day American English. Providence, RI: Brown Univer-sity Press.

MacKay, D. G. (1987). The organization of perception and action:A theory of action and other cognitive skills. New York:Springer-Verlag.

McClelland, J. L., & Rumelhart, D. E. (1988). Explorations inparallel distributed processing. Cambridge, MA: MIT Press.

Norris, D. G., & Brown, G. D. A. (1985). Race models and analogytheories: A dead heat? Reply to Seidenberg. Cognition, 20,155-168.

O'Seaghdha, P. G., Dell, G. S., Peterson, R. R., & Juliano, C.(1992). Modeling form-related priming effects in comprehensionand production. In R. Reilly & N. E. Sharkey (Eds.), Connection-ist approaches to language processing (Vol. 1, pp. 373—408).Hillsdale, NJ: Erlbaum.

Patterson, K. E., & Morton, J. (1985). From orthography tophonology: An attempt at an old interpretation. In K. E.Patterson, J. C. Marshall, & M. Coltheart (Eds.), Surfacedyslexia: Neuropsychological and cognitive studies of phonologi-cal reading (pp. 15-34). Hillsdale, NJ: Erlbaum.

Peereman, R., & Content, A. (1995). Neighborhood size effect innaming: Lexical activation or sublexical correspondences? Jour-nal of Experimental Psychology: Learning, Memory, and Cogni-tion, 21, 409-421.

Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson,K. E. (1996). Understanding normal and impaired word reading:Computational principles in quasi-regular domains. Psychologi-cal Review, 103, 56-115.

Reicher, G. M. (1969). Perceptual recognition as a function ofmeaningfulness of stimulus material. Journal of ExperimentalPsychology, 81, 274-280.

Richardson, J. T. E. (1976). The effects of stimulus attributes uponlatency of word recognition. British Journal of Psychology, 67,315-325.

Sakuma, N., Fushimi, T, & Tatsumi, I. (1997). Measurement ofnaming latency of Kana characters and words based on speechanalysis: Manner of articulation of a word-initial phonemeconsiderably affects naming latency. Japanese Journal ofNeuro-psychology, 13, 126-136.

Seidenberg, M. S., & McClelland, J. L. (1989). A distributed,developmental model of word recognition and naming. Psycho-logical Review, 96, 523-568.

Shallice, T., Warrington, E. K., & McCarthy, R. (1983). Readingwithout semantics. Quarterly Journal of Experimental Psychol-ogy: Human Experimental Psychology, 35A, 111-138.

Smith, N. V. (1973). The acquisition of phonology: A case study.Cambridge, England: Cambridge University Press.

Stemberger, J. P., & Treiman, R. (1986). The internal structure ofword-initial consonant clusters. Journal of Memory and Lan-guage, 25, 163-180.

Sternberg, S., Monsell, S., Knoll, R. L., & Wright, C. E. (1978).The latency and duration of rapid movement sequences: Compari-sons of speech and typewriting. In G. E. Stelmach (Ed.),Information processing in motor control and learning (pp.117-152). New York: Academic Press.

Taraban, R., & McClelland, J. L. (1987). Conspiracy effects inword pronunciation. Journal of Memory and Language, 26, 608-631.

Treiman, R., Mullennix, J., Bijeljac-Babic, R., & Richmond-Welty,E. E. (1995). The special role of rimes in the description, use, andacquisition of English orthography. Journal of ExperimentalPsychology: General, 124, 107-136.

Umeda, N. (1977). Consonant duration in American English.Journal of'the Acoustical Society of'America, 61, 846-858.

Van Orden, G. C., Pennington, B. F, & Stone, G. O. (1990). Wordidentification in reading and the promise of subsymbolic psycho-linguistics. Psychological Review, 97, 488-522.

Weekes, B. S. (1997). Differential effects of number of letters onword and nonword naming latency. Quarterly Journal of Experi-mental Psychology: Human Experimental Psychology, 50A,439-456.


Appendix A

Vocalic Offset Latencies and Percentage of Errors for Each Item in Experiment 1Simple

Low frequency

Word

bangbeckbibboardaledeergazegirthgustkinpaneparchteachtendtote

Latency

354339326367380357365376357344339337322345360

% error

0.012.58.30.08.30.00.08.34.20.0

12.58.34.20.08.3

High frequency

Word

bankbedbitboarddatedeepgamegirlgunkidpageparkteamtesttone

Latency

336322315324355350354360349334331333338350346

% error

8.34.20.04.20.00.00.04.24.24.28.30.0

16.70.08.3

Complex

Low frequency

Word

blastblondbrickbriskclunkdriftgrapegreedgropeplugpnzetracetraittrimtrump

Latency

332331331345336366336364375335346344343365346

% error

0.04.20.00.08.30.0

12.50.08.30.00.0

12.50.00.00.0

High frequency

Word

blackblockbridgebringclubdrinkgraygreengrowthpluspricetradetraintriptruck

Latency

338327327321353372354346368309321335331343348

% error

0.00.00.04.20.00.00.04.20.00.08.30.00.08.30.0

Appendix B

Acoustic Latencies and Percentage of Errors for Each Item in Experiment 2Simple

Low frequency

Word

sanesectseedseepsighsilksinesolesoothesub

Latency

464491464493453463479459494452

% error

0.00.00.02.80.00.0

13.90.00.00.0

High frequency

Word

samesendseaseemsightsixsizesourcesoonsun

Latency

441472465451481440456451436443

% error

2.80.00.00.02.80.00.02.80.02.8

Complex

Low frequency

Word

spadespecksportstabstainstakestenchstinkstompstork

Latency

428465446454454460481451452433

% error

0.05.60.00.02.80.02.80.00.00.0

High frequency

Word

spacespentspokestaffstaystagestepstillstockstore

Latency

448426440446465455441422425420

% error

0.00.00.02.82.82.82.80.00.02.8

Appendix C

Acoustic Latencies and Percentage of Errors for Each Item in Experiment 3

Word

saintsandseamsilksinksoapsocksoothesurfsurge

Simple

Latency

447438463449447414434451427461

Complex for si

% error

0.04.28.34.20.00.00.04.20.00.0

Word

slangslavesleekslickslideslipslobsloopslopeslum

Latency

446437434434424450415418445440

% error

0.00.00.04.20.00.00.08.30.00.0

Complex for sm

Word

smacksmartsmashsmearsmellsmeltsmirksmogsmokesmug

Latency

431449433425414437418438398436

% error

0.00.00.00.00.00.00.00.00.00.0

Received June 19, 1997Revision received December 29,1997

Accepted March 5,1998

Documents

Effect of onset cluster complexity in speeded naming: A test of rule-based approaches