Early and late Spanish-English bilinguals' acquisition of English word stress patterns *

Transcription

1 1 Early and late Spanish-English bilinguals' acquisition of English word stress patterns * Susan G. Guion Department of Linguistics University of Oregon Tetsuo Harada Department of East Asian Languages and Literatures University of Oregon J.J. Clark Department of Linguistics University of Oregon Short title: Stress patterns in bilinguals Address for correspondence: Susan G. Guion Department of Linguistics 1290 University of Oregon Eugene, OR guion@uoregon.edu

2 2 Abstract Guion et al. (in press) found that three factors affect English speakers' stress placement on bisyllabic non-words: syllabic structure, lexical class, and stress patterns of phonologically similar real words. The current replication and extension included three groups (N=30): native English speakers, early Spanish-English bilinguals, late Spanish-English bilinguals. Participants produced and gave perceptual judgments on 40 non-words of varying syllabic structures in noun and verb sentence frames. A regression analysis used the three factors to predict stress placement in production and perception. All three groups showed significant effects from stress patterns of phonologically similar real words and lexical class. The effect of syllabic structure for early bilinguals was slightly different from that of native speakers and late bilinguals showed greatly reduced effects. Late bilinguals exhibited more initial stress overall, possibly due to L1 transfer. These results run counter to the prediction made by Long (1990) about age effects on phonological acquisition. [150 words]

3 3 Many studies have shown that late learners tend to speak a second language with a heavier accent than early learners. The age of second language acquisition seems to be the most important predictor of degree of foreign accent (Flege, Yeni-Komshian, & Liu, 1999; see also Piske, MacKay, & Flege, 2001, for a review). 1 Drawing conclusions from a comprehensive review, Long (1990) proposed that a native-like accent would be impossible unless first exposure was quite early, before six years of age for most individuals, and no individual who began learning a second language after the age of 12 would be able to attain a native-like accent. A foreign accent may be caused by non native-like pronunciation of segmental elements (e.g., consonants and vowels) as well as prosodic elements (e.g., word stress and intonation) of the second language phonological system. While many studies have been conducted comparing the acquisition of segmental production by early and late learners (Flege, 1991, 1993; Flege, MacKay, & Meador, 1999; Flege, Munro, & MacKay 1995b; Guion, 2003; Munro, Flege, & MacKay, 1996; among others), relatively few studies (see review in the following section) have been conducted comparing prosodic acquisition by early and late learners. It is an aim of the current study to investigate the acquisition of elements of the English word stress system by early and late Spanish-English bilinguals. Another aim of the current study is to investigate the ability of early and late learners to acquire different aspects of the word stress system in English. It has been shown that late learners are more successful in acquiring some aspects of the grammatical system of the second language than others (Neville, 1999; Weber-Fox & Neville, 1999). Thus, it is probable that late bilinguals may be more likely to acquire some aspects of English word stress than others. The ability of early and late Spanish-English bilinguals to acquire the relatively complex patterns associated with syllabic structure and the relatively simple pattern associated with lexical class, as well as the ability for analogical extension of stress patterns from previously acquired words will be studied.

4 4 Previous Studies on Second Language Acquisition of English Stress The majority of previous work on second language acquisition of English stress has focused on the transfer of first language rules or on the learning of second language rules for stress placement. The rules of stress placement are taken from the phonological literature describing and modeling stress patterns based on syllable and foot structure. However, empirical investigation into the cognitive reality of the proposed phonological rules for native speakers is rarely conducted. Some studies have suggested that second language learners can acquire new stress rules, although the patterns may differ from the phonological models proposed for the target language. Erdmann (1973) had late German learners of English read lists of real English adjectives; some words were known and others were not. The participants showed patterns in their stress placement that differed from predictions made by both English and German rules of stress placement found in the phonological literature. Similarly, Mairs (1989) recorded native Spanish speakers who were studying English in the United States and who were proficient enough to hold a conversation in English reading English words, sentences and paragraphs which contained real English target words. Some words were stressed in a native-like fashion by the participants and others were not. The incorrect forms showed a pattern of stress placement that Mairs argued was a product of learning a modified set of English stress rules (modified with respect to native speakers). Other studies have found that the stress patterns of the first language may transfer into the second language. Archibald (1992) had adult native Polish speakers living in an Englishspeaking country who had studied English for one month to six years read real English words in isolation and in sentences. Participants also listened to English words and noted stress placement. The placement of stress showed regularities that could be attributed to the transfer of Polish stress patterns. Archibald (1993) investigated the production of real English words read by native Spanish adult learners of English and noted that errors patterned by syllabic structure type. Archibald proposed that the errors were due to the transfer of Spanish stress rules but noted

5 5 that, in the case of correctly stressed words, it was difficult to determine if transferred Spanish stress rules or (mis-) learned English stress rules were applying. The above studies focused on native speakers of stress languages and found evidence for some acquisition or transfer of stress patterns. However, in a study with native speakers of the non-stress languages Chinese and Japanese reading real English words, Archibald (1997) found that the errors did not seem to be related to syllabic structure or lexical class. In fact the errors did not have any readily discernable pattern. Thus, it seems that native speakers of a stress language are more likely to show evidence of learning stress patterns than first language speakers of a non-stress language. In contrast to the above studies, which focused on late bilinguals, Pater (1997) investigated the acquisition of stress patterns by early bilinguals and found that even early learners did not show much evidence of learning second language stress patterns. In this study, early French learners of English were recorded reading non-words, all in a noun context at the beginning of a sentence. The results showed little consistency in stress placement. For the most part, errors were non-systematic. In the few forms that showed systematic stress placement, patterns were similar to English in some cases and in other cases the patterns were unlike English or French. This study also included a native English control group. Interestingly, the stress placement for the control group did not follow many of the predictions for stress placement made by phonological models either. Pater suggested that native speakers were likely assigning stress based on analogy with real English words. Several aspects of the above studies make them difficult to interpret. First, most assume that the stress rules proposed in the phonological literature are actually at work in native speakers but do no empirical investigation to support this assumption. In the one case in which data from English speakers was also collected (Pater, 1997), it seems that factors other than the stress predictions made by phonological models were also at work. Second, the interpretation that correctly stressed words are a product of rule learning is not the only plausible interpretation: Word stress could be memorized on a word-by-word basis (c.f. Pater, 1997, for this critique).

6 6 Third, both correct and incorrect stress placement could be due either to analogy with similar lexical items or to knowledge of stress patterns. The studies provide no method for determining whether knowledge of stress patterns, word-by-word learning, or analogy with similar words best fits the cognitive reality of the learners. Fourth, none of the studies allowed for a comparison between early and late learners. Fifth, age of acquisition, length of exposure to, and/or proficiency in the target language were not reported in several of the studies. The possibility that English learners could detect and learn statistical patterns of stress placement has received relatively less attention than the investigation of phonological rule acquisition. As detailed below, English bisyllabic nouns are more likely to have initial syllable stress and bisyllabic verbs are more likely to have final stress. Native English speakers have been shown to be sensitive to this distributional property and to demonstrate effects of its knowledge in experiments with non-words (Baker & Smith, 1976; Guion, Clark, Harada & Wayland, in press). In a study specifically designed to investigate second language learners ability to learn statistical lexical patterns, Davis and Kelly (1997) tested both native and nonnative English speakers knowledge about the tendency for differing stress placement for nouns and verbs. In their first experiment, they took the important step of determining whether nonnative speakers were able to hear stress placement in English, a prerequisite for learning any stress pattern. They found that non-native English speakers were able to hear stress placement on spoken English words at near-native levels. This was the case for native speakers of non-stress languages such as Cantonese, Japanese, and Korean as well as for native speakers of stress languages such as Finnish, German, and Spanish. In addition, they found that age of arrival in an English-speaking country (range of 5-27 yrs.) and length of residence (range of yrs.) were not correlated with accuracy in perception. In a second experiment, Davis and Kelly presented the same participants with bisyllabic English non-words varying in stress placement and asked the participants to use the non-word in a sentence in any way they wanted. They found that, for both natives and non-natives, words with final stress were more likely to be used in a sentence as

7 7 verbs than words with initial stress were. Again, age of arrival and length of residence were not correlated with performance. Davis and Kelly (1997) conducted a third experiment with another set of nonnative speakers from a variety of linguistic backgrounds and with varying age of acquisition and length of residence in an English-speaking country. Subjects performed a speeded classification of bisyllabic real English words as nouns or verbs. Half of the nouns and verbs had initial stress. The nonnative participants made more errors and were slower classifying nouns with final stress than nouns with initial stress and verbs with initial stress than verbs with final stress. Davis and Kelly interpreted these results to mean that the nonnative speakers had learned the English nounverb stress difference. The current study extends research into the second language acquisition of English stress patterns in several ways. First, the rules for English stress described in phonological models are not simply assumed to be part of native speakers' knowledge. Instead, native English controls are used to determine what sort of stress patterns native English speakers have knowledge of. Second, two experimental groups are included in the study, early and late learners, and are matched on English proficiency. While length of residence and amount of education in an English environment is not equivalent between the two groups, all participants have been immersed in an English-speaking environment for a minimum of five years. Third, the predicted effects of English syllabic structure as well as knowledge of the noun-verb stress difference are independently manipulated. Fourth, the possibility that non-words are assigned stress by analogy to similar real words is investigated by collecting phonologically similar real words from the participants. A regression analysis is used to evaluate the relative contribution of three variables (syllabic structure, lexical class and phonological similarity) on the placement of stress in nonwords. Finally, many previous studies investigating knowledge of stress patterns have relied on production data alone (Davis & Kelly, 1997, notwithstanding). It is possible that learners have acquired knowledge of English stress patterns, but that this knowledge may not be robust enough

8 8 to use in an on-line production task. For this reason, a perceptual judgment task, in addition to a production task, is administered in the current study. Such a task may prove to be a more sensitive measure, as no production demands are made on the processing system. The Three Factors Under Investigation The three factors affecting stress placement on non-words to be investigated in the current study vary in their complexity of patterning. The factor of syllabic structure is the most complex, while the factor of lexical class is relatively simpler. The third factor concerns the effect of stress placement on real words that are phonologically similar to the stimulus nonwords. Each factor is motivated below and the effect that the factor had on stress placement with native English speakers in Guion et al. (in press), which used the same method as the current study, is reported. Syllabic Structure. Distributional descriptions of English stress patterns suggest that the structure of syllables within a word affects stress placement for that word. Syllable weight is an important determinant of stress placement. A syllable is considered heavy if it has a long vowel or is closed by a coda consonant or consonant cluster. A standard analysis of English predicts stress based on vowel length and number of coda consonants (see, e.g., Chomsky & Halle, 1968; Hayes, 1982). In the case of two-syllable nouns and verbs, the objects of investigation in the current study, stress regularly adheres to the following patterns, although there are many exceptions. For two-syllable verbs, the final syllable will receive main stress if it has a long vowel or ends in at least two consonants. If the final syllable does not have a long vowel or end in two or more consonants, the initial syllable is stressed. For two-syllable nouns, the final syllable will receive main stress only if it has a long vowel. Otherwise, the initial syllable will have the main stress. Examples are given in Table 1. Table 1 approx. here

9 9 However, the standard analysis of word stress does not give the whole picture of syllabic structure effects on stress placement in English. Long vowels display a statistical tendency to be stressed more often than short vowels, regardless of the presence of coda consonants. Guion et al. (in press) reported an investigation of the CELEX lexical database (Baayen, Piepenbrock, & Gulikers, 1995) whose results indicated that long vowels were roughly twice as likely to be stressed as short vowels (60% of long vowels vs. 35% of short vowels were stressed). Thus, patterns associated with syllabic structure consist of a complex array of variables including vowel length, number of coda consonants and metrical structure (i.e., position of the syllable within the word). In the previous study (Guion et al., in press), stress assignment by native English speakers was affected by the syllabic structure of the non-words, most notably vowel length. In both production and perception experiments, syllables with long vowels were more likely to be assigned stress than syllables with short vowels in non-words of otherwise comparable syllabic structure. In other words, non-words with the syllabic structure CVVCVCC 2 had more initial stress than CVCVCC non-words, and non-words with the syllabic structure CVCVVC had more final stress than CVCVC non-words. A weak effect of number of coda consonants was also observed for verbs. Namely, non-words with the syllabic structure CVCVCC had marginally more final stress than CVCVC non-words. However, this effect was reversed for nouns. In addition, the factor of syllabic structure made a significant, independent contribution to a regression model predicting stress placement in both the production and perception results. Lexical Class. Another factor affecting stress placement is lexical class (nouns vs. verbs). Sereno (1986) and Kelly and Bock (1988) investigated the asymmetrical stress distribution for nouns and verbs by collecting the dictionary stress assignment of bisyllabic nouns and verbs listed in the Francis and Kučera (1982) word frequency norms. Considering only pure nouns and verbs, i.e., words that did not have homographs in other lexical classes, Sereno found that bisyllabic

10 10 pure nouns are stressed on the initial syllable 76% of the time, whereas bisyllabic pure verbs are only stressed on the initial syllable 34% of the time. Similarly, Kelly and Bock found that of the pure nouns, 94% were stressed on the initial syllable and of the pure verbs, only 31% had stress on the initial syllable. Thus, patterns associated with lexical class consist of one variable, noun or verb, and are associated with that variable in a simple, consistent way: Nouns are more likely to have initial stress and verbs are more likely to have final stress. Guion et al. (in press) found that native English speakers demonstrated knowledge of this distributional pattern in both production and perception: Non-words in a noun frame were more likely to be stressed on the initial syllable than non-words in a verb frame, echoing the distributional characteristics of these two lexical classes. In addition, the factor of lexical class made a significant, independent contribution to a regression model predicting stress placement in both the production and perception results. Phonological Similarity. It has been shown that the stress patterns of known words play a role in stress assignment on new words for native English speakers. Baker and Smith (1976) investigated the predictions of stress placement on non-words made by syllabic structure patterns as well as analogical assignment of stress. They found that stress assignment based on analogy with a similar form predicted stress in many cases, even when the prediction went against stress patterns based on syllabic structure. Studies of machine learning have also demonstrated that the stress patterns of the nearest phonological neighbors can accurately predict stress placement on novel words: Using exemplarbased models, Daelemans, Gillis, and Durieux (1994) and Gillis, Daelemans, and Durieux (2000) have demonstrated that a learning algorithm based on similarity of new items to stored exemplars exhibits the same behavior found in human language learners and users. Namely, an exemplarbased similarity algorithm can generalize stress patterns beyond the data on which it was trained and assign stress to new words in a manner similar to that found for language learners.

11 11 In regression analyses on both production and perception data, Guion et al. (in press) found that the location of stress on a phonologically similar real word could predict the placement of stress on non-words by native English speakers. Importantly, the stress pattern of the phonologically similar word predicted stress independently from the factors of syllabic structure and lexical class. This indicates that a similarity metric may operate along dimensions of phonological similarity that are independent of lexical class and syllabic structure. These dimensions of phonological similarity may be at the level of the individual exemplar or of the segmental pattern shared by the real and non-words. Spanish Stress Patterns A brief description of Spanish stress patterns is presented in this section to allow for an evaluation of the Spanish-English bilinguals' results. Specifically, the description of Spanish stress patterns is needed to determine possible transfer effects as well as to determine which English patterns would be novel to a Spanish speaker learning English. In Spanish, the normal and most common stress pattern is for stress to fall on the final syllable of a word if it ends in a consonant and on the penultimate syllable if the final syllable ends in a vowel. Notable exceptions include some inflectional endings (i.e., nominal plural /-s/ and verbal third person plural /-n/), which do not condition stress on the final syllable. Example words are presented in Table 2. There are other stress patterns that occur much less frequently. Harris (1992) reports on frequency counts he made from dictionaries, newspapers and novels and reports that the stress pattern outlined here is by far the most common. Taking precautions to include only words for which alternate stress patterns were possible, Harris found that from 85% to 95% of the words in the corpora followed the regular stress pattern. Table 2 approx. here

12 12 An examination of two databases revealed that most words in Spanish end in a vowel. In a lexical database, Faitelson-Weiser (1987, p. 178) reports that out of a corpus of 180,975 Spanish words, 131,557 (73%) ended in a vowel. An analysis of a database of Spanish syllable types developed by Álvarez, Carreiras, and de Vega (1992) found that the token frequency of final syllables ending in a vowel, in words of two syllables or longer, was roughly twice that of final syllables ending in a consonant (717,485 vs. 436,844). Thus, given the regular pattern of Spanish stress in which words ending in a vowel will have penultimate stress, most words of two syllables or more will have penultimate stress. If this distributional pattern of stress placement forms part of the Spanish-speakers' knowledge of their language (which seems likely given the evidence that English speakers have knowledge about distributional patterns in their language) the bilingual participants in this study may transfer this pattern and use it to determine stress assignment on the English non-words in the study. If this were the case, more initial than final stress placement would be expected. It is also noteworthy that all lexical classes show the same adherence to the normal stress pattern. In other words, it is not the case that nouns are more or less likely to deviate from the norm and use an alternate stress pattern than verbs are (Harris, 1992). 3 Thus, if evidence of knowledge of the statistical distribution of stress placement across the lexical classes of noun and verb in English were found for the Spanish-English bilinguals, it would indicate that learning of the English distribution had taken place and that it would not be a transfer of knowledge from Spanish. Experimental Hypotheses The investigation reported here was designed to determine whether early learners (who began to learn English before the age of seven) and late learners (who began to learn English after the age of 14) are able to learn English stress patterns equally. These cut-off ages were chosen based on Long's (1990) review which found that phonological acquisition was least

13 13 affected by delays in language exposure before the age of seven and age of acquisition effects were seemingly inevitable after the age of 12. Some stress patterns, such as those associated with syllabic structure, are quite complex and others, such as the distributional tendency for nouns to have initial stress and verbs to have final stress, are simpler. Of interest is whether the complexity of the pattern is predictive of second language learners ability to abstract and learn the given pattern and whether early and late learners show equal ability to acquire simple versus complex patterns of stress placement. Given the results of the Davis and Kelly (1997) study in which no age of acquisition effects on stress patterns based on lexical class were found, it is predicted that both the early and later learners in the current study will demonstrate knowledge of the relatively simple lexical class stress patterns. The mixed results from studies investigating the relatively more complex stress patterns related to syllabic structure suggest that these pattern may be less likely to be acquired, especially by late learners. Another question addressed by this study is whether learners of English will show analogical extension of stress patterns from individual lexical items to non-word experimental stimuli and whether such analogical extension is used by both early and late bilinguals. As associative learning processes, such as analogy, are hypothesized to be available to learners of all ages (see, e.g., Neville, 1999), both early and late learners are predicted to show analogical effects. Finally, two experiments, one assessing production and the other perception of stress patterns, were conducted. It is hypothesized that the perceptual experiment may reveal knowledge of stress patterns not found in the production data because of the nature of the tasks. The perception task requires only preference judgments, and thus places fewer demands on the processing system than the production task, which also requires formation and production of a non-word.

14 14 Experiment 1: Production In this experiment, three groups of participants were asked to produce two syllable nonwords in both noun and verb sentence frames. The effects of lexical class and syllabic structure of the four non-word stimulus types were investigated. Method Participants. Thirty adults were paid to participate in the experiment. None of the participants reported being diagnosed with any language or reading disorders and all passed a pure tone hearing screening in both ears from 500 to 4000 Hz at octave intervals (26 at 20 db and 4 at 25 db). The participants were recruited based on native language and language experience and divided into three groups (n=10): Native English, Early Spanish-English Bilingual, and Late Spanish-English Bilingual. All participants in the Native English group learned English as a native language and no other language was spoken in the home during childhood. Most had some history of foreign language study in high school and/or college but none had lived abroad for more than three months at a time. The bilingual participants all spoke Spanish as their first language and had lived in the United States for at least five years. The Early Bilinguals had begun learning English at the age of 2.5 to 6 years. The age of acquisition for the Late Bilinguals ranged from 15 to 33 years old and was determined by age of first massive exposure to English when they moved to the United States. Speakers in both bilingual groups used English regularly on a daily basis at the time of the study. On average, the Early Bilingual group used English 85%, while the Late Bilinguals used English 75% of the time (see Table 3). It was important to the goals of the study that the bilingual participants in the two groups were of roughly equal proficiency in English. Since the Early Bilinguals were likely to be highly proficient, the Late Bilinguals needed to be quite proficient as well. To this end, late bilinguals with long lengths of residence (12.9 years on average) and high levels of education (all had at least some college in the United States) were recruited. In addition, a standardized test of

15 15 English proficiency, the Test of Adolescent and Adult Language (TOAL), was administered to all participants. This test was chosen due to its ability to test high-proficiency English-speakers at adult levels (Hammill, Brown, Larsen, & Weiderholt, 1994). Two of the TOAL subtests, those focusing on listening vocabulary and listening grammar, were administered. The two bilingual groups performed comparably on the TOAL tests. Table 1 presents the results. There was no significant difference between the two bilingual groups for the TOAL Test 1 [F(1,18) = 1.99, p >.05] or Test 2 [F(1,18) = 0.34, p >.05]. Note also that scores for all three groups were well below the ceiling score of 35. Table 3 approx. here An additional proficiency measure was devised to test specifically the ability of the Spanish-English bilinguals to correctly place stress on real English words. Eighteen words with stress patterns consistent with regular English stress patterns and 18 words of the same syllabic structure, but which were not consistent with the regular patterns, were selected. The words in the two groups did not differ in their token frequency. Appendix I contains a list of these words and their frequency information. The words were presented orthographically (i.e., in English spelling) and the bilingual participants were asked to first rate the confidence of their knowledge of the word s meaning and of their ability to pronounce the word on 5-point scales. 4 Then, the participants were recorded reading each of the words in the sentence frame I said this time. The words were later coded for correct stress placement. Only words with a rating of 5 on both scales were considered in the analysis. Table 4 presents the average proportion of correct stress placement by group for each of the two word types. Note that both groups of bilinguals display high accuracy (96% or higher) of stress placement on known English words. The results from this proficiency test indicate that both groups of bilingual speakers are highly accurate in stress placement on known English words, whether or not they are consistent with the regular stress patterns.

16 16 Table 4 approx. here Materials. The stimuli consisted of 40 two-syllable non-words presented as isolated stressed syllables. There were four types of non-words, each with a different syllabic structure (see Table 5). Take, for example, a non-word of Type 1 (CVV.CVCC), [be t st]. The first syllable began with a single consonant [b], then had a long vowel [e ], while the second syllable began with a single consonant [t], then had a short vowel [ ], and ended in two consonants [st]. The syllabic structures of the non-words are all possible and frequent in English. However, Spanish does not allow two consonants at the end of a syllable as some of the nonwords have (see Types 1 and 2). In addition, non-words were produced in an American English accent, which is unlike Spanish in many ways. Most notably, there is a longer voice onset time (aspiration) for voiceless stops (i.e., [p t k]) in English than Spanish and the vowel qualities differ as well. Thus the non-words are more English- than Spanish-like. Table 5 approx. here The individual syllables making up the non-words were produced by a native American English speaker and recorded on digital audio tape (DAT) with a high quality microphone in the frame Now I say. Thus each syllable was produced with sentential pitch accent and stress, creating an equal degree of stress and clarity in production for all syllables. The productions were digitized at khz (16 bit) on a personal computer. The syllables were then excised from the frame sentence and normalized to 50% peak intensity. This created individual stimuli of the same intensity. The same speaker also recorded the phrases I d like to and I d like a. The final to and a were produced in a reduced form (i.e., [t ] and [ ]). These phrases were also digitized and normalized. The phrase I d like to will be referred to as the verb frame and the phrase I d like a will be referred to as the noun frame.

17 17 Procedure. The three groups of participants were asked to concatenate words of four syllabic structure types (see Table 5), presented as isolated, stressed syllables, into a single word and say it in a frame sentence. Each of the non-words was presented twice, once with the noun frame and once with the verb frame. Two pseudo-randomized, counterbalanced blocks were used, making a total of 80 trials. Each non-word was presented only once in each block. Half of the productions in each block were in the noun frame, half in the verb frame. There was a short distracter task between the two blocks. Participants were given some practice trials using nontest items before the first block. Stimulus presentation was controlled by the software SPARCS (Smith, 1997) on a personal computer and played over high quality loud speakers in a sound attenuated room. Participants adjusted the presentation volume to a comfortable level before testing began. The participants wore a head mounted-microphone and their responses were recorded on DAT tape. For each trial, the participants were first presented with a frame sentence. Presentation of the frame was both aural and visual. After a 500 ms delay the two stressed, isolated syllables composing the non-word were presented with a 500 ms inter-stimulus interval. Presentation of non-words was aural only; no visual (i.e., no orthographic) representation was given. Participants could replay the trial if they wished. After responding, they pushed a button to continue to the next trial. The participants were instructed to take the two syllables, keep them in the same order, and make a single word from them. They were instructed to produce all the sounds of the syllables they heard. Additionally, participants were asked to say the word in the carrier phrase that had just been presented. They were asked to say the carrier phrase in the same way they had heard it. If the final word in the frame ( to or a ) was produced with stress during the practice block, the participant was stopped and corrected. All participants could easily produce the carrier phrase in the desired manner. It was important that all the non-words were produced in

18 18 the same metrical frame because of potential confounding effects of the frame on stress placement. The participants were also told to take their time and consider all possible responses. Coding and Reliability. A native English speaker listened to the taped responses and coded them as having either first or second syllable stress. Some of the non-words were produced with segmental content different from that in the stimulus. If the substitution did not change the syllable structure (e.g., [ve b kt] for [be b kt]), the response was counted. However, if the substitution did change the syllable structure (e.g., [b b kt] for [be b kt]), the response was counted as missing. Approximately 1% of the responses were discarded because they did not reproduce the syllable pattern presented to the participant. Rater reliability was assessed by recoding a subset of the data. Ten responses were arbitrarily chosen (to equally sample all the non-words) from each of the 30 participants, for a total of 300 items. These responses were digitized and removed from their sentential context. The same native English speaker, as well as another native English speaker, blindly recoded these responses. The intraclass correlation coefficient (McGraw & Wong, 1996) for the two ratings done by the same rater was quite high (ρ =.96). This indicates that the rater coded stress placement with a high degree of reliability. The high correlation also indicates that bias was minimal as the second coding was done blindly. When the ratings of the two raters were submitted to a single measure intraclass correlation (which indicates the reliability when a single rater is used, as was the case here), the correlation coefficient was also high (ρ =.86), indicating that the use of one rater was highly reliable. Statistical Analysis. The scores from the production task were submitted to a mixed-design analysis of variance (ANOVA) with three factors: Group (3), Lexical Class (2), and Syllabic Structure (4). The three levels for the factor Group are Native English, Early Bilingual, and Late Bilingual. The two levels for the factor Lexical Class are word produced in a noun frame and word produced in a verb frame. The four levels for the factor Syllabic Structure are defined as

19 19 the four non-word types in Table 5. For the subject (F1) and item (F2) analyses, repeated measures were used on Lexical Class, as the same non-word was produced both in a noun frame and in a verb frame. For the subject analysis, repeated measures were used on Syllabic Structure as well, in order to compare the productions by a given participant across the four non-word types. Results Figure 1 presents the mean proportion of initial syllable stress for the four non-word types, two sentence frames and three groups. Note that non-words produced in a noun frame were more often produced with initial syllable stress than those in a verb frame. Also a long vowel in the initial syllable (Type 1) conditioned more initial syllable stress, whereas a long vowel in the final syllable (Type 4) conditioned more final stress. In addition, the size of the effects is smaller for the Late Bilingual Group and overall this group has more initial syllable stress productions. Because some of the scores were greater than.90 or less than.10, the proportions were arcsine transformed, producing scores that were normally distributed and had a constant variance (see, e.g., Woods, Fletcher, & Hughes, 1986, p. 220). The transformed scores were then submitted to a mixed-design ANOVA with three factors: Group (3), Syllabic Structure (4), and Lexical class (2). The main effects of Group [F1(2,27) = 5.42, p <.05; F2(2,108) = 32.38, p <.05], Syllabic Structure [F1(3,81) = 50.63, p <.05; F2(3,108) = 40.84, p <.05] and Lexical Class [F1(1,27) = 50.31, p <.05; F2(1,108) = , p <.05] were all significant, as was the threeway interaction [F1(6,81) = 2.46, p <.05; F2(6,108) = 2.59, p <.05]. (Note that all 2-way interactions were also significant.) This indicates that the effects of Syllabic Structure and Lexical Class (and their interaction) affected stress placement differently across the three groups. Figure 1 approx. here

20 20 To further investigate these effects, separate one-way ANOVAs investigating the effect of Lexical Class for each of the four types, for each of the three groups were performed. Then separate one-way ANOVAs, and pair-wise comparisons, investigating the effect of Syllabic Structure for each lexical class for each of the three groups were performed. The effect of Lexical Class was significant (p <.05) for each of the four syllabic structure types for both the Native English group (F-values ranged from 13.8 to 42.39) and the Early Bilingual group (F-values ranged from 5.68 to 73.13). In the case of the Late Bilingual group, while the trend in the expected direction was found for all four types, the effect of Lexical class only reached significance for Type 3 in the item [F2(1,9) = 5.45, p <.05] but not the subject [F2(1,9) = 3.46, p >.05] analysis. The effect of Syllabic Structure was significant (p <.05) for all groups for words produced in both the noun and verb sentence frames (F-values ranged from 9.83 to 17.97). However, the pair-wise comparisons revealed differences between the groups. In the case of the words produced in the noun frame, both the Native English and Early Bilingual groups were found to have more initial syllable stress responses on Type 1 than the other three types (Tukey's p <.05). The Late Bilingual group, on the other hand, showed a different pattern. Initial syllable stress responses for Type 4 were less than the other three types (Tukey's p <.05). In the case of the words produced in the verb frame, both the Native English and Early Bilingual groups were found to have more initial syllable stress responses for Types 1 and 3 than Types 2 and 4 (Tukey's p <.05). On the other hand, the Late Bilingual group had fewer initial stress responses for Type 4 than the other three types (Tukey's p <.05). Discussion The results from the native English speakers largely replicate the findings from Guion et al. (in press) in which 17 native English speakers were tested on the same protocol. As in the previous study, non-words produced in a noun frame were found to have more initial syllable

21 21 stress than those produced in a verb frame. The syllabic structure effects due to vowel length were also replicated. Namely, for both noun frame and verb frame productions, more initial stress was found for CVVCVCC non-words (Type 1) than CVCVCC (Type 2) non-words. In addition, the verb frame productions showed more final syllable stress for CVCVVC non-words (Type 4) than CVCVC (Type 3) non-words. The previous study found this last effect for nouns as well. The marginal coda consonant effect found for verbs in the previous study was more robust here. Namely, the verb frame productions had more final syllable stress for CVCVCC non-words (Type 2) than CVCVC (Type 3) non-words. The replication of these effects provides strong evidence that native English speakers use lexical class and syllabic structure (i.e., vowel length and number of coda consonants) to determine stress placement during the production of non-words. The early Spanish-English bilinguals showed the same effects of lexical class and syllabic structure on stress placement in the non-words that the native English speakers did. This indicates that the early bilinguals used lexical class, vowel length and number of coda consonants in a way similar to the native English speakers when assigning stress to non-words in a production task. The late Spanish-English bilinguals, on the other hand, showed differences from the other two groups. First, the effect of lexical class was only found for CVCVC non-words (Type 3). This effect was found for all four types for the other two groups. Second, the effects of syllabic structure for the pair-wise comparisons were different. Namely, the initial long vowel effect (more initial stress for CVVCVCC than CVCVCC) was not observed, nor was the coda consonant effect (more final stress for CVCVCC than CVCVC). The final long vowel effect was, however, observed in both the noun and verb sentence frames. CVCVVC non-words (Type 4) had more final stress than any of the other three non-word types. Thus, the late Spanish- English bilinguals showed weaker and somewhat different effects of both lexical class and syllabic structure than the other groups.

22 22 Another difference between the late bilinguals and the other two groups was in the greater number of overall initial stress productions for this group. We suggest that this may be a transfer effect from Spanish and may reflect the overall preponderance of penultimate stress found across the Spanish lexicon (as discussed in the introduction). Experiment 2: Perception In this experiment, the same three groups from Experiment 1 participated. They were asked to report their preference for initial or final syllable stress on two syllable non-words (the same ones used in the previous experiment) that were aurally presented in noun and verb sentence frames. The effects of lexical class and syllabic structure on the four non-word stimulus types were investigated. Method Participants. The same 30 people, divided into the three groups of Native English, Early Spanish-English Bilingual, and Late Spanish-English Bilingual, participated. Materials. The same 40 non-words used in Experiment 1, listed in Table 5, were produced with stress on the initial and final syllable in each of the carrier frames I d like a and I d like to making a total of 160 sentences. The a or to of each phrase was produced in a reduced manner. The same speaker used in Experiment 1 produced these sentences. The productions were recorded on DAT tape. Then they were digitized at khz (16 bit) on a personal computer. Each phrase was edited into its own file. These files were then normalized to 50% peak intensity. Multiple productions of each sentence were produced and those considered to be the most similar in rate and speaking style by another native English speaker were selected for use as stimuli.

23 23 Procedure. All participants were run on Experiment 2 after finishing Experiment 1. Since the same words were used in both experiments, it is possible that practice effects may have influenced the results of Experiment 2. However, since the experiments were presented in the same order to all participants, it is likely that any such effects would have affected all participants equally. Participants were asked to listen to the prerecorded phrases in pairs that varied only in the stress placement on the non-word. In a given trial, the same sentence frame (noun or verb) was presented. They were instructed to listen to the two sentences and indicate which one sounded the most like a real English sentence to them. For example, consider the two representative trials below. Example Trial 1: [a d la k be bekt] (pause) [a d la k be bekt] Example Trial 2: [a d la k t b l n] [a d la k t b l n] (pause) Each target non-word was presented in two trials, once in a noun frame and once in a verb frame. Two pseudo-randomized, counterbalanced blocks were used, making a total of 80 trials. Each non-word was presented only once in each block. Half of the productions in each block were in a noun frame, half in a verb frame. The order of initial or final syllable stress within a trial was also controlled. For a given non-word, the order of stress presentation remained constant for both the noun frame and verb frame trials. For each of the four non-word types, five of the non-words were presented with initial stress, then final stress. Conversely, the other five were presented with final stress, then initial stress. There was a break of a few minutes between the two blocks. Participants were given some practice trials using non-test items before the first block.

24 24 Stimulus presentation and response recording was controlled by the software SPARCS (Smith, 1997) on a personal computer and played over high quality headphones in a sound attenuated room. Participants adjusted the presentation volume to a comfortable level before testing began. For each trial, the participants were presented two sentences with a 1500 ms interstimulus interval. They responded by indicating which sentence they preferred by pushing a button labeled first or second. This brought up the next trial after a 2000 ms delay. No repetition of the trial was allowed. Statistical Analysis. The same statistical design used for the production data was used for the perception data. Briefly, the scores from the perception task were submitted to a mixed-design ANOVA with three factors: Group (3), Syllabic Structure (4), and Lexical Class (2) with repeated measures on Lexical Class for the subject (F1) and item (F2) analyses, and repeated measures on Syllabic Structure as well for the subject analysis. Results The mean proportions of initial stress responses for each group are presented in Figure 2. Note that, as was the case with the production data, nouns tend to have more initial stress responses than verbs. The effect is largest for the Late Bilinguals. In addition, note that the pattern of initial stress preferences across the non-word syllabic structure types is quite similar for the Native English and Early Bilingual groups. The Late Bilinguals, on the other hand, show a different pattern. Also, the Late Bilinguals have more initial stress responses overall. The proportions of initial stress were arcsine transformed (because some of the scores were above.90 and some below.10) and then submitted to a mixed design ANOVA with three factors: Group (3), Syllabic Structure (4), and Lexical Class (2). Figure 2 approx. here

25 25 The main effects of Group [F1(2,27) = 4.53, p <.05; F2(2,108) = 37.97, p <.05], Syllabic Structure [F1(3,81) = 13.05, p <.05; F2(3,108) = 11.88, p <.05] and Lexical Class [F1(1,27) = 17.46, p <.05; F2(1,108) = 43.48, p <.05] were all significant. The two-way interactions of Group and Lexical class [F1(2,27) = 1.01, p >.05; F2(2,108) = 4.14, p <.05] and Group and Syllabic Structure [F1(6,81) = 1.96, p >.05; F2(6,108) = 3.2, p <.05] were significant for the item analysis but not the subject analysis. No other two-way interactions were significant. The three-way interaction was significant for the item analysis [F2(6,108) = 2.58, p <.05] but was not significant for the subject analysis [F1(6,81) = 1.42, p >.05]. As the three-way interaction was not robust, only the two-way interactions (found for the item analyses) of Group with Lexical Class and Group with Syllabic Structure were explored further. Separate ANOVAs testing the effect of Lexical Class for each group across the combined syllable structure types were performed. Then separate ANOVAs and pair-wise comparisons testing the effect of Syllabic Structure for each group on the combined noun and verb frame responses were performed. The effect of Lexical Class was found to be significant for all three groups: Native English [F2(1,39) = 9.87, p <.05], Early Bilingual [F2(1,39) = 7.59, p <.05], and Late Bilingual [F2(1,39) = 26.39, p <.05]. The source of the interaction seems to lie in the magnitude of the effect across the groups. Both the Native English and Early Bilinguals had about the same proportion greater first syllable responses in the noun than the verb frame (.11 greater for the Native English and.10 greater for the Early Bilinguals). The Late Bilinguals, on the other hand, showed a larger effect (.24 greater for the noun than the verb frame). The effect of Syllabic Structure was significant for all three groups: Native English [F2(3,36) = 8.33, p <.05], Early Bilingual [F2(3,36) = 4.72, p <.05], and Late Bilingual [F2(1,39) = 3.78, p <.05]. However, the pair-wise comparisons revealed differences between the groups. In the case of the Native English group, Tukey s comparisons (p <.05) on the combined noun and verb frame data showed more initial syllable stress preferences for Type 1 than Types 2 and 4 and more for Type 3 than Type 4. The Early Bilingual group showed more