Prosodic Disambiguation in First and Second Language Production: English and Korean

The prosodic system of English consists of two parameters, as summarized in (3). English marks word prominence with pitch accents, or well-defined pitch shapes, whose metrically strong (starred) tone is associated with words or syllables to be accented. For instance, an H* accent is typically associated with new information to be predicated in the discourse, while an L* accent in general marks information that is intended to be salient but not to be predicated by the speaker (for more information on the pitch accent inventory of English, see Pierrehumbert & Hirschberg, 1990). Pitch accents are phonetically realized as pitch movements and increased intensity. In addition, English marks boundaries of prosodic units, such as intermediate phrases (ip) or intonational phrases (IP), with phrase accents, boundary tones, final lengthening and pauses. An ip boundary is marked by a phrase accent, represented by H- or L-. An H- phrase accent indicates that the current phrase combines with the following phrase to form a larger phonological unit, and an L- phrase accent signals the separation of the current phrase from a subsequent phrase (Pierrehumbert & Hirschberg, 1990). IP boundary tones are labelled with the % symbol, as in H% and L%. An L% boundary tone generally marks a statement or a wh question, while an H% boundary tone typically marks yes-no questions (Pierrehumbert & Hirschberg, 1990).

In the example sentence in (4), the words cookie and kill have an H* pitch accent, marking new information to be predicated. The H- phrase accent on cookie marks ip boundaries, indicating the continuation of the larger phonological unit, IP. The sentence-final L% boundary tone marks an IP boundary.

        H* H-    H* L-L%  (Pierrehumbert & Hirschberg, 1990)

Word prominence and prosodic boundaries provide disambiguating cues when native English speakers comprehend spoken sentences with RC attachment ambiguity. Schafer et al. (1996) and E.-K. Lee and Watson (2011) showed that English speakers listening to sentences with RC attachment ambiguity were more likely to interpret a noun as the head of the RC when it carried a pitch accent (H* or L+H*) compared to when the other noun received the accent. Moreover, Jun and Bishop (2015b) also found that listeners were biased toward the more prominent noun as the head of the RC and that this effect of word prominence was persistent even when the boundary cues were directed toward the other syntactic phrasing. Clifton et al. (2002) created auditory stimuli varying in the relative strength of the two prosodic boundaries as in I met the daughter [1] of the colonel [2] who was on the balcony, based on which listeners responded to a forced-choice task indicating their interpretations of the sentences. The participants preferred the high attachment reading when the second boundary was stronger (no prosodic boundary at [1] and an ip boundary at [2]) than when the first boundary was stronger (an IP boundary at [1] and an ip boundary at [2]).

English speakers’ use of prosody for disambiguation has also been found in production studies. For example, Kraljic and Brennan (2005) investigated English speakers’ production of sentences with prepositional phrase (PP) attachment ambiguity, as in Put the dog [in the basket]_PP on the star, in which the PP in the basket can specify either the dog (modifier interpretation) or the location at which the dog is to be placed (goal interpretation). The results of their production experiment indicated that for modifier interpretations, the duration of the basket and the subsequent pause was longer than the duration of the dog and its subsequent pause by 17.3% of the utterance length. The opposite pattern was found for goal interpretations; the duration of the first NP and its subsequent pause was longer than that of the second NP and its subsequent pause by 10.5% of the utterance length. These results suggest that speakers resolve attachment ambiguity by placing a prosodic juncture (i.e., lengthened phrase-final word and its subsequent pause) at the location of boundaries of bigger syntactic phrases. A similar use of prosodic boundaries was found in the resolution of complex NP ambiguity ([television] [or radio and newspapers] vs. [television or radio] [and newspapers]) and closure ambiguity (Because her grandmother knitted pullovers / Kathy kept warm in the wintertime vs. Because her grandmother knitted / pullovers kept Kathy warm in the wintertime) (Allbritton et al., 1996; Price et al., 1991; Speer et al., 1996). However, few studies have investigated the use of word prominence cues in the production of ambiguous sentences by native English speakers.

In contrast to English, Korean does not have an accent system associated with lexical prominence (5). Instead, it marks boundaries of prosodic units, such as accentual phrases (AP) and IP. Similarly to those in English, these prosodic boundaries in Korean are marked by boundary tones, along with final lengthening and pauses. The final syllable of an AP typically carries a Ha boundary tone, and a La boundary tone occurs when it is flanked by two high tones at the end of a short AP (Jun, 2005a). The final syllable of an IP is marked by one of the nine boundary tones, L%, H%, LH%, HL%, LHL%, HLH%, LHLH%, HLHL%, LHLHL%, which have different semantic and pragmatic functions. Moreover, an IP-final syllable is about twice as long as the same syllable in the middle of an IP (Jun, 2005a).

For example, in (6), each word forms an AP, as represented by the Ha boundary tones, and the sentence-final L% boundary tone marks an IP boundary.

Korean speakers use prosodic boundaries to resolve RC attachment ambiguity, both in processing and production. Jun (2003) invited speakers of typologically diverse languages to produce a sentence with RC attachment ambiguity in their native language equivalent to Someone shot the servant of the actress who was on the balcony (NP1+NP2+RC for head-initial languages or RC+NP1+NP2 for head-final languages) with a pause or a phrase break before or after either NP2. The speakers were then asked to answer the question Who was on the balcony? for each production. The Korean speakers, as well as speakers of other head-final languages, preferred low attachment readings when there was a prosodic boundary between NP1 and NP2 and high attachment readings when there was a boundary between the RC and the adjacent NP. Although this was not a traditional processing task, in the sense that the informants were making judgments on their own guided productions, the results suggest that their judgments were influenced by prosodic boundary cues.

Moreover, Korean speakers use prosodic boundaries as a cue to disambiguation even when they are not explicitly directed to do so. In Baek and Yun (2018), the participants were instructed to read aloud sentences such as (7) to deliver a given meaning. Acoustic analysis of their productions showed a stronger prosodic juncture, realized as boundary tones, lengthening, pause, or pitch reset (a domain-initial reset of pitch to a high value after a continuous declination in the previous domain), after the RC when a high attachment reading was elicited, and the same cues were produced after NP1 when a low attachment reading was elicited. The authors also found that the strength of a prosodic juncture correlated with the number of syntactic phrase boundaries occurring in the same position, which suggests that a mapping between syntactic phrase boundaries and prosodic phrase boundaries helps ambiguity resolution in the case of RC attachment ambiguity in Korean.

Cross-linguistic differences in prosodic disambiguation might reasonably be expected to influence L2 learning. Yang (2010) studied the use of prosody for disambiguation by Taiwanese-English L2 speakers at two proficiency levels—advanced and limited. In a read-aloud task, the speakers produced ambiguous English sentences with coordinate structures, such as The little (a) dogs (b) and cats chased a ball, where the prosodic boundaries at (a) and (b) were expected to help resolve the ambiguity. This study found that the advanced L2 speakers produced strong boundary cues (pre-boundary lengthening and pause) at (a) when the adjective modified both of the following nouns, but these cues appeared at (b) when the adjective modified only the immediately following noun. This was the same pattern that the control group of native English speakers showed. In contrast, speakers with limited L2 proficiency produced both boundaries with a similar duration regardless of the intended interpretations and thus did not exhibit significant use of prosody for disambiguation.

Jackson and O’Brien (2011) elicited German sentences from English-German L2 speakers to examine their use of prosodic cues for the resolution of temporary PP attachment ambiguity (e.g., The manager thanked the secretary [with two children]_PP inducing noun attachment and . . . [with a bouquet of flowers]_PP inducing verb attachment). Their results showed that the L2 speakers used different prosodic cues for the different attachments, such as pre-boundary lengthening, pauses, and a phrase-initial pitch accent. Moreover, although the L2 speakers varied in their L2 German proficiency test scores, there was no significant relationship between their German proficiency and the extent to which they made use of these prosodic cues for disambiguation.

While Yang (2010) and Jackson and O’Brien (2011) made important contributions to our understanding of prosodic disambiguation by L2 speakers, their studies shared theoretical and analytical limitations. First, both did not compare the prosodic systems of the participants’ L1 and L2, which is necessary to create a testable hypothesis on the effects of L1 on L2 speech production. Second, both studies focused only on prosodic phrasing phenomena in disambiguation. When it comes to word prominence, another significant parameter in prosodic typology (Jun, 2014; Wagner & Watson, 2010), there has been a dearth of studies on how L2 speakers utilize this cue in producing ambiguous sentences in the target language.

There has been much research on the use of a single type of prosodic cue—either word prominence or prosodic boundaries—in ambiguity resolution. However, languages differ in terms of what parameters they have available in their prosodic systems, which should then influence what prosodic parameters are used for disambiguation. The first goal of the present study was to conduct cross-linguistic research comparing prosodic disambiguation in prosodically different languages, namely, English and Korean. On the assumption that languages make use of all prosodic parameters that their prosodic system makes available, English is expected to use both word prominence and prosodic boundary cues for disambiguation. That is, native English speakers’ productions of RC attachment ambiguity structure, NP1+NP2+RC, would differ with regard to both (a) the relative prominence of NP1 and NP2 and (b) the relative strength of prosodic boundaries after NP1 and NP2, depending on what readings they intend to deliver. On the contrary, Korean would only use prosodic boundaries since it lacks a prosodic parameter marking word prominence. Thus, when native Korean speakers produce RC+NP1+NP2 structures with different intended readings, their productions would differ in the relative strength of prosodic boundaries after RC and NP1 but not in the prominence of NP1 and NP2. To test these hypotheses, this study compared the prosodic properties of the two language groups’ productions of ambiguous RC structures.

Furthermore, this study also aimed to examine the impact of these cross-linguistic prosodic differences on L2 productions. The prosodic systems of English and Korean are relatively similar in the parameter of prosodic boundaries but differ with regard to the parameter of prosodic prominence. Thus, when Korean-English bilingual speakers produce sentences with RC attachment ambiguity in L2 English, they would be expected to use the relative strength of prosodic boundaries after NP1 and NP2 as native speakers do, but to fail to manipulate the prominence of NP1 and NP2 in a target-like manner. In order to test these hypotheses, Korean-English bilingual speakers were invited to produce RC attachment ambiguity structures in English, and their productions were compared to native English speakers’ productions.

To elicit the production of ambiguous sentences in English and Korean, 12 English sentences and 12 Korean sentences with RC attachment ambiguity were used as stimuli. The English sentences were adopted from previous studies (Jun & Bishop, 2015a, 2015b), and the Korean sentences were created by the author. To ensure that the sentences had the desired two readings, three native speakers of each language indicated their preferences for each sentence on a 5-point scale (e.g., question: Who was dishonest?; choices: must be the boss/more likely the boss/equally likely/more likely the clerk/must be the clerk). None of the sentences received more than one response at either end of the scale, confirming that they could be interpreted as ambiguous. Taking into account the raters’ comments, slight lexical changes were made to ensure the sentences were less biased towards either of the readings as well as easier to pronounce. The final list of stimuli is given in Appendix A.

Each target sentence was presented on a computer screen along with two short sentences representing two different interpretations, as shown in Figure 1. The task was to read the target sentence aloud twice, for one reading after the other, as indicated by the highlighted sentences. The order between the two attachment readings were counterbalanced across participants and across items. This choice of methodology was motivated by the results of a previous study (Baek & Yun, 2018), which showed that prosodic cues for disambiguation were saliently produced only when participants were explicitly attempting to differentiate competing interpretations in the absence of a disambiguating context. Similarly, Straub (1996) reported that the salience of disambiguating prosodic cues diminished when alternate sources of disambiguation information, such as biasing contexts, were present for the listeners.

Experimental sessions took place in a sound-treated recording room. All elicited utterances were recorded using a Zoom H6 digital recorder and a SM10A-CN dynamic head-mounted microphone at a sample rate of 44.1 kHz. At the end of the session, the participants completed a brief questionnaire on their demographic information and language background.

Eleven English speakers (8 females, 3 males) participated in the experiment. Their mean age was 20.2, ranging from 18 to 24, and they reported that they had grown up in the New York, Pennsylvania, or Delaware regions. Twelve Korean-English bilingual speakers (6 females, 6 males) also participated. All of them indicated Korean as their native language and English as their L2. Their age and language background information from the questionnaire is summarized in Table 1.

All participants were attending Stony Brook University. The native English speakers completed the task in English. The Korean-English bilingual speakers completed the task first in L1 Korean and then in L2 English.

During pre-processing of the recorded data, one female English speaker was excluded due to creaky voice quality, which interfered with accurate pitch measurement. Two female Korean speakers were excluded because they spoke the Kyeongsang dialect of Korean, which is prosodically different from Seoul Korean. The remaining data consisted of three language data sets (L1 English, Korean-English bilingual speakers’ L1 Korean and L2 English), each consisting of 240 utterances (12 sentences × 2 attachments × 10 speakers). Thirty-two utterances (8 L1 English tokens, 8 L1 Korean tokens, and 16 L2 English tokens) were excluded due to production errors, such as repetition of a word or production of an incorrect word.

The resulting 688 utterances were word-segmented by hand in Praat, version 6.1.07 (Boersma & Weenink, 2019). In each of the recorded English sentences, three intervals were segmented—NP1 (e.g., boss), NP2 (e.g., clerk), and RC (e.g., who was dishonest)—as well as pauses between them, if present. Similarly, in each of the recorded Korean sentences, three intervals were segmented: the predicate in the RC, NP1, and NP2 (e.g., pucilenhan, sonyeuy, and ennika), as well as pauses between them. After segmentation, the following acoustic measurements were extracted: (a) the duration of the intervals (ms), (b) the duration of pauses (ms), (c) the maximum and mean pitch of the intervals (Hz), and (d) the maximum and mean intensity of the intervals (dB). To eliminate the possible influence of speech rate, the duration measures were used to calculate the duration proportions of each interval or pause to the entire sentence. Pitch values in Hz were converted into semitones to control for individual variation ($12lo{g}_{2}F0/F_{ref}$, reference level = 1Hz). The use of mean intensity measurements was motivated by previous studies reporting a relationship between word mean intensity and word prominence, such as focus marking, across typologically diverse languages including English and Korean (A. Lee & Xu, 2012; Y. Lee et al., 2015; Y.-c. Lee & Xu, 2010; Wu & Xu, 2010).

Differences in measurements across attachment conditions (Low and High) and across word positions (NP1, NP2, and RC) were statistically analyzed by mixed-effects regression models using the lmer() function from the lme4 package (Bates et al., 2015) in R, version 3.6.1 (R Core Team, 2019). First, the L1 English data and Korean-English bilinguals’ L2 English data were compared for each dependent variable (word duration, pause duration, maximum and mean pitch, and maximum and mean intensity) by fitting a model with language group (English, Korean), word position (NP1, NP2, RC), and attachment (low attachment, high attachment) as fixed effects and a by-participant random intercept. Since the results indicated significant three-way interactions, as reported in the section reporting the results of the Korean-English bilingual speakers, each dataset (L1 English, Korean-English bilingual speakers’ L1 Korean and L2 English) was then analyzed separately to understand the prosodic cue use by each language group.

For each dataset, models were fit for each dependent variable with word position and attachment as fixed effects. The predictors were treatment-coded. The reference level for word position was the first interval for word and pause duration analyses (NP1 for English and RC for Korean) and the first NP for pitch and intensity analyses (NP1 for both English and Korean). The reference level for attachment was low attachment for all analyses. For the random effect structure, a full model was fit with by-participant and by-item random intercepts as well as by-participant and by-item random slopes for both fixed effects. When a model failed to converge, the random effect that captured the smallest variance was removed until the model fit reached convergence (Barr et al., 2013). The final models are reported in Appendix B. The p-values were obtained using the lmerTest package (Kuznetsova et al., 2017).

This section reports the results of the four acoustic measurements—word duration, pause duration, mean pitch, and mean intensity—extracted from the L1 English data. Figure 2 shows the means and standard errors of the four acoustic measurements for each interval by attachment condition. Table 3 summarizes the results of the linear mixed-effects analyses for each acoustic measure.

This section reports the results of the acoustic measurements extracted from the Korean-English bilingual speakers’ L1 Korean data. The following subsections each report the results for each prosodic measure: word duration, pause duration, mean pitch, and mean intensity. Figure 3 shows the means and standard errors of the four acoustic measurements for each interval by attachment condition. Table 4 summarizes the results of the linear mixed-effects analyses for each acoustic measure.

First, the L1 English data and the Korean-English bilingual speakers’ L2 English data were compared by running a model with language group, position, and attachment as fixed effects (reference level: English L1, NP1, low attachment) for each acoustic measurement. The results indicated significant three-way interaction effects for pause duration ($t=-3.8,p<.001$), mean pitch ($t=3.7,p<.001$), and mean intensity ($t=4.2,p<.001$). To better understand how these speakers used these prosodic cues, separate analyses were conducted on the Korean-English bilingual speakers’ L2 English data. This section reports the results for each prosodic measure: word duration, pause duration, mean pitch, and mean intensity. Figure 4 shows the means and standard errors of the four acoustic measurements for each interval by attachment condition. Table 5 summarizes the results of the linear mixed-effects analyses for each acoustic measure.

The first goal of this study was to compare English speakers and Korean speakers in terms of their use of prosodic cues to resolve ambiguity in their native languages. First, the English speakers used a combination of boundary cues and word prominence cues to disambiguate RC attachment. They inserted pauses between words to mark prosodic boundaries at different positions, depending on the intended attachment. When the RC was attached to the noun that was lower in the syntactic structure (NP2), a pause was inserted before NP2, marking a stronger prosodic juncture between (NP1) and (NP2 RC). When the RC was attached to the higher noun (NP1), a pause was inserted before RC, indicating a stronger prosodic juncture between (NP1 NP2) and (RC). Previous studies have reported similar use of prosodic boundaries by English speakers in the resolution of other types of syntactic ambiguities (Allbritton et al., 1996; Kraljic & Brennan, 2005; Price et al., 1991; Schafer et al., 2005; Speer et al., 1996). The English speakers also manipulated relative prominence of words for RC attachment ambiguity resolution. Nouns were produced with greater prominence (higher pitch and greater intensity) when they were heads of and modified by the RC than when they were not. The role of word prominence in English listeners’ comprehension of ambiguous structure has been observed in earlier processing studies (Jun & Bishop, 2015b; E.-K. Lee & Watson, 2011; Schafer et al., 1996). Adding to such studies on the effects of word prominence in processing, the current study found that English speakers manipulate word prominence for disambiguation in their utterances, suggesting that the effect of word prominence is present in production as well as in processing.

However, the English speakers did not show significant use of word duration for disambiguation. This result is likely due to the fact that lengthening can be used for both prosodic boundary marking and head prominence (Wagner & Watson, 2010). In the low attachment reading, (NP1) (NP2 RC), NP1 could be lengthened as a cue for the following prosodic boundary, while NP2 could also be lengthened for prominence as the RC head. Similarly, in the high attachment reading, (NP1 NP2) (RC), NP1 could be lengthened for head prominence, while NP2 could also be lengthened to mark the following prosodic boundary. There has been disagreement on the scope of phrase-final lengthening in English, with the locus of lengthening variously identified as segment, rhyme, syllable, and word (e.g., Cho et al., 2013; Klatt, 1975; Price et al., 1991; Turk & Shattuck-Hufnagel, 2007; Wightman et al., 1992). It is thus possible that further analysis of the data might show an effect of pre-boundary lengthening in domains smaller than the word.

In contrast to English, Korean relies mainly on boundary phenomena, including pre-boundary word lengthening and pauses, for disambiguation. When the RC modified the lower noun (NP1), NP1 was lengthened and followed by a longer pause, marking a stronger prosodic juncture between the two nouns, as in (RC NP1)(NP2). In contrast, when the RC was attached to the higher noun (NP2), RC was lengthened and followed by a longer pause, indicating a strong prosodic juncture between (RC) and (NP1 NP2). This finding conforms to the previous studies by Jun (2005b), who found pauses at phrase boundaries optionally accompanied by pre-boundary lengthening, as well as by Baek and Yun (2018), who found an effect of the number of syntactic phrase boundaries on the strength of prosodic juncture, reflected in pre-boundary lengthening and pauses. In the same vein, prosodic boundaries were also reported to affect syntactic ambiguity interpretation by Korean listeners (Jun, 2003; Kang & Speer, 2003).

In addition to the boundary cues, the Korean speakers manipulated the intensity of NP2 according to the intended reading. The mean intensity of NP2 was greater in the high attachment productions than in the low attachment productions. As NP2 is the head of the RC in the high attachment reading, it is likely that the greater intensity was used to mark head prominence. If this is true, it would suggest that even edge-marking languages, such as Korean, may make use of prosodic prominence cues to distinguish meaning. The use of prosodic prominence in Korean has also been reported in the domain of lexical ambiguity resolution. In Korean, an interrogative sentence containing a wh-phrase (e.g., nwuka ‘who’ or ‘anyone’) is ambiguous, as it can be interpreted as either a yes-no question, a wh question, or an incredulity question. Jun and Oh (1996) found that native speakers’ productions with different intended interpretations differed not only in prosodic phrasing but also in the pitch range and intensity peak on the wh-phrase, although these prominence cues were not used as consistently across speakers as prosodic phrasing, and prosodic phrasing cues had a stronger effect on the perception of the sentences than prominence cues. Similarly, since the effect size of the intensity use in this study was relatively small compared to that in the L1 English speakers’ productions ($t=2.157,p=0.031$ for L1 Korean vs. $t=-12.022,p<0.001$ for L1 English), further investigation would be needed before reaching any conclusion on whether Korean has a prosodic system of word prominence using intensity. Another possibility lies in that the Korean population used in this study are Korean-English bilingual speakers immersed in an English-speaking environment. As they are exposed to English on a daily basis, it is possible that their use of intensity for disambiguation is an effect of English prosody on their L1 Korean production. If this is true, it remains to be answered why only intensity and not pitch was the subject of L2-to-L1 transfer. I leave this to future studies.

The second goal of this study was to investigate the use of prosodic cues for disambiguation in English by Korean-English bilingual speakers. Given the similarities and differences between the prosodic systems of English and Korean, it was hypothesized that Korean-English bilingual speakers would use word duration and pause duration as boundary cues in the same way as native English speakers do and that their use of pitch and intensity as prominence cues would be different from that of native English speakers. However, the results of this study did not support these hypotheses.

The Korean-English bilingual speakers increased the duration, pitch, and intensity of a noun that was the head of the RC and decreased these measures for the other noun. This demonstrates that these speakers resolved ambiguity by manipulating relative word prominence (i.e., by placing pitch accents on the noun that heads the RC and by de-accenting the other noun), suggesting that they have successfully learned the inventory of the phonological categories of pitch accents in the English intonation system. However, the degrees to which they scaled these cues, represented by the interaction effect sizes in the statistical results, were considerably weaker compared to those in the English L1 speakers’ production. In the Korean-English bilingual speakers’ productions, the effect sizes of the interaction between word position (NP1 vs. NP2) and attachment condition (low vs. high) were $t=-2.322$ ($p=0.021$) for mean pitch and $t=-2.561$ ($p=0.011$) for mean intensity. In contrast, in the English L1 speakers’ productions, the effect sizes were $t=-7.302$ ($p<0.001$) for mean pitch and $t=-12.022$ ($p<0.001$) for mean intensity. In other words, although the Korean-English bilingual speakers have learned to use the phonological category of pitch accents, their scaling of acoustic cues to realize this category is still far from the native target. This finding is in line with the L2 Intonation Learning Theory (LILt), which postulates that different dimensions of L2 intonation pose different degrees of learning difficulty for L2 learners (Mennen, 2015). The LILt defines four dimensions of L2 intonation: the phonological dimension, the phonetic dimension, the semantic dimension, and the frequency dimension. These dimensions are expected to constitute different levels of difficulty in L2 learning depending on whether two languages are similar or different in each dimension. On the relative difficulty of learning of the phonological and phonetic dimensions, L2 speakers in earlier studies showed more difficulty in the phonetic dimension than in the phonological dimension, the same pattern found in the current study (Backman, 1979; O’Brien & Gut, 2010; Trofimovich & Baker, 2006).

Moreover, the Korean-English bilingual speakers did not show significant use of boundary marking cues for disambiguation, such as pre-boundary lengthening and pauses. Instead of using word duration as a cue for boundary marking as they do in their native language, they used duration as a cue for head prominence by lengthening a noun when it was the head of the RC (i.e., NP1 in the high attachment reading, NP2 in the low attachment reading). In addition, Korean-English bilingual speakers did not resolve ambiguity using pauses. Regardless of the attachment condition, they consistently inserted a longer pause after NP2, demarcating the prosodic phrases (NP1 NP2) and (RC). While it has been found that speakers in general tend to pause more often in their L2 than in their L1 (Riazantseva, 2001), it is still not clear from the literature at what juncture L2 speakers are more likely to pause. Moreover, this general tendency does not explain the current data, in which the grouping of words by syntactic junctures should differ depending on the intended syntactic structure (i.e., attachment of the RC). If it is the case that L2 speakers generally tend to pause more often at syntactic junctures, it is expected that their different attachment productions would differ in the relative pause durations at NP1-NP2 vs. NP2-RC boundaries. On the contrary, the current study found that they paused longer at NP2-RC boundaries regardless of the intended syntactic structure. This is a significant finding showing that they do not produce prosodic junctures to represent syntactic junctures. Rather, the prosodic grouping produced by the Korean-English bilingual speakers in this study, (NP1 NP2) (RC), is consistent with the general tendency in English for a prosodic break to be located close to the middle of the sentence, yielding an approximately symmetrical, or balanced, structure (Cooper & Paccia-Cooper, 1980; Fodor, 1998; Gee & Grosjean, 1983). Thus, the bilingual speakers seem to have followed this general tendency for prosodic bisection regardless of intended reading, failing to produce prosodic boundaries to mark corresponding syntactic boundaries for ambiguity resolution.

In short, the Korean-English bilingual speakers made better use of the L2 prosodic aspects that differ from their L1 prosody (word prominence) than those that are shared by their L1 and L2 (boundary marking). Some empirical studies on L2 segment learning also suggest that a novel L2 sound is more learnable than an L2 sound that has a similar counterpart in the L1 inventory. For example, in Aoyama et al.’s (2004) one-year longitudinal study, Japanese children showed a greater improvement in the perception and production of English /ɹ/, a novel sound to Japanese speakers, than that of English /l/, which is phonetically more similar to Japanese /r/. According to Flege’s (1995) Speech Learning Model, an L2 segment that is perceptually distinct from any L1 sound is more likely to be learned as a new independent phonetic category by L2 learners. On the contrary, an L2 sound that is perceptually similar to an L1 sound will not be perceived as a separate category and thus less successfully learned. The current findings suggest that this hypothesis of the Speech Learning Model can be adopted to L2 prosody learning. The pitch accent system in English is perceptually distinct from Korean prosody, which lacks a word prominence system. Prosodic boundaries in English, on the other hand, are phonetically similar to those in Korean. Consequently, the Korean-English bilingual speakers might have established new category representations for pitch accents, namely, relative pitch targets for tones and their alignment with lexical items, while having difficulty doing so for prosodic boundaries, such as target durations of a lengthened unit and its subsequent pause as well as the placement of prosodic boundaries in relation to syntactic boundaries. Therefore, a theory of L2 prosody learning must be able to account not only for learning difficulties predicted by typological differences across languages but also those arising from perceptual distances between L1 and L2 prosodic categories.

This study used a laboratory speech task with a construction that is relatively infrequent in spontaneous speech, which may raise a potential concern regarding the generalizability of the results. The motivation behind this design was to pay closer attention to the use of prosody while minimizing possible interference of segmental, lexical, and even contextual variables. The read-aloud task with no interlocutor created an environment in which prosody was the only cue that the speakers could freely use to communicate their intention. The fact that the results showed clear differences between L1 English and Korean-English bilingual speakers’ L1 Korean productions as well as between L1 English and Korean-English bilingual speakers’ L2 English productions highlights that prosody is indeed employed in different ways across languages even in the same context and can be an independent source of L2 learning difficulties.