Copvrieht 1992 bv the American Psychological Association. Inc. ' 0278-7393/92/S3.OO
Journal of Experimental Psychology: Learning, Memory', and Cognition 1992. Vol. 18. No'. 6. 1239-1250
Prelexical Facilitation and Lexical Interference in Auditory Word Recognition Louisa M. Slowiaczek and Mary Beth Hamburger
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
University at Albany, State University of New York Phonological priming effects were examined in an auditory single-word shadowing task. In 6 experiments, target items were preceded by auditorily or visually presented, phonologically similar, word or nonword primes. Results revealed facilitation in response time when a target was preceded by a word or nonword prime having the same initial phoneme when the prime was auditorily presented but not when it was visually presented. Second, modality-independent interference was observed when the phonological overlap between the prime and target increased from 1 to 3 phonemes for word primes but not for nonword primes. Taken together, these studies suggest that phonological information facilitates word recognition as a result of excitation at a prelexical level and increases response time as a result of competition at a lexical level. These processes are best characterized by connectionist models of word recognition.
One of the major concerns of contemporary research in psycholinguistics has been the study of the structure and organization of lexical information in memory. Examination of this issue can be found in early work on semantic memory (Anderson, 1976, 1983; Collins & Loftus, 1975; Collins & Quillian, 1969; Rosch, 1975) as well as more recent developments in word recognition (Elman & McClelland, 1986; Grosjean & Gee, 1987; Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978; McClelland & Elman, 1986; Pisoni & Luce, 1987; Pisoni, Nusbaum, Luce, & Slowiaczek, 1985; Slowiaczek, 1990; Taft & Forster, 1975, 1976). The accumulated results from these areas of research have led to several generalizations about the structure of lexical information and the processes used to access that information from the lexicon. It has generally been accepted, for example, that representations of semantically related information are linked in memory. This conclusion is supported by research on semantic memory that has consistently found evidence of semantic priming effects (Foss, 1982; Marcel, 1983; Meyer & Schva-
The research reported here was supported by National Institutes of Health Grant NS-12179 to Indiana University in Bloomington, a Faculty Research Award from the University at Albany, State University of New York to the first author, and National Institutes of Health Grant NS-29286-01 to the University at Albany, State University of New York. Part of this research was reported at meetings of the Psychonomic Society in Boston, Massachusetts, November 1985, and San Francisco, California, November 1991. We would like to thank David Pisoni, Roddy Roediger, Dave Balota, R. Scott Tindale, Laurie Feldman, Jim Jaccard. and several anonymous reviewers for valuable comments on earlier versions of this article. We also thank Victoria Kopper, Emily Soltano, and Merle Hamburger for technical assistance and Vijay Vaidyanathan and John Munson for developing the experimental programs. This article has benefited greatly from extensive and valuable discussions we have had with Jim Neely, and we thank him. Correspondence concerning this article should be addressed to Louisa M. Slowiaczek, Department of Psychology, University at Albany, State University of New York. Albany. New York 12222.
neveldt, 1971; Meyer, Schvaneveldt, & Ruddy, 1975; Neely, 1976, 1977; Schvaneveldt, Meyer, & Becker, 1976). In particular, Meyer and Schvaneveldt (1971) found that reaction times to classify a letter string (e.g., NURSE) as a word were faster when the preceding letter string was a semantically related word (e.g., DOCTOR) rather than an unassociated word (e.g., BUTTER). Consistent effects of semantic priming in lexical decision (Meyer & Schvaneveldt, 1971; Neely, 1976, 1977), identification (Jackson & Morton, 1984), naming (Bock, 1986; Keefe & Neely, 1990; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982), and speeded verification tasks (Rips, Shoben, & Smith, 1973; Schaeffer & Wallace, 1969, 1970; Wilkins. 1971) have provided a rich source of data for the development of memory models that emphasize the relationship between semantically related items (Anderson, 1976, 1983; Collins & Loftus, 1975). More recent work using various word recognition tasks has suggested that the structure and organization of lexical memory involves more than semantic relatedness. In particular, evidence suggests that the phonotactic (Luce, 1986; Pisoni et al., 1985), phonological (Pisoni et al., 1985; Slowiaczek, Nusbaum, & Pisoni, 1987), orthographic (Jakimik, Cole, & Rudnicky, 1985), and prosodic (Grosjean & Gee, 1987; Slowiaczek, 1987, 1990) structure of words may constitute organizing dimensions in the lexicon and thus play an important role in word recognition. The current investigation further explores the impact of phonological information in word recognition. The possible role of phonological information in word recognition has been outlined in several recent models that rely on structural characteristics beyond semantic relatedness (Columbo, 1986; Grosjean & Gee, 1987; Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978; McClelland & Elman, 1986). For example, cohort theory (Marslen-Wilson, 1987; Marslen-Wilson & Tyler, 1980; Marslen-Wilson & Welsh, 1978) proposes that early "acoustic-phonetic" information in a speech signal activates a list of word candidates that begin with the same acoustic-phonetic sequence. Thus, if the stimulus word /blaek/ is to be recognized, a list or cohort of word 1239
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
1240
LOUISA M. SLaWIACZEK AND MARYBETH HAMBURGER
candidates including /blaek/, /bliyk/, /bol/, /blaend/, /bluwn/, and so forth would initially be activated. In a second stage, word candidates that diverge from additional stimulus input or are inconsistent with contextual information are deactivated. When only one candidate remains activated in the cohort, that candidate represents the word to be recognized. Because of the initial activation of word candidates in Stage 1 of the model, word recognition depends on phonological information at the beginning of the stimulus word. Providing a preview of the relevant initial phonemes could thus facilitate the word recognition process. Specifically, a prime word that begins with a particular phoneme (e.g., /b/) would activate a cohort of words beginning with that acousticphonetic information (e.g., the phoneme /b/). If the subsequent target word also began with /b/, its representation would have been previously activated when the initial cohort was identified with the prime. When the target is presented, residual activation of the word's representation from the presentation of the prime may facilitate recognition of the target (see, e.g., Nusbaum & Slowiaczek, 1983; Slowiaczek et al., 1987). Inhibitory priming is not predicted in this model given Marslen-Wilson's (1990) claim that competition effects dissipate by the end of a word. As a result, candidates in a cohort activated by the prime would not compete with candidates in a cohort subsequently activated by the target. Thus, inhibitory priming from the prime to the target would not occur. Alternatively, Columbo (1986) proposed a connectionist model in which lexical units' are activated by grapheme units that are contained in the word to be recognized. In Columbo's model (as in related models by Elman & McClelland, 1986, and McClelland & Rumelhart, 1986), grapheme units send activation to lexical units that contain the relevant letter, but, in addition, similar lexical units (e.g., word and work) compete or inhibit each other. Thus, as in cohort theory, low-level units in Columbo's model (i.e., grapheme units) play a facilitatory role at an early (prelexical) stage. However, the natureof the architecture in Columbo's model also results in inhibition because of the competition among orthographically similar word units at a later (lexical) stage. As a result, both facilitation and inhibition may be obtained under certain conditions in a connectionist model. In the current investigation we propose a model similar to Columbo's with the exception that input to the model is acoustic rather than orthographic. Research on the processing of spoken words (e.g., Jakimik et al., 1985; Slowiaczek et al., 1987; Slowiaczek & Pisoni, 1986) has provided seemingly conflicting results with regard to the nature of phonological information in word recognition. For example, Slowiaczek et al. (1987) observed a phonological priming effect when subjects were asked to identify target words embedded in noise that were preceded by phonologically related primes. Specifically, the probability of identifying a word increased with the number of shared phonemes between the prime and the target. Based on these findings, Slowiaczek et al. (1987) argued that the phonological structure of an utterance can produce a facilitatory priming effect and that the magnitude of priming depends on the number of phonemes shared between the prime and the target words.
These results are consistent with both Marslen-Wilson's cohort model and Columbo's connectionist account.2 In a related study, however, Slowiaczek and Pisoni (1986) did not observe phonological priming effects in a lexical decision task. Although they did find priming for identical prime-target pairs, they found no facilitation in response to targets preceded by primes that shared word-initial phonological information. Slowiaczek and Pisoni (1986) suggested that the discrepancy between their results and those reported by Slowiaczek et al. (1987) was due to methodological differences between lexical decision and identification in noise. However, a theoretically more interesting account of the differences in these results is to posit different levels of processing within the word recognition system, such as those suggested by a connectionist model, that are differentially tapped by lexical decision and identification in noise. Specifically, in the first level, phonological information plays a facilitatory role in early (prelexical) processing, where phoneme units partially activate lexical units consistent with the phoneme. The facilitation observed when using the identification-in-noise task may be the result of tapping this prelexical level of processing. Moreover, this prelexical level is described in cohort theory and by the excitatory, graphemic connections in Columbo's connectionist model. The second level within the word recognition system involves the complete activation of word units within the lexicon. It is unclear whether operations at the lexical level are directly measured by identification-in-noise or lexical decision tasks. Research that examines processes at the lexical level is needed to evaluate the effects of phonological information at this stage. Third, a postlexical level involves decisions concerning the identification of lexical items. This level does not rely on phonological information. In addition, because they are believed to be influenced by factors unrelated to the initial activation of lexical units, traditional models of word recognition do not usually focus on processes at this level. In the present investigation we sought to provide evidence regarding the nature of phonological priming in auditory word recognition processes and how it reflects the phonological organization of information in lexical memory. In each of the experiments reported here we used a shadowing procedure in which subjects repeated an auditorily presented target item.3 ' Columbo's (1986) model can be contrasted with other connectionist accounts of word recognition in which a lexical level is not included in the architecture (e.g., Seidenberg & McClelland, 1989). 2 Phonological priming effects for primes and targets sharing wordfinal information were also observed by Slowiaczek, Nusbaum, and Pisoni (1987). These results were not consistent with Marslen-Wilson's cohort model. 'Typically, a shadowing task is one in which subjects repeat auditorily presented sentences as quickly and accurately as possible upon hearing them (e.g., Marslen-Wilson & Welsh, 1978), and shadowing errors and latencies are measured. The task used in the current series of experiments is also similar to naming, a procedure commonly used in studies of visual word recognition. In the naming task, subjects usually say aloud or name a visually presented picture (Oldfield & Wingfield, 1965) or word (Forster & Chambers, 1973). The time it takes to name the stimulus is assumed to include the time it takes to access the word in the lexicon. The decision to refer to the task used
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
PRELEXICAL FACILITATION
The shadowing task was chosen for the present investigation to address the differences in results that were obtained in studies using a lexical decision task and an identification-innoise task. First, unlike the identification-in-noise task, reaction times can be obtained with the shadowing task, providing a potentially more sensitive measure of the time course of word recognition processes. Second, to demonstrate an effect of phonological priming it is necessary to choose a task that is sensitive to the use of phonological information at the lexical level. The lexical decision task used by Slowiaczek and Pisoni (1986) is often considered to involve substantial postlexical processing (Balota & Chumbley, 1984; Jakimik et al., 1985; Neely, 1991; Neely & Keefe, 1989; Norris, 1986) and hence may involve decisions at a more abstract level of analysis. The shadowing task, on the other hand, may be more sensitive to the use of phonological information during prelexical and lexical stages of word recognition because fewer postlexical processes are involved. Finally, the shadowing task provides a way to measure subject responses in a more "natural" word recognition environment. It has been suggested (Radeau, Morais, & Dewier, 1989) that tasks such as the identification of words in white noise, as used by Slowiaczek et al. (1987), may induce subjects to use decision-making strategies not used in normal spoken word recognition because the stimuli are degraded by white noise. Thus, in the present series of experiments subjects shadowed (repeated aloud) single, auditorily presented target words. The target words were preceded by phonologically related primes that shared varying numbers of initial phonemes with the target. If information in lexical memory is related phonologically, target shadowing times should be influenced by primes sharing initial phonemes with the target. More specifically, if the architecture of lexical memory is similar to that proposed by connectionist theorists (i.e., Columbo, 1986; McClelland & Elman, 1986), responses should show a decrease in facilitation (or possibly inhibition) as the prime and target share more phonemes. If the architecture is more similar to the cohort model proposed by MarslenWilson and his colleagues (Marslen-Wilson, 1987, 1990; Marslen-Wilson & Tyler, 1980; Marslen-Wilson & Welsh, 1978), shadowing times should show more facilitation as the number of shared phonemes increases. Experiment 1A In Experiment 1 A, we used the shadowing task with phonologically related prime items to test whether lexical information in memory is organized on the basis of phonological relatedness and whether the nature of that organization results in the present investigation as shadowing is somewhat arbitrary given the similarities between naming a visually presented word, shadowing auditorily presented sentences, and the procedure used in the experiments reported here. In the current set of experiments the task is referred to as shadowing because this term has been used by other researchers (e.g., Radeau, Morais, & Dewier, 1989; Slowiaczek, 1990) using the same task. The similarities to the literature involving the naming of visually presented items, however, should be noted because a similar rationale supports the use of both procedures.
1241
in only facilitatory phonological priming, as suggested by Marslen-Wilson's (1987) cohort theory, or both facilitatory and inhibitory phonological priming, as suggested by Colombo's (1986) connectionist model. Although several studies have reported facilitation alone when primes and targets share the initial phoneme (Goldinger, Luce, Pisoni, & Marcario, 1992; Slowiaczek et al., 1987), one study (using different procedures) has revealed possible inhibition as the number of shared phonemes increases (Radeau et al., 1989). Thus, it was important to determine whether both of these effects could be obtained under the same conditions.
Method Subjects. Twenty subjects were obtained from a paid subject file maintained in the Speech Research Laboratory at Indiana University. Subjects were paid $3.50 for their participation in the experiment. All subjects were native speakers of English with no reported history of hearing loss or speech disorder. Materials. One hundred monosyllabic words, four or five phonemes in length, were selected as target items for the present experiment. In addition, each of the 100 target words was paired with five types of primes. The primes were monosyllabic words related to a particular target in the following ways: (a) identical, (b) with the first three phonemes in common, (c) with the first two phonemes in common, (d) with the first phoneme in common, and (e) unrelated— no phonemes in common. The mean durations of the different stimulus items were as follows: targets—767 ms, "three-phoneme" primes—845 ms, "two-phoneme" primes—844 ms, "one-phoneme" primes—813 ms, and unrelated primes—832 ms. Table 1 lists examples of the targets and corresponding primes used in the experiment. A male speaker recorded the target and prime items in a soundattenuated booth (Industrial Acoustics Corporation, Model No. 106648) on one track of an audiotape. The recordings were made using an Electro Voice DO54 microphone and an Ampex AG500 tape deck. The stimulus items were produced in the carrier sentence "Say the word please." The stimulus items were then digitized at a sampling rate of 10 kHz using a 12-bit analog-to-digital converter, low-pass filtered at 4.8 kHz, and excised from the carrier sentence using a digitally controlled speech waveform editor (WAVES) on a PDP 11/34 computer (Luce & Carrell, 1981). Each target and each corresponding prime were stored digitally in separate stimulus files on a computer disk for later presentation to subjects in the experiment. Procedure. Subjects were tested individually. The presentation of stimuli was controlled by a PDP 11/34 computer. Signals were output via a 12-bit digital-to-analog converter and low-pass filtered at 4.8 kHz. Subjects heard the stimuli at 80 dB (SPL) over a pair of TDH 39 headphones. Subjects participated in two sessions of the experiment. In one session, the target words were presented in isolation (unprimed). In the other session, each target word was preceded by a prime word (i.e., identical, three-phoneme, two-phoneme, one-phoneme, or unrelated). Subjects participated in the second session immediately following participation in Session 1. The order of participation in the primed and unprimed sessions was counterbalanced across subjects. The subject's task in both sessions was to repeat the target item as quickly and accurately as possible out loud into a microphone. Latency from the onset of the target until the onset of the subject's vocalization was recorded. A typical trial sequence in the primed session proceeded as follows: The trial started with the presentation of the warning phrase "Get Ready for Next Trial" on a video display monitor positioned in front
1242
LOUISA M. SLOWIACZEK AND MARYBETH HAMBURGER
Table 1 Examples of Target Stimuli and Their Corresponding Word Primes Used in Experiments 1A and IB Primes
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Target still plague green
Identical still plague green
3 Phonemes stiff played grief
of the subject. One second after the warning phrase, the subject heard the prime item. Subjects were instructed to listen to the prime item but not to respond to it in any way. Five hundred milliseconds after the offset of the prime, the subject heard the target item over the headphones. The subject responded by repeating the target item out loud into the microphone. The subject's vocal response triggered a voice key, which stopped a clock that was controlled by an experimental program. An experimenter seated next to the subject in the testing room recorded the subject's response for each trial on a response sheet to determine response accuracy. A typical trial in the unprimed session proceeded in the same manner except the presentation of the prime was omitted from the trial sequence and the target word occurred 1 s after the warning phrase. Four subjects were tested on each of five prime-target stimulus sets for a total of 20 subjects. Each session of the experiment consisted of 100 trials. The target words were presented in random order within each session. During the primed session an equal number of words (20) was primed in each of the five priming conditions. Subjects never heard the same target or prime item twice on any of the 100 trials within a particular session of the experiment, although the same targets were used in both the primed and unprimed sessions. Across groups of subjects, every target appeared equally often in each of the five priming conditions.
Results and Discussion Response accuracy and response latency were analyzed separately. The percentages of correctly shadowed words for the unprimed session and each of the five priming conditions in the primed session were determined for each subject. An error in repeating a particular target resulted in the elimination of the response time for that item in both the unprimed and primed sessions for that subject. The mean error rate for the unprimed session was .03, and for the primed session it was .01. The error rates for the five priming conditions in the primed session are presented in Table 2. Additional analyses were not conducted on the error rates because of the consistently low number of errors. Unprimed versus primed sessions. The mean shadowing time for the unprimed session was 852 ms. The mean shadowing times for the five priming conditions in the primed session are presented in Table 2. Analyses of variance (Session Order x Session x Priming Condition) by subjects (Fi) and by items (F2) were conducted on the response time data; in these analyses each target in the unprimed session was assigned to the priming condition in which it occurred in the primed session for a particular subject. The degrees of freedom were modified using the Huynh-Feldt correction to adjust for possible violations of the sphericity assumption common in
2 Phonemes steep plead grope
1 Phoneme smoke pants goals
Unrelated dream dance clump
repeated measures designs (Jaccard & Ackerman, 1985; Maxwell & Avery, 1982). The main effect of session order was not significant in the subject analysis, F,(l, 18) = 1.93, MS, = 187,820, although it was significant in the item analysis, F2(l, 94) = 925, MSe = 4,003, p < .001. The main effect of session was significant in both analyses, F,(l, 18) = 8.80, MSe = 15,216, p < .008, F 2 (l, 94) = 265, MS, = 4,996, p < .001. The main effect of priming condition was not found in either the subject, Fi(4, 72) = 1.58, MS, = 1,089, or item analyses, F2 < 1.0. Of the possible interactions, only the Session Order x Session, JFI(1, 18) = 4.65, MSe = 15,216,/?
Journal of Experimental Psychology: Learning, Memory', and Cognition 1992. Vol. 18. No'. 6. 1239-1250
Prelexical Facilitation and Lexical Interference in Auditory Word Recognition Louisa M. Slowiaczek and Mary Beth Hamburger
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
University at Albany, State University of New York Phonological priming effects were examined in an auditory single-word shadowing task. In 6 experiments, target items were preceded by auditorily or visually presented, phonologically similar, word or nonword primes. Results revealed facilitation in response time when a target was preceded by a word or nonword prime having the same initial phoneme when the prime was auditorily presented but not when it was visually presented. Second, modality-independent interference was observed when the phonological overlap between the prime and target increased from 1 to 3 phonemes for word primes but not for nonword primes. Taken together, these studies suggest that phonological information facilitates word recognition as a result of excitation at a prelexical level and increases response time as a result of competition at a lexical level. These processes are best characterized by connectionist models of word recognition.
One of the major concerns of contemporary research in psycholinguistics has been the study of the structure and organization of lexical information in memory. Examination of this issue can be found in early work on semantic memory (Anderson, 1976, 1983; Collins & Loftus, 1975; Collins & Quillian, 1969; Rosch, 1975) as well as more recent developments in word recognition (Elman & McClelland, 1986; Grosjean & Gee, 1987; Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978; McClelland & Elman, 1986; Pisoni & Luce, 1987; Pisoni, Nusbaum, Luce, & Slowiaczek, 1985; Slowiaczek, 1990; Taft & Forster, 1975, 1976). The accumulated results from these areas of research have led to several generalizations about the structure of lexical information and the processes used to access that information from the lexicon. It has generally been accepted, for example, that representations of semantically related information are linked in memory. This conclusion is supported by research on semantic memory that has consistently found evidence of semantic priming effects (Foss, 1982; Marcel, 1983; Meyer & Schva-
The research reported here was supported by National Institutes of Health Grant NS-12179 to Indiana University in Bloomington, a Faculty Research Award from the University at Albany, State University of New York to the first author, and National Institutes of Health Grant NS-29286-01 to the University at Albany, State University of New York. Part of this research was reported at meetings of the Psychonomic Society in Boston, Massachusetts, November 1985, and San Francisco, California, November 1991. We would like to thank David Pisoni, Roddy Roediger, Dave Balota, R. Scott Tindale, Laurie Feldman, Jim Jaccard. and several anonymous reviewers for valuable comments on earlier versions of this article. We also thank Victoria Kopper, Emily Soltano, and Merle Hamburger for technical assistance and Vijay Vaidyanathan and John Munson for developing the experimental programs. This article has benefited greatly from extensive and valuable discussions we have had with Jim Neely, and we thank him. Correspondence concerning this article should be addressed to Louisa M. Slowiaczek, Department of Psychology, University at Albany, State University of New York. Albany. New York 12222.
neveldt, 1971; Meyer, Schvaneveldt, & Ruddy, 1975; Neely, 1976, 1977; Schvaneveldt, Meyer, & Becker, 1976). In particular, Meyer and Schvaneveldt (1971) found that reaction times to classify a letter string (e.g., NURSE) as a word were faster when the preceding letter string was a semantically related word (e.g., DOCTOR) rather than an unassociated word (e.g., BUTTER). Consistent effects of semantic priming in lexical decision (Meyer & Schvaneveldt, 1971; Neely, 1976, 1977), identification (Jackson & Morton, 1984), naming (Bock, 1986; Keefe & Neely, 1990; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982), and speeded verification tasks (Rips, Shoben, & Smith, 1973; Schaeffer & Wallace, 1969, 1970; Wilkins. 1971) have provided a rich source of data for the development of memory models that emphasize the relationship between semantically related items (Anderson, 1976, 1983; Collins & Loftus, 1975). More recent work using various word recognition tasks has suggested that the structure and organization of lexical memory involves more than semantic relatedness. In particular, evidence suggests that the phonotactic (Luce, 1986; Pisoni et al., 1985), phonological (Pisoni et al., 1985; Slowiaczek, Nusbaum, & Pisoni, 1987), orthographic (Jakimik, Cole, & Rudnicky, 1985), and prosodic (Grosjean & Gee, 1987; Slowiaczek, 1987, 1990) structure of words may constitute organizing dimensions in the lexicon and thus play an important role in word recognition. The current investigation further explores the impact of phonological information in word recognition. The possible role of phonological information in word recognition has been outlined in several recent models that rely on structural characteristics beyond semantic relatedness (Columbo, 1986; Grosjean & Gee, 1987; Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978; McClelland & Elman, 1986). For example, cohort theory (Marslen-Wilson, 1987; Marslen-Wilson & Tyler, 1980; Marslen-Wilson & Welsh, 1978) proposes that early "acoustic-phonetic" information in a speech signal activates a list of word candidates that begin with the same acoustic-phonetic sequence. Thus, if the stimulus word /blaek/ is to be recognized, a list or cohort of word 1239
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
1240
LOUISA M. SLaWIACZEK AND MARYBETH HAMBURGER
candidates including /blaek/, /bliyk/, /bol/, /blaend/, /bluwn/, and so forth would initially be activated. In a second stage, word candidates that diverge from additional stimulus input or are inconsistent with contextual information are deactivated. When only one candidate remains activated in the cohort, that candidate represents the word to be recognized. Because of the initial activation of word candidates in Stage 1 of the model, word recognition depends on phonological information at the beginning of the stimulus word. Providing a preview of the relevant initial phonemes could thus facilitate the word recognition process. Specifically, a prime word that begins with a particular phoneme (e.g., /b/) would activate a cohort of words beginning with that acousticphonetic information (e.g., the phoneme /b/). If the subsequent target word also began with /b/, its representation would have been previously activated when the initial cohort was identified with the prime. When the target is presented, residual activation of the word's representation from the presentation of the prime may facilitate recognition of the target (see, e.g., Nusbaum & Slowiaczek, 1983; Slowiaczek et al., 1987). Inhibitory priming is not predicted in this model given Marslen-Wilson's (1990) claim that competition effects dissipate by the end of a word. As a result, candidates in a cohort activated by the prime would not compete with candidates in a cohort subsequently activated by the target. Thus, inhibitory priming from the prime to the target would not occur. Alternatively, Columbo (1986) proposed a connectionist model in which lexical units' are activated by grapheme units that are contained in the word to be recognized. In Columbo's model (as in related models by Elman & McClelland, 1986, and McClelland & Rumelhart, 1986), grapheme units send activation to lexical units that contain the relevant letter, but, in addition, similar lexical units (e.g., word and work) compete or inhibit each other. Thus, as in cohort theory, low-level units in Columbo's model (i.e., grapheme units) play a facilitatory role at an early (prelexical) stage. However, the natureof the architecture in Columbo's model also results in inhibition because of the competition among orthographically similar word units at a later (lexical) stage. As a result, both facilitation and inhibition may be obtained under certain conditions in a connectionist model. In the current investigation we propose a model similar to Columbo's with the exception that input to the model is acoustic rather than orthographic. Research on the processing of spoken words (e.g., Jakimik et al., 1985; Slowiaczek et al., 1987; Slowiaczek & Pisoni, 1986) has provided seemingly conflicting results with regard to the nature of phonological information in word recognition. For example, Slowiaczek et al. (1987) observed a phonological priming effect when subjects were asked to identify target words embedded in noise that were preceded by phonologically related primes. Specifically, the probability of identifying a word increased with the number of shared phonemes between the prime and the target. Based on these findings, Slowiaczek et al. (1987) argued that the phonological structure of an utterance can produce a facilitatory priming effect and that the magnitude of priming depends on the number of phonemes shared between the prime and the target words.
These results are consistent with both Marslen-Wilson's cohort model and Columbo's connectionist account.2 In a related study, however, Slowiaczek and Pisoni (1986) did not observe phonological priming effects in a lexical decision task. Although they did find priming for identical prime-target pairs, they found no facilitation in response to targets preceded by primes that shared word-initial phonological information. Slowiaczek and Pisoni (1986) suggested that the discrepancy between their results and those reported by Slowiaczek et al. (1987) was due to methodological differences between lexical decision and identification in noise. However, a theoretically more interesting account of the differences in these results is to posit different levels of processing within the word recognition system, such as those suggested by a connectionist model, that are differentially tapped by lexical decision and identification in noise. Specifically, in the first level, phonological information plays a facilitatory role in early (prelexical) processing, where phoneme units partially activate lexical units consistent with the phoneme. The facilitation observed when using the identification-in-noise task may be the result of tapping this prelexical level of processing. Moreover, this prelexical level is described in cohort theory and by the excitatory, graphemic connections in Columbo's connectionist model. The second level within the word recognition system involves the complete activation of word units within the lexicon. It is unclear whether operations at the lexical level are directly measured by identification-in-noise or lexical decision tasks. Research that examines processes at the lexical level is needed to evaluate the effects of phonological information at this stage. Third, a postlexical level involves decisions concerning the identification of lexical items. This level does not rely on phonological information. In addition, because they are believed to be influenced by factors unrelated to the initial activation of lexical units, traditional models of word recognition do not usually focus on processes at this level. In the present investigation we sought to provide evidence regarding the nature of phonological priming in auditory word recognition processes and how it reflects the phonological organization of information in lexical memory. In each of the experiments reported here we used a shadowing procedure in which subjects repeated an auditorily presented target item.3 ' Columbo's (1986) model can be contrasted with other connectionist accounts of word recognition in which a lexical level is not included in the architecture (e.g., Seidenberg & McClelland, 1989). 2 Phonological priming effects for primes and targets sharing wordfinal information were also observed by Slowiaczek, Nusbaum, and Pisoni (1987). These results were not consistent with Marslen-Wilson's cohort model. 'Typically, a shadowing task is one in which subjects repeat auditorily presented sentences as quickly and accurately as possible upon hearing them (e.g., Marslen-Wilson & Welsh, 1978), and shadowing errors and latencies are measured. The task used in the current series of experiments is also similar to naming, a procedure commonly used in studies of visual word recognition. In the naming task, subjects usually say aloud or name a visually presented picture (Oldfield & Wingfield, 1965) or word (Forster & Chambers, 1973). The time it takes to name the stimulus is assumed to include the time it takes to access the word in the lexicon. The decision to refer to the task used
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
PRELEXICAL FACILITATION
The shadowing task was chosen for the present investigation to address the differences in results that were obtained in studies using a lexical decision task and an identification-innoise task. First, unlike the identification-in-noise task, reaction times can be obtained with the shadowing task, providing a potentially more sensitive measure of the time course of word recognition processes. Second, to demonstrate an effect of phonological priming it is necessary to choose a task that is sensitive to the use of phonological information at the lexical level. The lexical decision task used by Slowiaczek and Pisoni (1986) is often considered to involve substantial postlexical processing (Balota & Chumbley, 1984; Jakimik et al., 1985; Neely, 1991; Neely & Keefe, 1989; Norris, 1986) and hence may involve decisions at a more abstract level of analysis. The shadowing task, on the other hand, may be more sensitive to the use of phonological information during prelexical and lexical stages of word recognition because fewer postlexical processes are involved. Finally, the shadowing task provides a way to measure subject responses in a more "natural" word recognition environment. It has been suggested (Radeau, Morais, & Dewier, 1989) that tasks such as the identification of words in white noise, as used by Slowiaczek et al. (1987), may induce subjects to use decision-making strategies not used in normal spoken word recognition because the stimuli are degraded by white noise. Thus, in the present series of experiments subjects shadowed (repeated aloud) single, auditorily presented target words. The target words were preceded by phonologically related primes that shared varying numbers of initial phonemes with the target. If information in lexical memory is related phonologically, target shadowing times should be influenced by primes sharing initial phonemes with the target. More specifically, if the architecture of lexical memory is similar to that proposed by connectionist theorists (i.e., Columbo, 1986; McClelland & Elman, 1986), responses should show a decrease in facilitation (or possibly inhibition) as the prime and target share more phonemes. If the architecture is more similar to the cohort model proposed by MarslenWilson and his colleagues (Marslen-Wilson, 1987, 1990; Marslen-Wilson & Tyler, 1980; Marslen-Wilson & Welsh, 1978), shadowing times should show more facilitation as the number of shared phonemes increases. Experiment 1A In Experiment 1 A, we used the shadowing task with phonologically related prime items to test whether lexical information in memory is organized on the basis of phonological relatedness and whether the nature of that organization results in the present investigation as shadowing is somewhat arbitrary given the similarities between naming a visually presented word, shadowing auditorily presented sentences, and the procedure used in the experiments reported here. In the current set of experiments the task is referred to as shadowing because this term has been used by other researchers (e.g., Radeau, Morais, & Dewier, 1989; Slowiaczek, 1990) using the same task. The similarities to the literature involving the naming of visually presented items, however, should be noted because a similar rationale supports the use of both procedures.
1241
in only facilitatory phonological priming, as suggested by Marslen-Wilson's (1987) cohort theory, or both facilitatory and inhibitory phonological priming, as suggested by Colombo's (1986) connectionist model. Although several studies have reported facilitation alone when primes and targets share the initial phoneme (Goldinger, Luce, Pisoni, & Marcario, 1992; Slowiaczek et al., 1987), one study (using different procedures) has revealed possible inhibition as the number of shared phonemes increases (Radeau et al., 1989). Thus, it was important to determine whether both of these effects could be obtained under the same conditions.
Method Subjects. Twenty subjects were obtained from a paid subject file maintained in the Speech Research Laboratory at Indiana University. Subjects were paid $3.50 for their participation in the experiment. All subjects were native speakers of English with no reported history of hearing loss or speech disorder. Materials. One hundred monosyllabic words, four or five phonemes in length, were selected as target items for the present experiment. In addition, each of the 100 target words was paired with five types of primes. The primes were monosyllabic words related to a particular target in the following ways: (a) identical, (b) with the first three phonemes in common, (c) with the first two phonemes in common, (d) with the first phoneme in common, and (e) unrelated— no phonemes in common. The mean durations of the different stimulus items were as follows: targets—767 ms, "three-phoneme" primes—845 ms, "two-phoneme" primes—844 ms, "one-phoneme" primes—813 ms, and unrelated primes—832 ms. Table 1 lists examples of the targets and corresponding primes used in the experiment. A male speaker recorded the target and prime items in a soundattenuated booth (Industrial Acoustics Corporation, Model No. 106648) on one track of an audiotape. The recordings were made using an Electro Voice DO54 microphone and an Ampex AG500 tape deck. The stimulus items were produced in the carrier sentence "Say the word please." The stimulus items were then digitized at a sampling rate of 10 kHz using a 12-bit analog-to-digital converter, low-pass filtered at 4.8 kHz, and excised from the carrier sentence using a digitally controlled speech waveform editor (WAVES) on a PDP 11/34 computer (Luce & Carrell, 1981). Each target and each corresponding prime were stored digitally in separate stimulus files on a computer disk for later presentation to subjects in the experiment. Procedure. Subjects were tested individually. The presentation of stimuli was controlled by a PDP 11/34 computer. Signals were output via a 12-bit digital-to-analog converter and low-pass filtered at 4.8 kHz. Subjects heard the stimuli at 80 dB (SPL) over a pair of TDH 39 headphones. Subjects participated in two sessions of the experiment. In one session, the target words were presented in isolation (unprimed). In the other session, each target word was preceded by a prime word (i.e., identical, three-phoneme, two-phoneme, one-phoneme, or unrelated). Subjects participated in the second session immediately following participation in Session 1. The order of participation in the primed and unprimed sessions was counterbalanced across subjects. The subject's task in both sessions was to repeat the target item as quickly and accurately as possible out loud into a microphone. Latency from the onset of the target until the onset of the subject's vocalization was recorded. A typical trial sequence in the primed session proceeded as follows: The trial started with the presentation of the warning phrase "Get Ready for Next Trial" on a video display monitor positioned in front
1242
LOUISA M. SLOWIACZEK AND MARYBETH HAMBURGER
Table 1 Examples of Target Stimuli and Their Corresponding Word Primes Used in Experiments 1A and IB Primes
This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Target still plague green
Identical still plague green
3 Phonemes stiff played grief
of the subject. One second after the warning phrase, the subject heard the prime item. Subjects were instructed to listen to the prime item but not to respond to it in any way. Five hundred milliseconds after the offset of the prime, the subject heard the target item over the headphones. The subject responded by repeating the target item out loud into the microphone. The subject's vocal response triggered a voice key, which stopped a clock that was controlled by an experimental program. An experimenter seated next to the subject in the testing room recorded the subject's response for each trial on a response sheet to determine response accuracy. A typical trial in the unprimed session proceeded in the same manner except the presentation of the prime was omitted from the trial sequence and the target word occurred 1 s after the warning phrase. Four subjects were tested on each of five prime-target stimulus sets for a total of 20 subjects. Each session of the experiment consisted of 100 trials. The target words were presented in random order within each session. During the primed session an equal number of words (20) was primed in each of the five priming conditions. Subjects never heard the same target or prime item twice on any of the 100 trials within a particular session of the experiment, although the same targets were used in both the primed and unprimed sessions. Across groups of subjects, every target appeared equally often in each of the five priming conditions.
Results and Discussion Response accuracy and response latency were analyzed separately. The percentages of correctly shadowed words for the unprimed session and each of the five priming conditions in the primed session were determined for each subject. An error in repeating a particular target resulted in the elimination of the response time for that item in both the unprimed and primed sessions for that subject. The mean error rate for the unprimed session was .03, and for the primed session it was .01. The error rates for the five priming conditions in the primed session are presented in Table 2. Additional analyses were not conducted on the error rates because of the consistently low number of errors. Unprimed versus primed sessions. The mean shadowing time for the unprimed session was 852 ms. The mean shadowing times for the five priming conditions in the primed session are presented in Table 2. Analyses of variance (Session Order x Session x Priming Condition) by subjects (Fi) and by items (F2) were conducted on the response time data; in these analyses each target in the unprimed session was assigned to the priming condition in which it occurred in the primed session for a particular subject. The degrees of freedom were modified using the Huynh-Feldt correction to adjust for possible violations of the sphericity assumption common in
2 Phonemes steep plead grope
1 Phoneme smoke pants goals
Unrelated dream dance clump
repeated measures designs (Jaccard & Ackerman, 1985; Maxwell & Avery, 1982). The main effect of session order was not significant in the subject analysis, F,(l, 18) = 1.93, MS, = 187,820, although it was significant in the item analysis, F2(l, 94) = 925, MSe = 4,003, p < .001. The main effect of session was significant in both analyses, F,(l, 18) = 8.80, MSe = 15,216, p < .008, F 2 (l, 94) = 265, MS, = 4,996, p < .001. The main effect of priming condition was not found in either the subject, Fi(4, 72) = 1.58, MS, = 1,089, or item analyses, F2 < 1.0. Of the possible interactions, only the Session Order x Session, JFI(1, 18) = 4.65, MSe = 15,216,/?
