How Lexical Stress Affects Speech Segmentation and Interactivity: Evidence from the Migration Paradigm

JOURNAL OF MEMORY AND LANGUAGE ARTICLE NO.

36, 87–116 (1997)

ML962472

How Lexical Stress Affects Speech Segmentation and Interactivity: Evidence from the Migration Paradigm SVEN L. MATTYS

AND

ARTHUR G. SAMUEL

SUNY Stony Brook Current models of word recognition differ on the potential influence of lexical knowledge on sublexical processes. The present study addresses this issue through a paradigm based on the perceptual ‘‘migration’’ of linguistic units between two stimuli presented simultaneously, resulting in an illusory percept. For example, ‘‘kinrtrorverrsy’’ and ‘‘bosrglorrarfe’’ were played at the same time, with subjects judging, in one condition, if ‘‘controversy’’ was presented and, in another condition, if ‘‘bisrglorrarfe’’ was presented. The results showed that, in dichotic listening, the migration of the vowel, leading to erroneous detection of the prespecified target, occurred less often with real word targets than with nonsense word targets. However, when the two items of the pairs were played in both ears at different amplitudes so that they were perceived to be closer to each other than in the dichotic situation, the lexical effect only remained when the mispronounced phoneme was in an unstressed syllable. This observation suggests that the mispronunciation of stressed syllables of words impairs lexical access in such a way that no lexical influence on early processing stages is possible. In contrast, the mispronunciation of unstressed syllables does not affect lexical access, thus allowing the lexical influence to take place. These results are problematic for temporally sequential (left-to-right) models of lexical access. Rather, they support models that include the stress patterns of words as an important aspect of both lexical access and the interaction between sublexical and lexical knowledge. q 1997 Academic Press

Two important psycholinguistic issues have often been examined in separate arenas: speech segmentation and interaction between levels of processing. The first issue concerns the way that the continuous speech wave is segmented into words. Although there are a number of segmental and supra-segmental cues to word boundaries in continuous speech (Cooper & Sorenson, 1977; Lehiste, 1960,

1964; Nakatani & Schaffer, 1978), those cues do not appear to be stable and reliable (Klatt & Stevens, 1973; Lehiste, 1972; Nakatani & Dukes, 1977; Reddy, 1976). Faced with this apparent absence of effective boundary cues in the signal itself, some speech scientists have hypothesized that speech segmentation is not so much a cognitive reality as a theoretical concept. By conceptualizing speech segmentation as the result of word recognition instead of its starting point, they finessed the problematic uninterrupted nature of the speech signal (Marslen-Wilson, 1987; McClelland & Elman, 1986; Pisoni & Luce, 1987). An alternative viewpoint is that there may be indicators of word boundaries in the acoustic signal such as allophonic variation and prosodic cues specific to a given language (Church, 1987; Cutler, 1994; Cutler & Norris, 1988; Grosjean & Gee, 1987; Mehler, Dommergues, Frauenfelder, & Segui, 1981; Otake, Hatano, Cutler, & Mehler, 1993), as well as supra-lexical information such as syntax (Marslen-Wil-

This research was supported by Air Force Office of Scientific Research Grant 91-0378, by the National Institute of Mental Health, and by the Belgian American Educational Foundation. Special thanks are due Anne Cutler, Re´gine Kolinsky, and an anonymous referee for extremely constructive comments on previous versions of the present paper. Thanks are also due Donna Kat for her programming, her control of data treatment, and her excellent advice. We also thank Heather Bortfeld for her recording, Debra Gimlin for her proofreading and her unexpected French–English bilingual skills, and Lee Wurm for his constant availability and pertinent remarks. Address reprint requests to Sven L. Mattys, Department of Psychology, SUNY Stony Brook, Stony Brook, NY 11794-2500. E-mail: [email protected]. 87

0749-596X/97 $25.00 Copyright q 1997 by Academic Press All rights of reproduction in any form reserved.

AID

JML 2472

/

a005h$$141

12-06-96 11:18:59

jmla

AP: JML

88

MATTYS AND SAMUEL

son & Tyler, 1980; Tanenhaus & Lucas, 1987). These findings do not necessarily rule out models of word recognition in which segmentation is a by-product of lexical access. They simply suggest that segmentation may interact with lexical access to facilitate word recognition. Thus, the first issue concerns speech segmentation as a time-related aspect of speech processing. The second question addresses a more static aspect of word recognition. Generally, the recognition system is viewed as stratified into several layers of processing (e.g., phonetic features, phonemes, syllables, lexical units). The debate focuses on the number of layers needed by the system to perform word recognition adequately and how the layers interact, if they do. Models of word recognition differ with regard to the possible contribution of the mental lexicon to lower levels of analysis such as phonetic feature extraction, phoneme classification, and intermediate unit assembly. In some models, spoken words are processed through these levels hierarchically, without any backward influence. In others, the different processing stages interact, without waiting for the preceding stages to produce their outputs. This debate opposes, in an extreme way, autonomous/modular models to interactive/ connectionist ones. A major difference between these models is whether the lexicon affects lower levels of processing or not. According to autonomous models, spoken word recognition is accomplished through several functionally autonomous modules, each having its own domain of computation. For instance, in the Race model (Cutler & Norris, 1979; Cutler, Mehler, Norris, & Segui, 1987), two independent processing routes can be used: the phonemic route, that carries out the computation of sublexical representations on the basis of the sensory input, and the lexical route, that activates the phonological information that corresponds to a lexical unit. The two routes are activated in parallel during phoneme detection tasks and, presumably, during conventional speech situations. However, in certain circumstances, depending on the utility, one route may be preferred over the other.

AID

JML 2472

/

a005h$$141

According to Race, the variability in processing time across different classes of linguistic stimuli is due to the relative time course of the two competing routes. The main aspect of autonomous models is that there are no top-down connections from the lexical to the prelexical level and, hence, lexical knowledge cannot influence phoneme perception. The interactive alternative may be illustrated by Trace, a parallel distributed model developed by McClelland and Elman (1986). In Trace, lexical, syntactic, and semantic knowledge exert their influence on prelexical processes from the very onset of the spoken message. In particular, lexical knowledge influences phoneme perception. In fact, all of the models listed above posit that lexical knowledge has some influence (often facilitatory) during speech processing (Blank & Foss, 1978; Cairns & Shu, 1980; Marslen-Wilson & Tyler, 1980; Salasoo & Pisoni, 1985). The point on which they make different predictions is the locus (and the moment) in the system where this influence takes place. For the proponents of Race and autonomous models in general (Cairns, 1984; Fodor, 1983; Fodor, Bever, & Garrett, 1974; Forster, 1979; Massaro, 1989; Massaro & Cohen, 1991; Norris, 1986; Tanenhaus, Carlson, & Seidenberg, 1985), lexical effects are the consequence of postaccess mechanisms, i.e., mechanisms intervening after a lexical candidate has been isolated on a purely acousticphonetic basis. So, for autonomous models, lexical effects occur after the word has been accessed. For Trace and interactive models in general (Glucksberg, Kreuz, & Rho, 1986; Marslen-Wilson & Welsh, 1978; McClelland, 1991; Morton, 1969; Samuel, 1981, 1986), lexical effects begin to appear as the spoken message is heard. Lexical information feeds down to the prelexical level and aids perceptual processing. Therefore, for interactive models, lexical effects can occur before the word has been accessed. The Cohort model (Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978), sharing features with both Race and Trace, divides lexical access into two stages. The first stage, called

12-06-96 11:18:59

jmla

AP: JML

89

MIGRATIONS AND SPEECH PROCESSING

activation, involves the mapping of acousticphonetic information onto word-form representations stored in the lexicon. The second stage is the selection, from the subset (or cohort) of activated candidates, of the candidates that still match the unfolding sensory input. Marslen-Wilson (1978) claims that word recognition can theoretically occur when the cohort is reduced to one candidate. This point in the word is called the ‘‘uniqueness point’’ (UP). For instance, the UP of trespass is p since trespass is the only word in the lexicon that starts with tresp. Like Trace, the Cohort model allows lexical knowledge to participate on-line in isolating the best candidate through the selection stage. It is thus possible for recognition to occur before the entire input is available. Moreover, the lexical representations have their own specifications that determine their activation likelihood. For example, a high-frequency word would require less goodness-of-fit with the input to be activated and, therefore, enter the cohort more readily than a lower-frequency candidate. However, as in Race, lexical knowledge has no influence on the acoustic-phonetic analysis itself. The activation of the candidates is automatic, autonomous, and data-driven, whereas the selection process is partially lexicon- and contextdriven. As such, Cohort belongs more to autonomous than interactive models, at least during the activation/selection stage of processing, that is, up to the uniqueness point. It is still unclear what happens to phonemes near the end of a word. The last segments of a word may not be analyzed as thoroughly as earlier ones are. For example, when subjects have to repeat incoming speech aloud, staying as close in time as possible to the input, they fluently restore mispronunciations more often in third than in first syllables (Marslen-Wilson & Welsh, 1978). This result suggests that the later part of a word may be subject to perceptual negligence. However, late mispronunciations are detected faster than early ones (Cole, 1973; Cole & Jakimik, 1978, 1980). Here, in contrast, the processing of later sounds seems to be enhanced by the activated word and not

AID

JML 2472

/

a005h$$142

attenuated by it. It is interesting to envisage a left-to-right model of word recognition with the features of the cohort model in addition to top-down connections from the lexicon to prelexical levels that are activated only after the recognition point. We will refer to such a hybrid as a left-to-right, semi-interactive model. We previously pointed out that segmentation and interactivity have seldom been studied together. More exactly, models that focus on interactivity between levels of processing often consider segmentation to be a by-product of word recognition. One exception is Grosjean and Gee’s (1987) model that treats segmentation and interactivity as rhythmbased and, hence, as language-dependent phenomena. In this model, segmentation is viewed as a process itself (and not a by-product of word recognition) that takes place in collaboration with lexical access. During running speech, the perceptual system takes advantage of probabilistic markers in the signal, such as prosody, to start the segmentation process. This idea is similar to Cutler and Norris’s (1988) Metrical Segmentation Strategy (MSS). These authors proposed that English listeners start lexical access at the beginning of every strong syllable, irrespective of its temporal location. This could be a fairly successful strategy given the high occurrence of first-syllable-stressed words in English. The role of strong syllables in spoken word recognition has been emphasized in a number of studies (Bradley, 1980; Cutler, 1976, 1986; Cutler & Butterfield, 1992; Cutler & Clifton, 1985; van Heuven, 1985; Slowiaczek, 1990, 1991). In addition to segmentation principles, Grosjean and Gee’s (1987) model includes hypotheses about the interaction between lexical and prelexical processes. In this conceptualization, interactivity mainly takes place on weak syllables, after a cohort of candidates has been activated on the basis of a word’s strong syllable in a bottom-up fashion. There are two notable points in this assumption. First, contrary to most models of word recognition, words are not processed in a strictly

12-06-96 11:18:59

jmla

AP: JML

90

MATTYS AND SAMUEL TABLE 1 PREDICTIONS MADE BY FOUR MODEL TYPES ABOUT THE INFLUENCE OF LEXICAL KNOWLEDGE ON THE PRELEXICAL STAGES OF SPEECH PROCESSING Strong syllable on

First syllable

Later syllable

Location of top-down effects in the word

Early

Late

Early

Late

Autonomous models L-to-R semi-interactive models Stress-based models Interactive models

No No No Yes

No Yes Yes Yes

No No Yes Yes

No Yes No Yes

Note. Effects that are expected to be greater in the late position than in the early position have been underlined.

left-to-right fashion. Instead, recognition is guided by the stress patterns of words. A stressed syllable encountered in a word at time t drives lexical access so that the information occurring before as well as after this syllable can contribute to recognition of the word at time t / 1. Recognition thus proceeds forward (sequentially) and backward, with perceptual emphasis put on stressed syllables. This view reduces the theoretical importance of word onsets, consistent with findings by Connine, Blasko, and Hall (1991). Second, because interactivity is related to the stress pattern and stress pattern is a language-specific feature, interactivity is potentially language-dependent too. These two aspects make this stress-based model an interesting alternative to fully autonomous models like Race, fully interactive models like Trace, and left-to-right semi-interactive models derived from Cohort. Table 1 summarizes the predictions made by fully autonomous models, left-to-right semi-interactive models, stress-based models, and fully interactive models regarding the influence of lexical knowledge on prelexical levels of speech processing. We have included a variable labeled ‘‘position in the word,’’ with two levels: Early and Late. The Early position represents a location before the uniqueness point and the Late position represents a location after the uniqueness point. This parameter is critical for left-to-right semi-interactive models that define the uniqueness point as the recognition point of the word,

AID

JML 2472

/

a005h$$142

the moment at which autonomous processes switch to interactive processes. It is also important for stress-based models because these models suggest that lexical access can be initiated early or late in the word, depending on the stress pattern. Position is also relevant for interactive models. Indeed, it makes it possible to examine the lexical impact on prelexical stages at the beginning of the word, when the network has not received much top-down confirmation yet, and later in the word, when the system starts stabilizing on a unique output. For interactive models, the lexical effect should be present in both positions, but it should be greater late in the word than early because of the increase in lexical activation, a prediction that has been empirically supported by Frauenfelder, Segui, and Dijkstra (1990) and by Pitt and Samuel (1995). THE METHODOLOGICAL CHALLENGE These models have been investigated using several paradigms, including phoneme monitoring, phoneme classification, gating, shadowing, mispronunciation detection, phonemic restoration, and word spotting. All of these methods have revealed reliable lexical effects. So, at first glance, these effects support models allowing at least some interactivity. However, this conclusion is valid only if these methods actually tapped a prelexical stage. If they tapped a lexical or postlexical stage, the lexical effects would not imply a true topdown influence of the lexicon on prelexical

12-06-96 11:18:59

jmla

AP: JML

91

MIGRATIONS AND SPEECH PROCESSING

processing stages but, rather, some lexical factor or postlexical decision bias. Researchers are thus confronted with an equation with two unknowns: the architecture of the system and the functioning of the method used to explore this architecture. In order to shed some light on this ambiguity, Samuel (1981) tried to distinguish between what is due to prelexical processing and what is due to postlexical bias, using Signal Detection Theory (Green & Swets, 1966). This theory produces two indices: d * and Beta. d * assesses the discriminability of the observer (e.g., when the observer has to indicate if a signal was present or absent in a complex stimulus pattern) and Beta reflects any response bias due to the observer’s response strategies. d * has generally been interpreted as reflecting pure perception or, at least, a cognitive operation that takes place before a lexical representation has been retrieved. In contrast, Beta is assumed to reflect a postlexical response tendency. With this working hypothesis, Samuel (1981) found a lexical effect on phoneme restoration, an effect on the perceptual index d *: when words or pseudowords were presented with a phoneme either replaced by noise or mixed with noise, subjects’ discrimination between truly intact stimuli and those with noise replacement was worse when the item was a word than when it was a pseudoword; this effect was focused late in the words (see Table 1). Thus, if d * reflects perception or a prelexical process, these results support models in which lexical knowledge exerts its influence on perception or, at least, on sublexical processes in a genuine top-down fashion. However, the one-to-one correspondence between d * and perception on the one hand and Beta and postperception on the other may not be as straightforward as is generally assumed. First, the bias parameter Beta may reflect effects at a perceptual level, in which case it may be called ‘‘perceptual bias’’ (Bertelson & Radeau, 1976; Ratcliff, McKoon, & Verwoerd, 1989; Samuel, 1996). Second, isolating d * from Beta tells us only that we can obtain a measure of performance uncontami-

AID

JML 2472

/

a005h$$142

nated by subjects’ strategies. It does not ensure that this measure is generated at the perceptual level. Discrimination could be due in part to postlexical processes, independent of any response bias. Even though d * indexes discriminability, discriminability does not necessarily index perception (Norris, 1995). One way to show that d * effects are not occurring at a postlexical level is to demonstrate that, for a given technique, such postlexical effects are implausible on logical grounds. For example, Samuel (1996) argued that, at a postlexical level, where information about words has been retrieved, more processing resources are available for words than for pseudowords. Therefore, the processing of words should benefit from the lexicon whereas the processing of pseudowords should not. This is at odds with the observed negative lexical effect (i.e., lower discrimination for words than for pseudowords). Thus, the lexical effect on d * observed for phonemic restoration is likely to be prelexically, rather than postlexically, based. To investigate the relationship between sublexical processes and the lexicon with all response bias partialled out of the data, an optimal technique must meet two conditions: First, it must be based on a discrimination situation so that d * can be computed. Second, it must be possible to demonstrate that this index measures effects at a sublexical level. In this paper, we describe and use the Migration of Linguistic Units Technique that may meet these two criteria. The migration paradigm originates from studies in the domain of visual perception (Treisman & Gelade, 1980). Treisman and Schmidt (1982) observed that, when presented with red squares and green circles, subjects frequently experienced the presence of a red circle or a green square. According to the authors, such erroneous recombination, or migration, was evidence that visual attributes, such as shape and color, had been separately extracted from the visual field at some, presumably early, stage of processing. Thus, migrations can be thought of as the perceptual

12-06-96 11:18:59

jmla

AP: JML

92

MATTYS AND SAMUEL

blending of information coming from different sources. Treisman’s conjunction error technique has recently been adapted to speech processing. Kolinsky and her colleagues (Kolinsky, 1992; Kolinsky & Morais, 1992; Kolinsky, Morais, & Cluytens, 1995), in several experiments conducted in French, elicited migrations from dichotic pairs of spoken pseudowords (two pseudowords were presented simultaneously, one to each ear). The stimuli were constructed so that they could lead to the illusory perception of a word target by migration of some linguistic properties from one stimulus of the pair to the other (e.g., the two pseudoword stimuli ‘‘kirjou’’ /kiZu/ and ‘‘borton’’ /bɔtɔ˜ / can lead to the erroneous perception of the word target ‘‘bijou’’ /biZu/ by what the authors term a migration of the first consonant). Kolinsky et al. (1995) observed that the migration rate of syllables was higher than that of the initial consonant (and its phonetic features) or the first vowel. They suggested that a high migration rate for a possible speech unit provides evidence that this unit is extracted from the sound wave at an early stage of speech processing. This technique has so far been primarily used to study the units of speech perception across languages (Kolinsky & Morais, 1993; Morais, Kolinsky, Cluytens, & Paiva, in preparation, reported in Morais & Kolinsky, 1994; Morais, Kolinsky, & Nakamura, in press). The migration rate has been assessed by computing the discrimination index d *. d * is obtained using the false detection rate for a target in a dichotic stimulus pair where the target is not present (but could be erroneously perceived by migration of some linguistic fragment) and the correct detection rate of this target in a dichotic pair where the target is one member of the pair. There is some indirect evidence that the migration technique could reveal early codes of speech processing. First, conjunction errors, as originally described in Treisman’s feature integration theory of attention (Treisman & Gelade, 1980), have been used to study the nature of perceptual codes. Kolinsky (1988,

AID

JML 2472

/

a005h$$143

1989, 1994) has provided evidence that the erroneous recombination of extracted features, leading to illusory percepts, arose at a relatively early processing stage. She showed that there was no strong developmental effect in the occurrence of illusory conjunctions. Conjunctions were as frequent in young children or unschooled adults as in highly educated adults. In addition, illusory conjunctions seemed to be insensitive to both strategy and context because neither expectancies about the material nor familiarity with it hampered the phenomenon. These observations suggest that illusory conjunctions arise at a stage unaffected by higher-level knowledge. Similarly, phonetic feature blendings, that, to some extent, are the ‘‘ancestor’’ of the generalized migration paradigm, seem to tap a level where the acoustic signal is checked against phonetic categories at a relatively early processing stage (Cutting, 1976; Day, 1968; Halwes, 1969; Pisoni & McNabb, 1974; Tartter & Blumstein, 1981). There is also logical and empirical evidence against a postlexical explanation of the migration phenomenon. A postlexical locus would imply that the false detection of a target occurs after the two stimuli of a pair have been lexically processed. Specifically, one stimulus would be identified as the target in which a segment is missing while the other stimulus would be processed as a pseudoword that carries the missing segment. The false detection of the target would be the result of some postlexical reconciliation between the target and the stimuli. Practically, the missing segment would be extracted from one representation and inserted into the other. This postlexical segment-splicing hypothesis is at odds with a number of studies on the speech perception mechanisms in illiterate and ex-illiterate adults (adults who learned to read after the usual age). These studies have shown that illiterates have no awareness of speech as sequences of phones whereas literates do. In other words, illiterates are not able to perform the phoneme manipulation discussed above. For instance, Portuguese illiterates are unable to delete or add a phone at the beginning of

12-06-96 11:18:59

jmla

AP: JML

93

MIGRATIONS AND SPEECH PROCESSING

a spoken pseudoword—e.g., ‘‘porsa’’ becoming ‘‘orsa’’ after subtraction of the initial segment and, conversely, ‘‘orsa’’ becoming ‘‘porsa’’ after addition of an initial /p/ (Morais, Cary, Alegria, & Bertelson, 1979; see also Read, Zhang, Nie, & Ding, 1986, and Morais, Bertelson, Cary, & Alegria, 1986, for similar results). However, critically, Portuguese literates and illiterates yield the same phonetic feature blending rate (Morais, Castro, & Kolinsky, 1991; Morais, Castro, Scliar-Cabral, Kolinsky, & Content, 1987) and the same d * migration pattern, namely that consonants are the units that migrate the most (Castro, Vicente, Morais, Kolinsky, & Cluytens, 1995; Morais & Kolinsky, 1994; Morais, Kolinsky, Cluytens, & Paiva, in preparation; Morais, Kolinsky, & Paiva, unpublished). In short, illiterates experience consonant migrations to the same extent that literates do but they are not able to manipulate these consonants in tasks that appear to require the same segmentation skills as those needed if migrations were postlexically-based. This discrepancy suggests that migrations are not naturally accounted for by some postlexical mechanism. Instead, the similarity of the migration pattern between literates and illiterates suggests that the migration technique taps an early processing stage, just as Treisman and her colleagues have interpreted visual feature migrations in terms of early feature analysis. Because the migration technique can be used to compute d *, and there is empirical evidence that this technique may tap early processes, it should be a useful tool for studying sublexical processing. Moreover, like other paradigms, the migration technique allows us to test the lexical influence at different locations in the words. This factor is important to assess the different models shown in Table 1. FACILITATORY VERSUS INHIBITORY LEXICAL KNOWLEDGE Most experimental techniques show that lexical knowledge usually helps speech processing. For example, in a noisy setting, subjects are better at reporting high-frequency words than low-frequency words (Broadbent,

AID

JML 2472

/

a005h$$143

1967). Similarly, target phonemes are detected faster in words than in pseudowords (Rubin, Turvey, & van Gelder, 1976; Cutler, Mehler, Norris, & Segui, 1987). However, in some tasks, lexical knowledge can have a negative or inhibitory effect on performance. For example, as mentioned earlier, subjects fail to detect mispronunciations more often when the mispronunciations occur late in the word (i.e., when lexical activation is relatively strong) than in initial position (Cole, Jakimik, & Cooper, 1978). Similarly, when subjects repeat a mispronounced word back aloud in a shadowing task, they erroneously restore late mispronunciations more often than early ones (Marslen-Wilson & Welsh, 1978). Finally, the most striking report of a detrimental lexical effect comes from phonemic restoration studies (e.g., Samuel, 1981). These results, taken together, illustrate that the influence of the lexicon on word recognition can be seen in both positive and negative performance shifts. The migration technique is a young paradigm. Because it shares some features with both mispronunciation detection tasks and phonemic restoration, the way lexicality might show itself in the migration paradigm is hard to predict. Only a few studies based on migrations have addressed the issue of lexicality (Kolinsky, 1992; Kolinsky & Morais, 1992; Kolinsky, Mattys, Morais, Cluytens, & Evinck, in preparation). These studies were conducted in French and produced heterogeneous results. Kolinsky (1992) compared the migration rate of several linguistic units in bisyllabic word and pseudoword targets. She found differences in migration rates among the units but the pattern was the same for word targets and pseudoword targets. Thus, in this case, the lexical status of the targets had no effect on migrations. In this experiment, the migration rate was computed on the percentage of false detections alone because the design did not allow the computation of d * values. Kolinsky and Morais (1992) devised a similar experiment in which target-present trials were included so that d * values were available. They observed that the d * migration rate varied not only across units but also between

12-06-96 11:18:59

jmla

AP: JML

94

MATTYS AND SAMUEL

word and pseudoword targets: The migration rate was higher with word targets than with pseudoword targets. This lexical influence is inhibitory in the sense that the increased migration rate reflects a higher false alarm rate. This result may be explained as follows: A word target activates a representation in the lexicon that, in turn, can bias the processing of the subsequent input toward the word target. In contrast, a pseudoword target, because it has no lexical representation, would not favor the detection of any particular target. Therefore, the misperception of a pseudoword target would occur less often than the misperception of a word target. However, Kolinsky, Mattys, Morais, Cluytens, and Evinck (in preparation), in an experiment using bisyllabic stimuli, found no lexical effect of the target for consonant- and syllable-based migrations and a trend toward a facilitatory lexical effect for vowel migrations; the migration rate of vowels was higher with pseudoword targets than with word targets. In this case, the lexical representation activated by a word target might impose strict expectations of what is considered an acceptable token of the word. Pseudoword targets, with no stored representations, would not impose such strict rules of concordance. As a result, the misperception of a pseudoword target would occur more often than the misperception of a word target, a facilitatory lexical influence. The latter two experiments thus indicate that the migration paradigm can generate lexical facilitation as well as lexical inhibition. It is possible that the valence of the lexical effect (facilitatory vs inhibitory) may be affected by a number of procedural factors. For instance, the nature of the migration unit could influence the occurrence of lexical effects. In the Kolinsky, Mattys, Morais, Cluytens, and Evinck (in preparation) study, the type of unit interacted with lexicality. Vowels yielded a facilitatory lexical effect whereas consonants and syllables did not. Moreover, the length of the items could be an important aspect in migrations. So far, migration studies have only used relatively short sequences (in Kolinsky’s work, all the stimuli were bisyllabic).

AID

JML 2472

/

a005h$$143

The use of longer items could reveal an impact on the valence of the lexicality effect. In the experiments reported here, we examined three factors that could potentially affect the migration rate: the lexicality of the target (word vs pseudoword), the position of the migrating unit in the word (early vs late) and the stress pattern of the stimuli (first syllable vs third syllable). Experiment 1 explores the migration phenomenon in a situation where the subjects have to detect a pre-specified auditory target (word or pseudoword) in a pair of stimuli played dichotically. The lexical effect is analyzed and the direction of this effect is discussed (facilitation vs inhibition). Experiment 2 examines the lexical influence on migrations in four alternative listening situations. The perceptual distance between the two stimuli is manipulated by blending their respective amplitudes between the two ears. Experiment 2 investigates the degree to which linguistic migrations are affected by the perceptual distance between the complementary sources of information. Experiment 3 provides an empirical and methodological confirmation of an interesting result obtained in Experiment 2, while Experiment 4 tests a resource-based account of this finding. EXPERIMENT 1 Method Subjects. Twenty-eight undergraduate students of the State University of New York at Stony Brook participated in the experiment. Most were recruited in an Introductory Psychology class. All were native English speakers with no known hearing problems. They received payment or course credit for their participation. Materials. Thirty word targets were chosen, all of which started with a stop consonant. Fifteen were stressed on the first syllable and fifteen were stressed on the third syllable. Among those stressed on the first syllable, seven had four syllables, five had three syllables, one could either be classified as four or five syllables (‘‘capitalism’’), and two could be classified as having three or four (‘‘plagia-

12-06-96 11:18:59

jmla

AP: JML

95

MIGRATIONS AND SPEECH PROCESSING TABLE 2 EXAMPLE OF A SET OF TRIALS PRESENTED TO THE SUBJECTS Target-absent pairs Experimental

Target-present pairs Control

Word target: ‘‘controversy’’ /kɑntrəvÆrsi/ Early Late

kintroversy–bosrglorrarfe /kIntrəvÆrsi/–/bɑsglorɑfe/ kontrovarsy–bisrglorrerfe /kɑntrəvɑrsi/–/bIsglorÆfe/

controversy–bisrglorrarfea /kɑntrəvÆrsi/–/bIsglorɑfe/ controversy–bisrglorrarfeb /kɑntrəvÆrsi/–/bIsglorɑfe/

koontroversy–basrglorrarfe /kuntrəvÆrsi/–/bæsglorɑfe/ kontrovairsy–bisrglorrurfe /kɑntrəversi/–/bIsglorufe/

Pseudoword target: ‘‘bisrglorrarfe’’/bIsglorɑfe/ Early Late

kintroversy–bosrglorrarfe /kIntrəvÆrsi/–/bɑsglorɑfe/ kontrovarsy–bisrglorrerfe /kɑntrəvɑrsi/–/bIsglorÆfe/

controversy–bisrglorrarfea /kɑntrəvÆrsi/–/bIsglorɑfe/ controversy–bisrglorrarfeb /kɑntrəvÆrsi/–/bIsglorɑfe/

koontroversy–basrglorrarfe /kuntrəvÆrsi/–/bæsglorɑfe/ kontrovairsy–bisrglorrurfe /kɑntrəversi/–/bIsglorufe/

Note. This set is based on the original pair ‘‘controversy’’ /kɑntrəvÆrsi/–‘‘bisrglorrarfe’’/bIsglorɑfe/. a The two items of this pair were synchronized on the onset of the vowel of the first syllable. b The two items of this pair were synchronized on the onset of the vowel of the third syllable.

rism’’ and ‘‘preferable’’). Among those stressed on the third syllable, fourteen had four syllables and one could either be classified as a threeor four-syllable word (‘‘diagnosis’’). All of the words were recorded in their citation form, that is, with the full number of syllables. Words were paired across the stress manipulation in terms of the identity of the stressed vowel. For instance, for the word target ‘‘controversy’’ /kɑntrəvÆrsi/ (stressed on the first syllable), the corresponding word target of the other set was ‘‘correspondent’’ /kɑrspɑndənt/ (stressed on the third syllable). The two targets thus had the same vowel in the stressed syllable. The mean frequency of words in the first set was 18.3 (SD, 23.1); in the second set, it was 17.0 (SD, 19.0) (Kucera & Francis, 1967). For each of the 30 word targets, a pseudoword target was constructed. Table 2 presents an example stimulus set, based on the ‘‘controversy’’ /kɑntrəvÆrsi/–‘‘bisrglorrarfe’’ /bsglorɑfe/ pair. Each pseudoword was matched to its paired word in terms of the phonological categories of its constituent phonemes. Specifically, for the consonants, a stop sound in the word target led to another stop, at the same location, in the

AID

JML 2472

/

a005h$$143

matched pseudoword target (the stops’ category encompassed the phonemes /b, d, g, p, t, k/). In the same way, the presence of a fricative, nasal, or liquid consonant in a word target was matched by the presence of another fricative, nasal, or liquid consonant in the corresponding pseudoword target (this category encompassed the phonemes /f, v, s, z, m, n, l, rZ, ʃ/). Moreover, to avoid perceptual confusion in the dichotic pairs, we tried to maximize the phonetic contrast between the two phonemes at the same location. For example, for any /p/ in a word target, a phonetically contrasting stop, such as a /g/, was assigned to the corresponding location in the pseudoword target. As for the choice of the vowels, we were guided by perceptual phonetic distance tables. Furthermore, a pseudoword was assigned the stress pattern of its matched word, including primary stress and any secondary stress. Using these construction rules, the pseudoword target associated with the word target ‘‘controversy’’ was ‘‘bisrglorrarfe’’ where the stress was put on the first syllable. All of the targets are listed in the Appendix. For each of these 30 pairs, four new pairs were generated. Two of these new pairs assessed the migration of the vowel of the first

12-06-96 11:18:59

jmla

AP: JML

96

MATTYS AND SAMUEL

syllable (early) whereas the two others addressed the migration of the vowel of the third syllable (late). Four words had a schwa or a vocalic /l/ in their third syllable (‘‘preferable,’’ ‘‘delicacy,’’ ‘‘terminal,’’ ‘‘dangerous’’). All the other words had a full vowel in both their first and their third syllables (American phonetic transcription after Lewis, 1972). For each of the two positions, an experimental pair was created by exchanging the vowels of the syllable in question between the word and the pseudoword. That is, from the original pair ‘‘controversy’’–‘‘bisrglorrarfe,’’ two experimental pairs were created that were, respectively, ‘‘kintroversy /kntrəvÆrsi/–basrglor rarfe /bɑsglorɑfe/’’ (early) and ‘‘kontrovarsy /k ɑntr əv ɑ rsi/–bis r glo r rer fe /b  sglor Æfe/’’ (late). These pairs were of primary interest in this design because they were constructed in such a way that, if migration of the vowel occurred, it could lead to the misperception of either the target ‘‘controversy’’ or the target ‘‘bisrglorrarfe.’’ However, to ascertain that this erroneous perception, when it happened, reflected a true migration of vowels and not perceptual confusion due to the complexity of the listening situation, each experimental pair was accompanied by a control pair. The control pairs were identical to the experimental pairs, except that the critical vowels of the control pairs were different from the corresponding critical vowels in the experimental pairs. In this manner, the control pairs lacked the crucial attribute that, had it migrated, would have allowed the target illusion to occur. So, a false detection of the target in a control situation revealed the failure of the system to adequately perceive the phonemes of the complex dichotic signal. Hence, the control trials were regarded as a baseline condition for the migration phenomenon. As an example, the two control pairs derived from the original targets ‘‘controversy’’–‘‘bisrglor rarfe’’ and controlling for ‘‘kintroversy–bosrglorrarfe’’ (experimental early) and ‘‘kontrovarsy–bisrglorrerfe’’ (experimental late) were, respectively, ‘‘koontroversy /kuntrəvÆrsi/–basrglorrarfe /bæsglorɑfe/’’ and

AID

JML 2472

/

a005h$$144

‘‘kontrovairsy /kɑntrəversi/–bisrglorrurfe /bsglorufe/.’’1 The selection of the pseudoword targets as well as their sets of four derived pseudowords was constrained by an additional consideration. In order to limit the number of lexical candidates that these pseudowords could have generated on-line, we attempted to construct pseudowords that were as ‘‘pure’’ as possible. We chose pseudowords with lexical cohorts that quickly shrank to zero candidates. The pseudowords of the present experiment diverged phonetically from all words in the lexicon on or before their fourth phoneme (we considered the stress pattern but not the syllabic structure in the establishment of the cohort). In order to compute the discriminability score d * for each of the four conditions presented above (early experimental, early control, late experimental, late control), two target-present pairs were included in the design. These pairs consisted of the two original targets (e.g., ‘‘controversy–bisrglorrarfe’’). The only difference between them was the locus on which the two members of the pairs were synchronized. In one pair, the two items were synchronized on the onset of the vowel of the first syllable (early target–present condition). In the other pair, the two items were synchronized on the onset of the vowel of the third syllable (late target–present condition). These two target–present pairs provided the 1 The construction rules used for the control trials in our experiment were slightly different from those followed in Kolinsky et al.’s studies. In their control pairs, one of the stimuli was always identical to the stimulus that minimally departed from the target. For example, given the target word ‘‘bijou’’ /biZu/ and the experimental pair ‘‘kirjou–borton’’ /kiZu/–/bɔtɔ˜ /, the pair controlling for the migration of the first consonant was ‘‘kirjou–dor ton’’ /kiZu/–/dɔtɔ˜ /. This manipulation enabled them to focus on the migration of the sole consonant /b/, while keeping the carrier stimulus ‘‘kirjou’’ /kiZu/ found in the corresponding experimental pair. In our design, because our principal interest was the lexical issue, the control pairs were constructed so as to minimize the difference between word and pseudoword trials: the only difference between a word trial and a pseudoword trial was the lexicality of the target.

12-06-96 11:18:59

jmla

AP: JML

97

MIGRATIONS AND SPEECH PROCESSING

hit rate for the early and the late condition, respectively. The entire design included 30 sets similar to that presented in Table 2. In 15 sets, targets and stimuli were stressed on the first syllable and, in the other 15 sets, they were stressed in the third syllable. Thus, there were 360 trials in all [3 Types (target–absent/experimental, target–absent/control, and target–present) 1 2 Positions (early and late) 1 2 Lexical forms (word and pseudoword) 1 2 Stress positions (first syllable and third syllable) 1 15 Sets/Position]. They were all presented to all the subjects. The experiment was preceded by 24 practice trials that were constructed out of four sets of items not used in the experiment and assembled according to roughly the same principle (the pairing algorithm was slightly modified to allow subjects to practice on various trial types without excessively increasing the number of practice trials). The targets were recorded by a female native American English speaker and the stimuli of the dichotic pairs were recorded by a male native speaker. The voice contrast was intended to avoid detection responses based on pure acoustic matching between the target and the relevant stimulus of the pair. All the items were recorded in a sound attenuating chamber. They were low-pass filtered at 4.8 kHz, digitized at a 10 kHz sampling rate (12 bit A/D) and stored on the disk of a 486/33 computer. On output, the items were converted to analog form (12-bit D/A, 10 kHz), low-pass filtered (4.8 kHz) and played over high-quality headphones at a comfortable listening level (approximately 70 dB SPL). Within a pair, stimuli were synchronized at the onset of the critical vowel (either in the first syllable or in the third syllable). Procedure. Trials were presented randomly and ear assignment was randomly distributed among pairs. Headphone assignment to the ears was counterbalanced between subjects. Subjects were tested in groups of up to three at a time in a sound attenuating chamber. They were told that, on each trial, they would hear a target pronounced by a female voice fol-

AID

JML 2472

/

a005h$$144

lowed by two stimuli pronounced by a male voice and played dichotically. They were also instructed that targets and pairs could be meaningful or meaningless. They were asked to pay attention to both stimuli (i.e., to both ears) in order to decide whether the target had been presented or not. They gave their response for each trial through a response board with two labeled buttons (Yes for ‘‘present’’ and No for ‘‘absent’’). Subjects were encouraged to respond as quickly and accurately as possible. On each trial, there was a silent interval of approximately 500 ms between the binaurally presented target and the dichotically presented test pair. When the subjects had responded, there was a pause of 1 s before the next target was presented. If a subject failed to respond within 3 s, the program moved on to the next trial. The correctness and latencies of the responses were recorded. Results False alarms, hit rates, and d *s can be seen in Table 3. They were calculated for each of the 16 cells generated by the experimental design: Type of trial (experimental vs control), Lexicality of the target (words vs pseudowords), Position of the critical phoneme (early vs late), and Stress (first syllable vs third syllable). Migration rates, obtained from the d * differences between control and experimental trials, are also reported. From the hypothesis that phonemes may lose their ear specificity when presented dichotically and, as a result, erroneously combine, one can predict that control trials constitute a situation of higher discriminability than experimental trials (Treisman & Souther, 1986; Kolinsky, 1992). Indeed, on control trials, the phonemes making up the target are not all distributed among the two stimuli of the pair whereas, on experimental trials, all the sounds are present. Thus, discrimination between presence and absence of the target is likely to be better on control trials than on experimental trials. The working hypothesis is therefore that control trials should yield greater d *s than experimental trials. In this way, the migration

12-06-96 11:18:59

jmla

AP: JML

98

2.15 0.05

16.0 15.0

2.20

80.9

Exp. Ctrl.

1.40 0.10

34.9 17.1

0.34 2.20

80.2

25.4

1.86

31.9

1.50

81.1

Exp. Ctrl. Exp. Ctrl.

Stress 1

1.83 0.13

19.3

JML 2472

/

The target-present trials were the same for both experimental and control conditions. Migration rate is computed as dc*trl 0 de*xp .

1.84 2.75 00.10

2.65 2.61 00.39

2.22

a005h$$144

b

00.15

2.19 00.35

19.4 84.3 28.0 16.7 91.7 16.7 11.0 18.3

84.8

Exp. Ctrl.

FA HRa d* Migration rateb

AID

a

9.3

2.90

87.4

7.4

3.05

16.4

1.96

76.4

Exp. Ctrl. Exp. Ctrl. Exp. Ctrl. Exp.

Stress 3 Stress 1

Early

Ctrl.

Stress 1

Late

Stress 3

Stress 1

Early

Stress 3

Pseudoword targets Word targets

MEAN RESULTS ON THREE MEASURES IN THE DICHOTIC SITUATION: PERCENTAGE OF FALSE ALARM (FA), PERCENTAGE OF CORRECT DETECTION OR HIT RATE (HR), AND d * (EXPERIMENT 1)

TABLE 3

Late

Stress 3

MATTYS AND SAMUEL

rate can be thought of as the difference between d * on control trials and d * on experimental trials. A four-way analysis of variance was performed on the d *s, examining Type (experimental vs control), Lexicality (word targets vs pseudoword targets), Position (early vs late), and Stress (first syllable vs third syllable). A Lexical effect was present, with better discriminability for words than for pseudowords, F(1,27) Å 105.17, p õ .001 (2.53 vs 1.89). There was no Type effect, F(2,27) Å 1.62, ns (2.18, control vs 2.23, experimental), but a cross-over interaction between Type and Lexicality indicated opposite effects of Type for word and for pseudoword targets, F(1,27) Å 21.02, p õ .001. The migration phenomenon emerged only for pseudoword targets, d * being higher on control than experimental trials, F(1,27) Å 4.98, p õ .05 (1.96 vs 1.81). With word targets, there was an inverse-migration effect, F(1,27) Å 27.30, p õ .001, that is, d * was lower on control than experimental trials (2.40 vs 2.65). This interaction between Type and Lexicality was found both early and late in the stimuli, and in both first-syllable- and third-syllable-stressed stimuli. Discriminability was also affected by Stress. d * was greater for items stressed on the third syllable than on the first, F(1,27) Å 63.74, p õ .001 (2.47 vs 1.94). However, an interaction between Stress and Position, F(1,27) Å 16.25, p õ .001, revealed that the Stress effect was substantial in the late position, F(1,27) Å 106.28, p õ .001 (2.57 vs 1.73), but was only moderate early in the items, F(1,27) Å 3.12, p õ .10 (2.37 vs 2.15). Migration rates can be visually inspected in Fig. 1. The migration rate was defined as the difference between d * on control trials and d * on experimental trials (d *ctrl 0 d *exp). This index corresponds to the difference in discriminability between trials in which not all the sounds of the target are present in the stimuli (control trials) and trials in which all the sounds of the target are present in the stimuli (experimental trials). Better discriminability on control trials than on experimental trials, reported as a positive migration rate, signals migrations. Note

12-06-96 11:18:59

jmla

AP: JML

99

MIGRATIONS AND SPEECH PROCESSING

FIG. 1. Migration rates in the dichotic situation (Experiment 1).

that this index does not reflect overall accu- listening situation (dichotic) and the way the racy. pairs of stimuli were constructed. Due to the The principal conclusion one can draw from clear spatial segregation of the two stimuli these d * results is that there was a lexical ef- when played dichotically, the familiar word tarfect on migration. Words resisted migrations gets might have drawn more attention than more than pseudowords did. In other words, pseudoword targets. If this is true, the addimigrations were subject to a lexical facilita- tional attention to words would have especially tory effect. The migration phenomenon was favored experimental trials (i.e., fewer migraabsent, and even reversed (see below) with tions) because the critical vowel in the experiword targets, whereas it showed up with pseu- mental condition presented a stronger phonetic doword targets. This lexical effect was the contrast with the original vowel than the same early in the words and late in the words, matched control vowel did. This is because the and it was not affected by the stress pattern vowel in the experimental condition came from of the stimuli. the corresponding pseudoword target, and was chosen to make the pseudoword target highly Discussion contrastive with the word target. For example, Experiment 1 shows that, in dichotic lis- the phonetic contrast between the first vowels tening, the migration index behaves differ- in ‘‘kintroversy’’ /kntrəvÆrsi/ (experimental) ently with words and with pseudowords. The and ‘‘controversy’’ /kɑntrəvÆrsi/ was greater direction of the lexical effect on migrations than that in ‘‘koontroversy’’ /kuntrəvÆrsi/ (conindicates lexical facilitation and not lexical in- trol) and ‘‘controversy’’ /kɑntrəvÆrsi/. Therehibition. Thus, the lexicon appears to exert a fore, it would have been easier for the subjects facilitatory influence on sublexical processing. to reject ‘‘kintroversy’’ (experimental) than However, one feature of the results compli- ‘‘koontroversy’’ (control) when they were precates these conclusions. We did not expect to sented with the target ‘‘controversy.’’ As a refind a reverse migration effect for words (i.e., sult, with word targets, the number of false lower d * on control trials than on experimental detections would have been lower on experitrials) because the experimental situation was mental trials than on control trials, generating designed to allow d *-lowering migrations, a reverse migration effect. This phenomenon whereas the control case was not. This reverse would not have been present for pseudoword migration effect with word targets may be due targets because of the assumption that pseuto the combination of two factors: the type of dowords did not receive as much attention as

AID

JML 2472

/

a005h$$144

12-06-96 11:18:59

jmla

AP: JML

100

MATTYS AND SAMUEL

words. Our own observations, and reports by some subjects, were that, in this complex dichotic task, there was a tendency to try to listen to one ear’s stimulus, and that words were favored in this choice. This attentional issue is a serious concern because it could mean that the facilitatory lexical effect we observed on migration was the consequence of an attentional phenomenon generated by words and not the result of top-down lexical influence on sublexical analysis. In order to reduce the attentional asymmetry between the two members of the pairs, we designed a set of listening situations in which the perceived distance between the two stimuli was smaller. In Experiment 2, instead of presenting the stimuli dichotically, we created the impression that they were closer to each other, more toward the center of the head. This was achieved by presenting the two stimuli of the pair to both ears at different amplitudes. Four different situations were set up along this spatial proximity variable. In the situation we called ‘‘1/3,’’ the subjects heard the two stimuli of the pairs in both ears, but, in each ear, one of the two items was played at a third of its original amplitude. Thus, one stimulus was louder than the other in one ear, while the other stimulus was louder in the other ear. The impression created in this situation was that the two stimuli were coming from distinctively different positions but somewhat closer to each other than in the dichotic situation. In the same way, ‘‘1/2,’’ ‘‘1/1.5,’’ and ‘‘1/1’’ situations were devised so that the stimuli were, respectively, perceived closer and closer to each other. Note that, in the 1/1 situation, the two items were played in the two ears at the same amplitude. This situation was designed to eliminate the spatial cue entirely. Experiment 2 allows us to examine the evolution of the lexical effect along the spatial proximity variable. EXPERIMENT 2 Method In the four situations tested in Experiment 2, the experimental design and procedures

AID

JML 2472

/

a005h$$144

were largely identical to those of Experiment 1. Only differences are mentioned below. Subjects. One hundred twelve undergraduate students of the State University of New York at Stony Brook participated in Experiment 2. They were native English speakers with no known hearing problems. They received course credit for their participation. Twenty-eight subjects were assigned to each of the four listening situations. Materials. The targets and the pairs were the same as in Experiment 1. The only difference was the relative amplitudes of the two stimuli in each ear. In the 1/3 situation, in one ear, an item was played at a third of the amplitude of the other item. The ratio was reversed in the other ear. In the 1/2 situation, the amplitude ratio was one to two. In the 1/1.5 situation, the amplitude ratio was one to one and a half. In the 1/1 situation, the amplitude ratio was 1 to 1, that is, the two items were played at the same amplitude in the two ears. Procedure. Subjects in each of the four situations were told that, on each trial, they would hear a target pronounced by a female voice followed by two stimuli pronounced by a male voice and played simultaneously. They were asked to pay attention to both items in order to decide whether the target had been presented or not. Descriptive Results The evolution of the lexical effect (i.e., the difference between word and pseudoword scores) on the migration rate (d *ctrl 0 d *exp) along the spatial variable can be seen in Fig. 2. In this figure, Position and Stress were collapsed together. Not shown in the figure is that the increase in spatial proximity affected the overall level of performance. The false alarm rates were 19, 25, 26, 28, and 40% for the dichotic (Experiment 1), 1/3, 1/2, 1/1.5, and 1/1 situations, respectively. The corresponding mean percentage of correct detection was 83, 88, 84, 85, and 73%. Taken together, false alarms and hits yielded d * discrimination values of 2.21,

12-06-96 11:18:59

jmla

AP: JML

101

MIGRATIONS AND SPEECH PROCESSING

FIG. 2. Migration rates as a function of the spatial proximity between the stimuli of the pairs (Experiment 2).

2.27, 2.08, 1.96, and 1.05 for the five listening situations, respectively. Furthermore, as can be derived from Fig. 2, the increase in spatial proximity also affected the overall migration rate. Collapsing across word and pseudoword targets, the migration rates were 00.05, 00.09, 00.02, 0.12, and 0.30 in the listening situations ranging from dichotic to 1/1. Thus, the reduction of spatial cues increased the migration rate in a rather steady way. Listeners were more likely to experience migration of the vowel when the two stimuli in the pair were spatially confusable. This general increase of migration was expected because, with fewer (or no) spatial cues to inform the subjects where the sounds were, the phonemes were in fact more and more acoustically superimposed in the signal.2 However, spatial proximity seemed to 2 Cutting’s (1976) classic study of ‘‘six fusions’’ used the contrast between dichotic and binaural conditions as one method to discover different levels of analysis. All of the conditions in Cutting’s study involved much simpler stimuli than those of the current study, generally single CV or CCV syllables. Cutting found that with such simple stimuli, binaural presentation enhanced fusion at the simplest acoustic level, while dichotic presentation promoted fusion involving more phonological codes. Our results indicate that with complex, long stimuli, binaural presentation maximizes the likelihood of migrations.

AID

JML 2472

/

a005h$$145

affect words and pseudowords differently. The migration rate for pseudowords was relatively unaffected, while for words a substantial change was observed. The facilitatory lexical effect that we observed in Experiment 1 (dichotic listening) disappeared in Experiment 2 as the spatial cues became less available. Thus far, the results indicate that subjects’ performance was facilitated by the lexicality of the target when there was a possibility that the system could be attentionally biased toward words (dichotic listening). When this possibility was eliminated, the overall lexical effect dissipated and words no longer yielded different scores than pseudowords. We were also interested in whether or not this trend was present early and late in the stimuli and whether or not it depended upon the stress pattern. We examined the evolution of the lexicality effect along the spatial proximity variable separately for the early and late positions and for the two locations of stress. The general pattern remained the same: there was a global increase in migrations and a decrease of the lexical effect as one moved from the dichotic situation to the 1/1 situation. However, if, instead of inspecting the Position and the Stress factors independently, we combined them and looked at the lexical effect when the critical vowel was in the strong syllable as opposed to when it was not, two contrasting patterns of results emerged. When the critical vowel was in the stressed syllable, that is, when Position and Stress were ‘‘consistent,’’ the lexical effect was greatly reduced in the 1/2 situation, and totally disappeared (i.e., the migration rate was actually higher for pseudowords than for words) in the 1/1.5 and 1/1 situations (Fig. 3, top). In contrast, when the critical vowel was in the weak syllable, that is, when Position and Stress were ‘‘inconsistent,’’ the lexical effect was only mildly modulated by spatial proximity (Fig. 3, bottom). Even in the 1/1 situation, the lexical effect was significant, F(1,27) Å 3.48, p Å .073. Thus, words consistently induced fewer migrations than pseudowords did. Subjects were less likely to experience migrations in words than in pseudowords when the mispro-

12-06-96 11:18:59

jmla

AP: JML

102

MATTYS AND SAMUEL

FIG. 3. Migration rates as a function of the spatial proximity between the stimuli of the pairs. The upper graph shows the migration rates on consistent trials (the mispronunciation is in the strong syllable). The lower graph shows the migration rates on inconsistent trials (the mispronunciation is in the weak syllable) (Experiment 2).

nunciation was in the weak syllable, regardless of position. This distinction between consistent and inconsistent trials was very constant up to the 1/1.5 situation, that is, the situation in which most of the spatial information was eliminated but the two stimuli were not acoustically superimposed. In the 1/1 situation, this contrast was less marked. The difference in the evolution of the lexical effect along the spatial variable between consistent and inconsistent trials suggests that the lexical influence on speech processing is modulated by lexical stress. If, as we have

AID

JML 2472

/

a005h$$145

suggested, the migration paradigm provides an index of sublexical processing, then this interaction between Lexicality and Stress is problematic for autonomous and left-to-right semi-interactive models. It may be problematic, although probably to a lesser extent, for interactive models too. Further Results: Toward a Stress-Based Model of Lexical Access Figure 4 shows the breakdown of the results for listening situations 1/3, 1/2, 1/1.5, and

12-06-96 11:18:59

jmla

AP: JML

103

MIGRATIONS AND SPEECH PROCESSING

FIG. 4. Migration rates in the 1/3, 1/2, 1/1.5, and 1/1 situations (Experiment 2).

1/1. Figure 1 can be considered the initial graph in this series. As the two items of the pairs were presented closer to each other so that spatial cues became less available, the stress pattern had a progressively stronger influence on the lexical effect. This can be seen in the development of the interaction between Stress and Lexicality along the several spatial proximity steps. More specifically, the interaction built up in a complementary way early in the items (left graphs) and late in the items (right graphs). Let us first examine the evolution of the interaction between Stress and Lexicality late

AID

JML 2472

/

a005h$$145

in the words (right graphs). We will treat the right graph in dichotic listening (Fig. 1) as the first step of this progression. Under dichotic listening, there was no interaction between Stress and Lexicality, as shown by the parallelism between the word and the pseudoword bars. The lexical effect was present in firstsyllable- and third-syllable-stressed items to the same extent. This parallelism progressively broke down as we inspected the subsequent situations (moving down the column in Fig. 4). In other words, an interaction appeared between Stress and Lexicality, as we moved along the spatial proximity variable.

12-06-96 11:18:59

jmla

AP: JML

104

AID

JML 2472

/

a005h$$145

1.64 0.11

30.0 27.9

1.75

80.5

Exp. Ctrl.

1.22 0.11

39.7 21.1

0.56 1.94

76.7

36.4

1.38

36.8

1.33

80.4

Exp. Ctrl. Exp. Ctrl.

1.57 0.08

27.2

The target-present trials were the same for both experimental and control conditions. Migration rate is computed as dc*trl 0 de*xp .

1.98 2.34 0.17 2.51 2.41 0.09 2.50

AP: JML

b

0.17 2.30 00.32

21.6 89.3 31.4 31.0 93.1 25.8 26.7 22.6

92.6

Exp. Ctrl.

FA HRa d* Migration rateb

jmla

a

16.2

2.51

89.2

21.7

2.34

24.8

1.65

76.8

Exp. Ctrl. Exp. Ctrl. Exp. Ctrl. Exp.

Stress 3 Stress 1

Early

Ctrl.

Stress 1

Late

Stress 3

Stress 1

Early

Stress 3

Pseudoword targets

Stress 1

Late

12-06-96 11:18:59

Word targets

MEAN RESULTS ON THREE MEASURES IN THE 1/1.5 SITUATION: PERCENTAGE OF FALSE ALARM (FA), PERCENTAGE OF CORRECT DETECTION OR HIT RATE (HR), AND d * (EXPERIMENT 2)

TABLE 4

This interaction shows that, at the end point of the series, the lexical effect remained only on items stressed in their first syllables. It disappeared on items stressed in their third syllables. Inspection of the interaction between Stress and Lexicality early in items (left graphs) revealed a similar, complementary, pattern. In dichotic listening (Fig. 1, left graph), the lexical effect was present equally in both firstsyllable- and third-syllable-stressed items. Stress and Lexicality started interacting in the next situations (down the left column of Fig. 4) in such a way that a lexical effect only remained in items stressed in the third syllable. This progression was very consistent up to and including the 1/1.5 situation. The 1/1 situation departed from this regular pattern. Here, the lexical effect disappeared not only in items stressed in the first syllable but also in items stressed in the third syllable. The interplay between Stress and Lexicality, and the outlying result obtained early in the items in the 1/1 situation, need to be examined in more detail. Situation 1/1.5 showed the clearest indication of the interaction pattern. These results are presented in Table 4. When the strong syllable was mispronounced (early/first and late/third), detection of the error was not helped by the lexicon. Words yielded as many migrations as pseudowords did. In contrast, when one of the weak syllables was mispronounced (early/third and late/first), detection of the error was helped by the lexicon. Words resisted migrations more than pseudowords did. A three-way analysis of variance was performed on the migration rates obtained in the 1/1.5 listening situation with the factors Lexicality (word targets vs pseudoword targets), Position (early vs late), and Stress (first syllable vs third syllable). The Type factor (experimental vs control) was not entered into the analysis as such. Rather, it was synthesized into a single migration rate indexed by the difference d *ctrl 0 d *exp , which served as the dependent variable. The three-way interaction between Lexi-

Stress 3

MATTYS AND SAMUEL

105

AID

JML 2472

/

a005h$$145

0.64

35.5

0.93

0.29

44.7 66.4

Exp. Ctrl.

0.77

27.2

0.61 0.88

56.1

46.2

0.27

41.0

1.06

0.29

51.0 75.7

Exp. Ctrl. Exp. Ctrl.

1.09

The target-present trials were the same for both experimental and control conditions. Migration rate is computed as dc*trl 0 de*xp .

1.04 0.53 1.57 1.37 0.15 1.52

AP: JML

b

0.45 1.14 1.14

0.00

43.2 78.9 43.7 53.1 81.9 34.5 42.1 35.8

82.3

Exp. Ctrl.

FA HRa d* Migration rateb

jmla

a

34.1

1.35

76.7

49.6

0.90

29.2

1.16

0.07

32.3 66.7

Exp. Ctrl. Exp. Ctrl. Exp. Ctrl. Exp.

Stress 3 Stress 1

Early

Ctrl.

Stress 1

Late

Stress 3

Stress 1

Early

Stress 3

Pseudoword targets

Stress 1

Late

12-06-96 11:18:59

Word targets

MEAN RESULTS ON THREE MEASURES IN THE 1/1 SITUATION: PERCENTAGE OF FALSE ALARM (FA), PERCENTAGE OF CORRECT DETECTION OR HIT RATE (HR), AND d * (EXPERIMENT 2)

TABLE 5

cality, Position, and Stress was significant, F(1,27) Å 12.07, p õ .005. It revealed that the interaction between Lexicality and Stress was different early in the items than late in the items, F(1,27) Å 2.94, p õ .10 and F(1,27) Å 9.62, p õ .005, respectively. In the early position, Lexicality and Stress interacted in such a way that the lexical effect was found only when the third syllable of the items was stressed, F(1,27) Å 5.70, p õ .05 (0.17, words vs 0.56, pseudowords) (Fig. 4, 1/1.5 situation, left graph). Conversely, in the late position, the lexical effect appeared only in first-syllable-stressed items, F(1,27) Å 23.47, p õ .001 (00.32, words vs 0.11, pseudowords) (Fig. 4, 1/1.5 situation, right graph). These results suggest that stressed syllables are critical in accessing the lexicon, whatever their positions in the words. If one of these stressed syllables is damaged (in this case, mispronounced), lexical access is impaired in such a way that the lexical information is no longer available to help sublexical processes in a top-down way. In contrast, if an unstressed syllable is damaged, with the stressed syllable intact, lexical access is successfully realized, allowing the lexical information to be used to detect mispronunciations. Note that this lexical effect can span forward (i.e., when the stress is in the first syllable and the mispronunciation is late in the word) as well as backward (i.e., when the stress is in the third syllable and the mispronunciation is early in the word). This is consistent with models of spoken word recognition that emphasize the use of a metrical segmentation strategy together with partial interactivity. In listening situation 1/1, the stimuli of the pairs were heard at the same amplitude in the two ears. In one important respect, the pattern of results in this condition departs from the trend observed in the other situations. The results are shown in Table 5. A three-way analysis of variance was performed on the migration rates (d *ctrl 0 d *exp) obtained in the 1/1 listening situation with the factors Lexicality (word targets vs pseudoword targets), Position (early vs late) and Stress (first syllable vs third syllable). It re-

Stress 3

MIGRATIONS AND SPEECH PROCESSING

106

MATTYS AND SAMUEL

vealed a significant three-way interaction between Lexicality, Position, and Stress, F(1,27) Å 5.09, p õ .05. Lexicality and Stress interacted only when the mispronunciation was in the late position, F(1,27) Å 7.43, p õ .05 (Fig. 4, 1/1 situation, right graph). In this case, a significant lexical effect was found only for items stressed in the first syllable, F(1,27) Å 4.69, p õ .05 (0.00, words vs 0.29, pseudowords). Thus, although Situation 1/1 shows a hint of the pattern of interaction observed in Situation 1/1.5 in the early position, it is greatly attenuated, and does not approach statistical reliability. When the two stimuli of the pairs are played at the same amplitude in the two ears, lexicality does not seem to have any influence on the way mispronunciations are detected early in the items. It only does so when mispronunciations are found late in the items (a forward lexical effect). This observation suggests that both stress and the time-course of uttered words (i.e., early vs late) are crucial variables in lexical access. Discussion The results obtained in Experiment 2 show that the lexical effect found in Experiment 1 is actually a function of the spatial proximity between the two members of the pairs. When the two items are presented dichotically, the system takes advantage of the representations stored in the lexicon to resist migrations. However, as we suggested above, this lexical effect may be modulated by attentional factors. Indeed, as we reduced the availability of an attentional focus by increasing the spatial proximity of the stimuli of the pairs, the lexical effect was reduced. In the 1/1.5 situation, in which the spatial cues are reduced to a minimum but in which there is no physical superimposition of the two stimuli, the lexical topdown influence is completely determined by the stress patterns of the items. When the strong syllable is altered (by changing its vowel), words and pseudowords display the same rate of migration. In this case, apparently, words cannot benefit from top-down lexical information. In contrast, when a weak

AID

JML 2472

/

a005h$$145

syllable is altered but the strong syllable remains intact, the information from the lexicon is available to lower levels, allowing the interpretation of the input signal to be affected by the stored representation of the word. In such cases, fewer detection errors occur. The results obtained in the 1/1.5 situation suggest two points. First, lexical access is initiated on strong syllables of words (Cutler & Norris, 1988), and this initial stage is not strongly influenced by the lexicon; it is primarily data-driven. For this reason, mispronunciation of the strong syllable impairs lexical access and, hence, weakens the lexical effect. Second, weak syllables of words, whether they are located before or after the strong syllable, are processed in a more interactive way. Specifically, the lexical candidates activated by the strong syllable of the input make processing of weak syllables more acute. This lexical influence can temporally spread forward onto weak syllables following the strong syllable, and it can spread backward onto weak syllables preceding the strong syllable. The results obtained in the 1/1 situation show a rupture in the trend observed along the spatial variable. When the two stimuli were played at the same amplitude in the two ears, the detection of mispronunciations in the first syllable was no longer helped by the lexicon, whatever the stress pattern. That is, the retroactive lexical effect observed in the other situations did not appear. If this absence of an interaction between Stress and Lexicality early in the words reflects a true time-course feature of word processing, then the left-toright aspect of lexical access, put forth by the Cohort model for instance, must also be considered a critical factor in modeling speech processing. Given that the results obtained in situation 1/1 violated the regular pattern observed across all the other situations, in Experiment 3, we attempted to replicate the outlying result using a slightly different way of presenting the two stimuli of a pair in the two ears. This time, the absence of spatial cues was arranged by presenting the stimuli binaurally (in situa-

12-06-96 11:18:59

jmla

AP: JML

107

/

a005h$$145

12-06-96 11:18:59

0.55 0.29

59.1

0.84

75.5 49.1

Ctrl.

0.74 0.25

59.2 38.6

0.46 0.86

68.0

54.0

0.40

50.3

0.99

80.6

Exp. Ctrl. Exp.

0.64 0.28

48.0

The target-present trials were the same for both experimental and control conditions. Migration rate is computed as dc*trl 0 de*xp .

1.11 1.08 0.43 1.51 1.13 0.34 1.47

AP: JML

b

0.56 1.30 00.19

43.0 82.4 49.6 53.1 83.6 38.2 52.2 84.3 42.5

FA HRa d* Migration rateb

jmla

a

37.8

1.45

80.5

54.8

0.89

39.4

0.92

69.3

Exp. Ctrl. Exp. Ctrl. Exp. Ctrl. Exp. Exp. Ctrl.

The results obtained in binaural listening are presented in Table 6 and depicted in Fig. 5. A four-way analysis of variance was performed on the migration rates obtained in the 1/1 situation from Experiment 2 and the binaural listening situation of the current experi-

JML 2472

Ctrl.

Stress 3 Stress 1

Results and Discussion

AID

Ctrl.

Stress 1 Stress 1

Late

Stress 3

Stress 1

Early

Stress 3

Pseudoword targets Early

In Experiment 3, the experimental design and procedure were largely identical to those of Experiments 1 and 2. Only differences are mentioned below. Subjects. Twenty-eight undergraduate students of the State University of New York at Stony Brook were subjects in the experiment. They were all native English speakers with no known hearing problems. They received course credit for their participation. Materials. The targets and the pairs were the same as in Experiments 1 and 2. The only difference was that the items were presented binaurally, that is, they were played on ‘‘mono’’ through the amplifier. The two items of the pairs were thus acoustically superimposed and came from the same spatial origin. Procedure. The procedure was the same as that described in Experiment 2.

TABLE 6

Method

Word targets

EXPERIMENT 3

MEAN RESULTS ON THREE MEASURES IN THE BINAURAL SITUATION: PERCENTAGE OF FALSE ALARM (FA), PERCENTAGE OF CORRECT DETECTION OR HIT RATE (HR), AND d * (EXPERIMENT 3)

Late

Stress 3

tion 1/1, their volumes were digitally averaged). In practice, we turned the channel distributor button to ‘‘mono’’ on the amplifier so that both stimuli of the pairs were presented to both ears. Although these two situations are generated from different electronic manipulations, they are perceptually identical and, therefore, should not yield any differences in the subjects’ performances. If the results we obtain from the binaural situation in Experiment 3 are similar to those of Situation 1/1 in Experiment 2, we can then conclude that the absence of an interaction between Lexicality and Stress early in the items is really tied to the particular listening condition created by the physical superimposition of the stimuli of the pairs.

Exp.

MIGRATIONS AND SPEECH PROCESSING

108

MATTYS AND SAMUEL

FIG. 5. Migration rates in the binaural situation (Experiment 3).

ment, with the factors Listening situation (1/1 vs binaural), Lexicality (word targets vs pseudoword targets), Position (early vs late) and Stress (first syllable vs third syllable). The type of Listening situation (1/1 vs binaural) proved to be a nonsignificant factor in all the tests except in one of the three-way interactions (Listening situation 1 Position 1 Stress, F(1,27) Å 6.75, p õ .05). The Lexicality of the items was not affected by the type of Listening situation in any way. Thus, the binaural results provide a replication of the results obtained in the 1/1 situation. To avoid confusion, we will refer to either one of them as the generic binaural situation. Taken together, the results of the two binaural situations show that, when the two items of the pairs are acoustically superimposed, the stress pattern shapes the lexical effect only when the mispronunciation is near the end of the word. There is no retroactive lexical effect on mispronunciations at the beginning of the word. This observation suggests that, along with the stress pattern, there is an important sequential component of lexical access. However, as suggested above, the binaural presentation of stimuli is, in essence, quite different from the other types of listening situations in that the phonemes are acoustically superimposed whereas, in the other situations, they are spatially distinct. This aspect could make the binaural listening situation so arduous that

AID

JML 2472

/

a005h$$145

resource-demanding aspects of lexical access—such as the retroactive lexical influence—would be more affected than less resource-demanding aspects—such as the forward lexical influence. Indeed, the retroactive lexical influence implies that the detection of mispronounced weak syllables occurring early in the word is assisted by lexical information obtained later in the word. This presupposes that the decision about the identity of the early phonemes is delayed until the word is accessed through its late strong syllable. Only at that moment can the retroactive lexical influence occur. While there is evidence in the literature for this type of retroactive process (Connine, Blasko, & Hall, 1991; Samuel, 1979), retaining this information is no doubt resource-demanding. In contrast, the detection of mispronounced weak syllables occurring late in the word is helped by previously accessed lexical information spread forward along the word. Thus, in this case, because no information must be held in a temporary buffer, fewer resources are needed. In the binaural situation, the retroactive lexical influence may be hampered because of the imprecision of the information contained in the buffer—due to the acoustical superimposition of the two items. The forward lexical influence would not suffer such a penalty because it does not require that any information be maintained in a buffer. Hence, it would be less

12-06-96 11:18:59

jmla

AP: JML

109

/

a005h$$145

12-06-96 11:18:59

1.07 0.06

40.8

1.13

75.0 39.0

Ctrl.

0.86 0.11

45.7 34.1

0.33 1.05

68.5

43.5

0.72

41.6

0.97

73.5

Exp. Ctrl. Exp.

1.25 00.08

33.4

jmla

1.29 1.34 0.34 1.68 1.35 00.12

1.23

The target-present trials were the same for both experimental and control conditions. Migration rate is computed as dc*trl 0 de*xp . b

0.35 1.66 00.37

34.3 83.7 46.3 50.9 87.6 38.6 41.7 82.8 45.9

AP: JML

a

36.8

1.66

85.5

48.9

1.31

35.4

1.17

71.5

Exp. Ctrl. Exp. Ctrl. Exp. Ctrl. Exp. Exp. Ctrl.

FA HRa d* Migration rateb

The results obtained in this 1/1.5 / Noise situation are presented in Table 7 and shown in Fig. 6. The overall false alarm, hit, and d *

JML 2472

Ctrl.

Stress 3 Stress 1

Results and Discussion

AID

Ctrl.

Stress 1 Stress 1

Late

Stress 3

Stress 1

Early

Stress 3

Pseudoword targets Early

Experiment 4 was almost identical to situation 1/1.5 in Experiment 2. The only difference was that white noise was added to the pairs of stimuli. Subjects. Twenty-eight undergraduate students of the State University of New York at Stony Brook were subjects in the experiment. They were all native English speakers with no known hearing problems. They received payment or course credit for their participation. Materials. The targets and the pairs were the same as in situation 1/1.5 in Experiment 2. The only difference was that the test pairs were presented with white noise added to them. White noise of a fixed amplitude was presented throughout each pair. The same amplitude was used for all pairs. It was determined empirically through pilot testing; we chose an amplitude level that produced an overall performance level equivalent to that of the binaural situation. The noise was added to the stimuli by adding a random number between 010 and /10 to every point of the signal. The digital root mean square amplitude of the stimuli, before adding the noise, was 54, with an average peak range of approximately {300. Procedure. The procedure was identical to that described in Experiment 2.

TABLE 7

Method

Word targets

EXPERIMENT 4

MEAN RESULTS ON THREE MEASURES IN THE 1/1.5/NOISE SITUATION: PERCENTAGE OF FALSE ALARM (FA), PERCENTAGE OF CORRECT DETECTION OR HIT RATE (HR), AND d * (EXPERIMENT 4)

Late

Stress 3

sensitive to a perceptual load such as that created by the binaural situation. If this hypothesis is correct, then adding a processing load, such as white noise, in the 1/1.5 situation should cause a loss of the retroactive lexical effect, but leave the forward lexical effect unscathed. In other words, the results should be similar to those obtained in the binaural situation. Experiment 4 tests this hypothesis by presenting the items from the 1/1.5 situation in a white noise background.

Exp.

MIGRATIONS AND SPEECH PROCESSING

110

MATTYS AND SAMUEL

FIG. 6. Migration rates in the 1/1.5 / Noise situation (Experiment 4).

scores were, respectively, 41%, 78%, and 1.23, compared to 40%, 73%, and 1.05 in the 1/1 situation and 48%, 78%, and 0.94 in the binaural situation. A four-way analysis of variance was performed on the migration rates obtained in the 1/1.5 / Noise situation and the binaural situation from Experiment 3, with the factors Listening situation (1/1.5 / Noise vs binaural), Lexicality (word targets vs pseudoword targets), Position (early vs late), and Stress (first syllable vs third syllable). The type of Listening situation (1/1.5 / Noise vs binaural) proved to be nonsignificant in all the tests. Thus, the 1/1.5 / Noise situation and the binaural situation yielded identical results. Specifically, the 1/1.5 / Noise situation produced a significant interaction between Position, Lexicality, and Stress, F(1,27) Å 9.18, p õ .005. The interaction between Lexicality and Stress was reliable only in the late position, F(1,27) Å 36.18, p õ .001 (in the early position, F(1,27) õ 1). In the late position, a lexical effect was found when stress was in the first syllable, F(1,27) Å 13.25, p õ .001. The reverse pattern was observed when stress was in the third syllable, F(1,27) Å 6.75, p õ .05. These results show that the addition of a processing load, such as white noise, eliminates retroactive lexical effects but does not affect forward lexical effects. This finding suggests that the absence of retroactive lexical

AID

JML 2472

/

a005h$$146

effects observed in the binaural situation is due to an inability to maintain a usable representation for the required time. Experiment 4 thus confirms the importance of the stress pattern during speech segmentation and lexical access, and shows that retroactive lexical help is a fragile component of speech processing. GENERAL DISCUSSION The experiments presented in this paper stemmed from the hope to unite two lines of inquiry in modeling spoken word recognition: the segmentation of the continuous speech stream and the potential top-down influence of the lexicon on sublexical processing. The experiments described in this paper demonstrated that lexical stress could be an important aspect at the intersection of these two concerns. The results obtained in the present research can be summarized as follows. Experiment 1 suggests that sublexical processing is affected by lexical knowledge in such a way that words resist phoneme migrations more than pseudowords do. We propose that this lexical facilitation is the result of top-down connections between lexical representations and sublexical stages of processing. These connections would make the sublexical processes more sensitive to minor discrepancies between the incoming signal and the stored lexical representations. As a result, subjects would be more prone to

12-06-96 11:18:59

jmla

AP: JML

111

MIGRATIONS AND SPEECH PROCESSING

detecting vowel mispronunciations, and hence to avoiding migrations, in words than in pseudowords. Experiment 2 shows that this topdown lexical effect does not take place evenly throughout the word. In the 1/1.5 situation, in which attentional biases are reduced to a minimum, the lexical effect remains only when the strong syllable of the word is intact. This result is an indication that strong syllables play a crucial role in lexical access. Therefore, we hypothesize that strong syllables are used to initiate the access to the lexicon, a procedure that entails proactive as well as retroactive processing. However, the results of the binaural situation seem to indicate a time-course aspect of lexical access in addition to the rhythmic aspect. Indeed, when both items of the pairs are presented in the two ears at the same amplitude, the retroactive lexical influence disappears. That is, the lexical facilitation does not occur at the beginning of words, even if the stressed syllable is intact. Experiment 3 replicates this finding. Nevertheless, it does not rule out the importance of the stress pattern during lexical access because Experiment 4 demonstrates that the absence of a retroactive lexical effect in binaural listening is primarily the consequence of a processing resource overload. Thus, the present data, as a whole, support a stress-based model of lexical access and segmentation consistent with the Metrical Segmentation Strategy (Cutler & Norris, 1988) and the model hypothesized by Grosjean and Gee (1987). In addition, they provide information about how and when lexical knowledge intervenes during lexical access. The (facilitatory) activation of top-down connections between the lexicon and lower levels occurs primarily on weak syllables, that is, when the quality of the signal is impoverished by a globally lower intensity and the presence of reduced vowels. It could be argued that our results did not tap a stage as early as we claim. Indeed, the influence of the perceptual distance on the pattern of lexical effects suggests that migrations are sensitive to attentional factors. When there is a possibility for the subjects to focus their

AID

JML 2472

/

a005h$$146

attention on the word of the pair (e.g., in the dichotic situation), the lexical effect is found in stressed and unstressed syllables to the same extent. It is only when attention is not biased toward the word that the interaction between Lexicality and Stress occurs. For us, this attentional result does not necessarily mean that migrations arise at a postlexical level or as a result of response bias. It indicates that lexical knowledge can help the processing of strong syllables too, provided that there is little or no interference due to an irrelevant message (the pseudoword). This is an indication of the fragility of the top-down lexical effect, not of its locus of intervention. Likewise, a certain fragility of the retroactive lexical effect is observed when noise is added to the signal. To date, the most compelling facts supporting the earliness of the migration technique are (a) the use of the bias-free index d * and (b) the numerous studies that show consistent migration patterns across preliterate infants, normal adults, ex-illiterate, and illiterate adults. However, because there is no absolute one-toone correspondence between, on the one hand, the locus of processing and, on the other hand, literacy and phonological awareness, we wish to remain open-minded on the locus to which the present findings must be attributed. When we compare the present results to the predictions made by the four major types of models presented in Table 1, it appears that fully autonomous models as well as fully interactive models have difficulty accounting for the results. The stress-based lexical effect we observed cannot be accounted for by autonomous models because these models do not allow any lexical influence before the word has been accessed in a fully bottom-up fashion. As for interactive models, top-down effects are expected to commence as soon as the first sounds of the word are pronounced and to increase steadily all along the word until one single candidate has been isolated. Our data did not show a systematic positional increase in migration. To account for the present data, a connectionist model like Trace would require significant modifications, such

12-06-96 11:18:59

jmla

AP: JML

112

MATTYS AND SAMUEL

as allowing the units to be differentially activated according to the clarity of the incoming input (e.g., strong vs weak syllables), incorporating some temporal backtracking feature, or adding a syllabic level duplicated in a layer of strong syllables and a layer of weak syllables. At this point, our data are best explained by a model that allows lexical access to be driven principally by the strong syllables present in the signal. Such a mechanism, which has been empirically supported in the Metrical Segmentation Strategy (MSS) put forth by Cutler and her associates (Butterfield & Cutler, 1988; Cutler, 1990; Cutler & Butterfield, 1992; Cutler & Carter, 1987; Cutler & Norris, 1988), appears to be both cognitively inexpensive and statistically effective. Indeed, few resources are required by a solely stress-based scanning of the signal. In addition, with 74% of all full vowels found in the initial syllables of lexical (content) words (Svartvik & Quirk, 1980), a strong syllablebased segmentation mechanism like MSS could be very effective for English. As suggested above, such a model theoretically involves both forward and retroactive temporal processing. Although both phenomena emerged in our data, we observed that the latter was more vulnerable to demanding listening conditions than the former. Given this pattern, left-to-right models of speech processing, such as Cohort, have gained some support from the present experiments. The asymmetry between the two mechanisms would probably stem from the low occurrence of words stressed in a noninitial syllable.

AID

JML 2472

/

a005h$$146

However, we showed that the way this infrequent situation is handled by the system is by retroactively processing the information encountered before the stressed syllable, with the help of lexical information made available when the late stressed syllable is processed. The fact that words stressed in a noninitial syllable are infrequent implies that the retroactive lexical effect occurs relatively infrequently as well. To conclude, what the present data have shown is that, despite the intrinsic sequential nature of speech production, lexical access does not always proceed in a left-to-right fashion. The stress patterns of words are used to set word boundaries in the continuous speech stream. Word onsets are postulated on each strong syllable and the lexical candidates activated thereupon help the processing of incoming sounds on-line. Right-to-left speech processing occurs when the stressed syllable does not start the word. This retroactive mechanism is less robust than forward top-down influences, but appears to operate well under moderate to good listening conditions. Our results help to reconcile competing theories of word recognition, because they require predominantly but not exclusively left-to-right mechanisms, stress-based mechanisms, and autonomous processing of some material (stressed syllables), as well as interactive influences (on unstressed syllables). Thus, rather than seeing these as competing theoretical constructs, we suggest that each mechanism plays a role in word recognition.

12-06-96 11:18:59

jmla

AP: JML

113

MIGRATIONS AND SPEECH PROCESSING

APPENDIX Targets Used in Experiments 1 through 4 Targets stressed on the first syllable

Targets stressed on the third syllable

controversy /`kɑntrə fi vÆrsi/ bisrglorrarfe /`bIsglo fi rɑfe/

coconuts /`kokə fi nöts/ bairderliksh /`bɑIdə fi lIkʃ/

correspondent / fi kɑrIs`pɑndənt/ birzalrgunrtolp / fi bIsɑl`guntɑlp/

diagnosis / fi dɑIəg`nosIs/ porekrlirvus / fi poək`lIvus/

dictionary /`dIkʃə fi neri/ koprsirlarzaw /`kɑpsI fi lɑzɔ/

difficulty /`dIfə fi költi/ karlerbusrkaw /`kɑlə fi buskɔ/

demolition / fi demə `lIʃən/ karrirmorsel / fi kɑrI `mɔsel/

politician / fi pɑlə `tIʃən/ gurmerpervis / fi gumə `pevIs/

kindergarten /`kIndər fi gɑrtən/ bufrtesrpirrdam /`buftəs fi pIrdəm/

caterpillar /`kætə fi pIlər/ birgusrtarmil /`bigus fi tɑmIl/

competition / fi kɑmpə `tIʃən/ dunrkirbarres / fi dunki`bɑres/

disadvantage / fi dIsəd`væntIdZ kolrigrfunrpats / fi kɑlIg`funpɑts/

plagiarism /`pleIdZə fi rIsəm/ drartsirlorfin /`drɑtsI fi lɑfən/

terminal /`tÆrmInəl/ gerlorfin /`geIlofin/

cancellations / fi kænsə `leIʃəns/ dulrfarzorrimf / fi dulfɑ `zɔrImf/

controversial / fi kɑntrə `vÆrʃəl/ belrgrarfirlas / fi belgrə `fIlɑs/

capitalism /`kæpItə fi lIzəm/ burtergrarfel /`butəgə fi rɑfəl/

tablespoon /`teIbəl fi spun/ durkilrstam /`dukIl fi stɑm/

propaganda / fi prɑpə `gænd/"/ glirdorpumrke / fi glIdo`pumke/

demonstration / fi demən`streIʃən/ porlimrglurzem / fi pɑləm`gluzəm/

generator /`dZenə fi reItər/ plirmorzurkor /`plImo fi zukor/

bakery /`beIkəri/ kortirma /`kɔtImɑ/

callisthenics / fi kælIs` θenIks/ birnelrgarreps / fi bInəl`gɑreps/

generations / fi dZenə `reIʃəns/ kirzarvurselve / fi kIzɑ `vuselv/

preferable /`prefərəbəl/ glirzarnurkef /`glIzənukef/

dangerous /`deIndZərəs/ parkrirmal /`pɑkrImɑl/

confidential / fi kɑnfI `denʃəl/ binrserkorfam / fi binsə `kɑfæm/

calculations / fi kælkjə `leIʃəns/ dulrpirsorfalm / fi dulpI `sɔfɑlm/

delicacy /`deləkəsi/ kurrorpurle /`kuropule/

comprehension / fi kɑmpri`henʃən/ dinrklerfusrtim / fi dInkle`fustəm/

Note. ɑ as in arm; I as in sit; i as in bee; e as in ten; æ as in hat; o as in go; u as in too; ö as in cup; Æ as in fur; ɔ as in jaw; ə Å schwa; ʃ as in shop; Z as in vision; θ as in thin; ` Å primary stress; fi Å secondary stress.

REFERENCES BERTELSON, P., & RADEAU, M. (1976). Ventriloquism, sensory interaction and response bias: Remarks on the paper by Choe, Welch, Gillford and Juola. Perception & Psychophysics, 19, 531–535. BLANK, M., & FOSS, D. (1978). Semantic facilitation and lexical access during sentence processing. Memory & Cognition, 6, 644–652. BRADLEY, D. (1980). Lexical representation and derivational relation. In M. Aronoff & M.-L. Kean (Eds.), Juncture. Saratoga, CA: Anma Libri.

AID

JML 2472

/

a005h$$147

BROADBENT, D. E. (1967). Word-frequency effect and response bias. Psychological Review, 74, 1–15. BUTTERFIELD, S., & CUTLER, A. (1988). Segmentation errors by human listeners: Evidence for a prosodic segmentation strategy. Proceedings of SPEECH ’88, Seventh Symposium of the Federation of Acoustic Societies of Europe (Vol. 3, pp. 827–833). Edinburgh, Scotland: Federation of Acoustic Societies of Europe. CAIRNS, H. S. (1984). Current issues in language comprehension. In R. Naremore (Ed.), Recent advances in language sciences. San Diego: College Hill Press.

12-06-96 11:18:59

jmla

AP: JML

114

MATTYS AND SAMUEL

CAIRNS, H. S., & SHU, J. (1980). Effects of prior context upon lexical access during sentence comprehension: A replication and reinterpretation. Journal of Psycholinguistic Research, 9, 319–326. CASTRO, S.-L., VICENTE, S., MORAIS, J., KOLINSKY, R., & CLUYTENS, M. (1995). Segmental representation of Portuguese in 5- and 6-year olds: Evidence from dichotic listening. In I. Hub Faria & J. Freitas (Eds.), Studies on the acquisition of Portuguese. Proceedings of the first Lisbon meeting on child language (pp. 1–16). Lisbon: Colibri. CHURCH, K. W. (1987). Phonological parsing and lexical retrieval. Cognition, 25, 53–69. COLE, R. A. (1973). Listening for mispronunciations: A measure of what we hear during speech. Perception & Psychophysics, 11, 153–156. COLE, R. A., & JAKIMIK, J. (1978). Understanding speech: How words are heard. In G. Underwood (Ed.), Strategies of information processing. London: Academic Press. COLE, R. A., & JAKIMIK, J. (1980). A model of speech perception. In R. Cole (Ed.), Perception and production of fluent speech. Hillsdale, NJ: Erlbaum. COLE, R. A., JAKIMIK, J., & COOPER, W. E. (1978). Perceptibility of phonetic features in fluent speech. Journal of the Acoustical Society of America, 64, 44–46. CONNINE, C. M., BLASKO, D. G., & HALL, M. (1991). Effects of subsequent sentence context in auditory word recognition: Temporal and linguistic constraints. Journal of Memory and Language, 30, 234– 251. COOPER, W. E., & SORENSON, J. M. (1977). Fundamental frequency contours at syntactic boundaries. Journal of the Acoustical Society of America, 62, 683–692. CUTLER, A. (1976). Phoneme monitoring reaction time as a function of preceding intonation contour. Perception & Psychophysics, 20, 55–60. CUTLER, A. (1986). Forbear is a homophone: Lexical prosody does not constrain lexical access. Language and Speech, 29, 201–220. CUTLER, A. (1990). Exploiting prosodic probabilities in speech segmentation. In G. T. M. Altmann (Ed.), Cognitive models of speech processing: Psycholinguistic and computational perspectives (pp. 105– 121). Cambridge, MA: MIT Press. CUTLER, A. (1994). The perception of rhythm in language. Cognition, 50, 79–81. CUTLER, A., & BUTTERFIELD, S. (1992). Rhythmic cues to speech segmentation: Evidence from juncture misperception. Journal of Memory and Language, 31, 218–236. CUTLER, A., & CARTER, D. M. (1987). The predominance of strong initial syllables in the English vocabulary. Computer Speech and Language, 2, 133–142. CUTLER, A., & CLIFTON, C. E. (1985). The use of prosodic information in word recognition. In H. Bouma & D. G. Bouwhuis (Eds.), Attention and performance X. Hillsdale, NJ: Lawrence Erlbaum Associates.

AID

JML 2472

/

a005h$$147

CUTLER, A., MEHLER, J., NORRIS, D., & SEGUI, J. (1987). Phoneme identification and the lexicon. Cognitive Psychology, 19, 141–177. CUTLER, A., & NORRIS, D. (1979). Monitoring sentence comprehension. In W. E. Cooper & E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett (pp. 113–134). 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, 113–121. CUTTING, J. E. (1976). Auditory and linguistic processes in speech perception: Inference from six fusions in dichotic listening. Psychological Review, 83, 114– 140. DAY, R. S. (1968). Fusion in dichotic listening. Dissertation Abstracts International, 29, 2649B. (University Microfilms No. 69-211) FODOR, J. A. (1983). The modularity of mind. Cambridge, MA: MIT Press. FODOR, J. A., BEVER, T. G., & GARRETT, M. F. (1974). The psychology of language. New York: McGrawHill. FORSTER, K. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper & E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, NJ: Erlbaum. FRAUENFELDER, U. H., SEGUI, J., & DIJKSTRA, T. (1990). Lexical effects in phonemic processing: Facilitatory or inhibitory? Journal of Experimental Psychology: Human Perception and Performance, 16, 77–91. GREEN, D. G., & SWETS, J. (1966). Signal detection and theory of psychophysics. New York: Wiley. GROSJEAN, F., & GEE, J. P. (1987). Prosodic structure and spoken word recognition. Cognition, 25, 135–155. GLUCKSBERG, S., KREUZ, R. J., & RHO, S. H. (1986). Context can constrain lexical access: Implications for models of language comprehension. Journal of Experimental Psychology: Learning, Memory and Cognition, 12, 323–335. HALWES, G. T. (1969). Effects of dichotic fusion on the perception of speech. Status report to speech research, Supplement, September, Haskins Laboratories. HEUVEN, V. J. VAN (1985). Perception of stress pattern and recognition: Recognition of Dutch words with incorrect stress position. Journal of the Acoustical Society of America, 78, S21. KLATT, D. H., & STEVENS, K. H. (1973). On the automatic recognition of continuous speech: Implications from a spectrogram reading experiment, IEEE Trans. Audio Electroacoustic, 21, 210–217. KOLINSKY, R. (1988). La se´parabilite´ des proprie´te´s dans la perception des formes, Vol. II. Unpublished doctoral dissertation, Universite´ Libre de Bruxelles, 1988.

12-06-96 11:18:59

jmla

AP: JML

115

MIGRATIONS AND SPEECH PROCESSING KOLINSKY, R. (1989). The development of separability in visual perception. Cognition, 33, 243–284. KOLINSKY, R. (1992). Conjunction errors as a tool for the study of perceptual processes. In J. Alegria, D. Holender, J. Morais, & M. Radeau (Eds.), Analytic Approaches to Human Cognition, Amsterdam: North-Holland. KOLINSKY, R., MATTYS, S., MORAIS, J., CLUYTENS, M., & EVINCK, S. On the nature of speech migration errors: A study of the influence of acoustic and lexical variables. Manuscript in preparation. KOLINSKY, R., MORAIS, J., & VERHAEGHE, A. (1994). Visual separability: A study on unschooled adults. Perception, 23, 471–486. KOLINSKY, R., & MORAIS, J. (1992). Repre´sentations interme´diaires dans la reconnaissance de la parole: Apports de la technique de cre´ation de mots illusoires. In Actes des 19e`mes Journe´es d’Etude sur la Parole, Groupe de la Communication Parle´e, Brussels, May 19–22 (pp. 129–133). KOLINSKY, R., & MORAIS, J. (1993). Intermediate representations in spoken word recognition: A cross-linguistic study of word illusions. Proceedings of the 3rd European conference on speech communication and technology, Eurospeech’93, Berlin, Germany (pp. 731–734). KOLINSKY, R., MORAIS, J., & CLUYTENS, M. (1995). Intermediate representations in spoken word recognition: Evidence from word illusions. Journal of Memory and Language, 34, 19–40. KUCERA, H., & FRANCIS, W. (1967). Computational analysis of present-day American English. Providence, RI: Brown University Press. LEHISTE, I. (1960). An acoustic-phonetic study of internal open juncture. Phonetica, 5, (Suppl.) S. Karger, Basel, Switzerland. LEHISTE, I. (1972). The timing of utterances and linguistic boundaries. Journal of the Acoustical Society of America, 51, 2018–2024. LEWIS, J. W. (1972). A concise pronouncing Dictionary of British and American English. London: Oxford University Press. MARSLEN-WILSON, W. D. (1987). Functional parallelism in spoken word-recognition. Cognition, 25, 71–102. MARSLEN-WILSON, W. D., & TYLER, L. (1980). The temporal structure of spoken language understanding. Cognition, 8, 1–71. MARSLEN-WILSON, W. D., & WELSH, A. (1978). Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology, 10, 29–63. MASSARO, D. W. (1989). Testing between the TRACE model and the fuzzy logical model of speech perception. Cognitive Psychology, 21, 398–421. MASSARO, D. W., & COHEN, M. M. (1991). Integration versus interactive activation: The joint influence of stimulus and context in perception. Cognitive Psychology, 23, 558–614.

AID

JML 2472

/

a005h$$147

MCCLELLAND, J. M. (1991). Stochastic interactive processes and the effect of context on perception. Cognitive Psychology, 23, 1–44. MCCLELLAND, J. L., & ELMAN, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 19, 1–86. MEHLER, J., DOMMERGUES, S., FRAUENFELDER, U. H., & SEGUI, J. (1981). The syllable’s role in speech segmentation. Journal of Verbal Learning and Verbal Behavior, 20, 298–305. MORAIS, J., BERTELSON, P., CARY, L., & ALEGRIA, J. (1986). Literacy training and speech segmentation. Cognition, 24, 45–64. MORAIS, J., CARY, L., ALEGRIA, J., & BERTELSON, P. (1979). Does awareness of speech as a sequence of phones arise spontaneously? Cognition, 7, 323–331. MORAIS, J., CASTRO, S.-L., SCLIAR-CABRAL, L., KOLINSKY, R., & CONTENT, A. (1987). The effects of literacy on the recognition of dichotic words. Quarterly Journal of Experimental Psychology A, 39, 451–465. MORAIS, J., CASTRO, S.-L., & KOLINSKY, R. (1991). La reconnaissance des mots chez les adultes illetre´s. In R. Kolinsky, J. Morais, & J. Segui (Eds.), La reconnaissance des mots dans diffe´rentes modalite´s sensorielles: Etudes de psycholinguistique cognitive (pp. 59–80). Paris: Presses Universitaires de France. MORAIS, J., & KOLINSKY, R. (1994). Perception and awareness in phonological processing: The case of the phoneme. Cognition, 50, 287–297. MORAIS, J., KOLINSKY, R., CLUYTENS, M., & PAIVA, M. Speech attribute migration: A language-dependent but literacy-independent phenomenon. Manuscript in preparation. MORAIS, J., KOLINSKY, R., & NAKAMURA, M. (in press). The psychological reality of speech units in Japanese. In T. Otake & A. Cutler (Eds.), Phonological structure and language processing: Cross-linguistic studies. Berlin: Mouton de Gruyter. MORAIS, J., KOLINSKY, R., & PAIVA, M. The phoneme’s perceptual reality exhumed: Studies in Portuguese. Unpublished manuscript. MORTON, J. (1969). Interaction of information in word recognition. Psychological Review, 76, 165–178. NAKATANI, L. H., & DUKES, K. D. (1977). Locus of segmental cues for word juncture. Journal of the Acoustical Society of America, 62, 714–719. NAKATANI, L. H., & SCHAFFER, J. A. (1978). Hearing ‘words’ without words: Prosodic cues for word perception. Journal of the Acoustical Society of America, 63, 234–245. NORRIS, D. G. (1986). Word recognition: Context effects without priming. Cognition, 22, 93–136. NORRIS, D. G. (1995). Observation. Signal detection theory and modularity: On being sensitive to the power of bias models of semantic priming. Journal of Experimental Psychology: Human Perception and Performance, 21, 935–939. OTAKE, T., HATANO, G., CUTLER, A., & MEHLER, J.

12-06-96 11:18:59

jmla

AP: JML

116

MATTYS AND SAMUEL

(1993). Mora or syllable? Speech segmentation in Japanese. Journal of Memory and Language, 32, 258–278. PISONI, D. B., & LUCE, P. A. (1987). Acoustic-phonetic representations in word recognition. Cognition, 25, 21–52. PITT, M. A., & SAMUEL, A. G. (1995). Lexical and sublexical feedback in auditory word recognition. Cognitive Psychology 29, 149–188. PISONI, D. B., & MCNABB, S. D. (1974). Dichotic interactions of speech sounds and phonetic feature processing. Brain and Language, 1, 251–364. RATCLIFF, R., MCKOON, G., & VERWOERD, M. (1989). A bias interpretation of facilitation in perceptual identification. Journal of Experimental Psychology: Learning, Memory and Cognition, 15, 378–387. READ, C., ZHANG, Y., NIE, H., & DING, B. (1986). The ability to manipulate speech sounds depends on knowing alphabetic spelling. Cognition, 24, 31–44. REDDY, R. (1976). Speech recognition by machine: A review. Proc. IEEE, 64, 501–531. RUBIN, P., TURVEY, M. T., & VAN GELDER, P. (1976). Initial phonemes are detected faster in spoken words than in spoken nonwords. Perception & Psychophysics, 19, 394–398. SALASOO, A., & PISONI, B. (1985). Interaction of knowledge sources in spoken word identification. Journal of Memory and Language, 24, 210–231. SAMUEL, A. G. (1979). Speech is specialized, not special (Doctoral dissertation, University of California, San Diego, 1979). Dissertation Abstracts International, 40-08B. SAMUEL, A. G. (1981). Phonemic restoration: Insights from a new methodology. Journal of Experimental Psychology: General, 110, 474–494. SAMUEL, A. G. (1986). The role of the lexicon in speech perception. In E. Schwab & H. Nusbaum (Eds.), Pattern perception by human and machines: Speech per-

AID

JML 2472

/

a005h$$147

ception (Vol. 1, pp. 89–111). New York: Academic Press. SAMUEL, A. G. (1996). Does lexical information influence the perceptual restoration of phonemes? Journal of Experimental Psychology: General, 125, 28–51. SLOWIACZEK, L. M. (1990). Effects of lexical stress in auditory word recognition. Language and Speech, 33, 47–68. SLOWIACZEK, L. M. (1991). Stress and context in word recognition. Journal of Psycholinguistic Research, 20, 465–481. SVARTVIK, J., & QUIRK, R. (1980). A corpus of English conversation. Lund: Gleerup. TANENHAUS, M. K., CARLSON, G. N., & SEIDENBERG, M. S. (1985). Do listeners compute linguistic representations? In D. Dowty, L. Kartunnen, & A. Zwicky (Eds.), Natural Language parsing: Psychological, theoretical and computational perspectives. New York: Cambridge University Press. TANENHAUS, M. K., & LUCAS, M. M. (1987). Context effects in lexical processing. Cognition, 25, 213–234. TARTTER, V. C., & BLUMSTEIN, S. E. (1981). The effects of pitch and spectral differences on phonetic fusion in dichotic listening. Journal of Phonetics, 9, 251– 259. TREISMAN, A., & GELADE, G. (1980). A feature integration theory of attention. Cognitive Psychology, 12, 97–136. TREISMAN, A., & SCHMIDT, H. (1982). Illusory conjunctions in the perception of objects. Cognitive Psychology, 14, 107–141. TREISMAN, A., & SOUTHER, J. (1986). Illusory words: The roles of attention and top-down constraints in conjoining letters to form words. Journal of Experimental Psychology: Human Perception and Performance, 12, 3–17. (Received May 30, 1995) (Revision received November 11, 1995)

12-06-96 11:18:59

jmla

AP: JML