Journal of Experimental Psychology:

Difficulty in Learning Similar-Sounding Words: A Developmental Stage or a General Property of Learning?

Bozena Pajak Duolingo Inc., Pittsburgh, Pennsylvania

Sarah C. Creel and Roger Levy University of California, San Diego

How are languages learned, and to what extent are learning mechanisms similar in infant native-language (L1) and adult second-language (L2) acquisition? In terms of vocabulary acquisition, we know from the infant literature that the ability to discriminate similar-sounding words at a particular age does not guarantee successful word–meaning mapping at that age (Stager & Werker, 1997). However, it is unclear whether this difficulty arises from developmental limitations of young infants (e.g., poorer working memory) or whether it is an intrinsic part of the initial word learning, L1 and L2 alike. In this study, we show that adults of particular L1 backgrounds—just like young infants— have difficulty learning similar-sounding L2 words that they can nevertheless discriminate perceptually. This suggests that the early stages of word learning, whether L1 or L2, intrinsically involve difficulty in mapping similarsounding words onto referents. We argue that this is due to an interaction between 2 main factors: (a) memory limitations that pose particular challenges for highly similar-sounding words, and (b) uncertainty regarding the language’s phonetic categories, because the categories are being learned concurrently with words. Overall, our results show that vocabulary acquisition in infancy and adulthood shares more similarities than previously thought, thus supporting the existence of common learning mechanisms that operate throughout the life span. Keywords: word learning, spoken word recognition, nonnative speech perception, second language

acquisition

Humans are able to learn languages throughout their life spans. But how similar are the learning mechanisms for infants acquiring their native language (L1) and adults learning a second language (L2)? Little work has been done to connect these two literatures, reflecting the underlying assumption of a lack of developmental continuity in terms of language learning (see, e.g., White, Yee, Blumstein, & Morgan, 2013, for discussion). Instead, infants and adults have been assumed to use qualitatively different mecha-

nisms to process and learn languages, largely following the critical period hypothesis (Lenneberg, 1967; Johnson & Newport, 1989). However, recent work has shown that while age of L2 acquisition negatively correlates with achieved proficiency, there are signs of developmental continuity in language learning and similarities between infant and adult acquisition (Birdsong, 2009; Birdsong & Molis, 2001; Hakuta, Bialystok, & Wiley, 2003; Werker & Tees, 2005; White et al., 2013). For example, it has been shown that infants and adults rely on similar statistical learning mechanisms to segment words out of a continuous speech stream (e.g., Saffran, Aslin, & Newport, 1996; Saffran, Newport, & Aslin, 1996)orto learn phonetic categories (e.g., Maye & Gerken, 2000; Maye, Weiss, & Aslin, 2008; Maye, Werker, & Gerken, 2002; Pajak & Levy, 2011; see Pajak, Fine, Kleinschmidt, & Jaeger, in press, for a review and further discussion), and are similarly affected by word familiarity during lexical processing of newly learned words (White et al., 2013). White et al. argued that these parallel results for infants and adults might reflect common mechanisms that operate throughout development, thus highlighting the need for greater interaction between the infant and the adult language learning literatures. Herein, we compared infant and adult language learning by

This article was published Online First March 10, 2016. Bozena Pajak, Duolingo Inc., Pittsburgh, Pennsylvania; Sarah C. Creel,

Department of Cognitive Science, University of California, San Diego; Roger Levy, Department of Linguistics, University of California, San This document is copyrighted by the American Psychological Association or one of its allied publishers.Diego. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.Supported by the National Institutes of Health (Training Grants T32-

DC000041 and T32-DC000035) from the Center for Research in Language at the University of California San Diego and the Center for Language Sciences at University of Rochester, respectively, and research grants from the University of California San Diego Academic Senate and the Alfred P. Sloan Foundation. For their helpful feedback, we thank Eric Bakovic, Klinton Bicknell, the Computational Psycholinguistics Lab at the University of California San Diego, the audience of LSA 86, and four anonymous CogSci 2012 reviewers. Rafi Feliciano, K. Michael Brooks, and Hannah Byers-Straus helped with data collection. We are grateful to Eugene Carsey for permission to use his photographs of mushrooms (http://www .eugenecarsey.com/). Correspondence concerning this article should be addressed to Bozena

http://www.eugenecarsey.com/ .eugenecarsey.com/).

considering one aspect that is crucial at the initial stages of acquisition: the encoding of phonetic detail during word learning. Learning words requires not only remembering a label for a given referent, but also forming a phonetically rich representation of that label by segmenting the word into individual sounds. The detailed phonetic representation is especially important for similarsounding words (e.g., bin vs. pin), because successful learning 1377

Pajak, Duolingo Inc., 5533 Walnut Street, 3rd Floor, Pittsburgh, PA 15232. E-mail: bozena@duolingo.com

E-mail: bozena@duolingo.com mailto:bozena@duolingo.com

crucially relies on the ability to distinguish between the words based on subtle acoustic–phonetic cues. Thus, the learner must be able to perceptually discriminate the sounds that distinguish between words (e.g., [b] vs. [p]), and ignore any irrelevant variability between instances of the same phonetic category (e.g., multiple exemplars of the word bin). The ability to discriminate among similar sounds is thus a necessary condition for successfully learning words distinguished by those sounds. But is it a sufficient condition? In this article, we investigated this question for adult learners by taking advantage of the influence of L1 background (here, Mandarin and Korean) on adult perceptual discrimination abilities. We examined to what extent the L1-driven differences in adult speech sound discrimination are associated with differences in word learning ability, building on a small existing body of work in this area (Creel & Dahan, 2010; Creel, Aslin, & Tanenhaus, 2006; Silbert et al., 2015). In the case of infant language learning, we know from the

literature that discrimination does not guarantee successful learning of similar-sounding words: Despite the ability at age 14 months to perceptually discriminate between similar sounds (e.g., b and d), 14-month-olds have been shown to confuse newly learned words differentiated by those sounds (e.g., bih and dih Pater, Stager, & Werker, 2004; Stager & Werker, 1997), unless there is additional contextual information, or less demanding learning conditions (Ballem & Plunkett, 2005; Fennell, Waxman, & Weisleder, 2007; Fennell & Werker, 2003; Rost & McMurray, 2009; Swingley & Aslin, 2002; Thiessen, 2007; Yoshida, Fennell, Swingley, & Werker, 2009). The initial explanation proposed for this result was a limited resource hypothesis (Stager & Werker, 1997; Werker, Fennell, Corcoran, & Stager, 2002): Because attending to fine phonetic detail while learning new words is computationally very demanding, young infants—who have limited attentional and cognitive resources—might have difficulty accessing full phonetic detail when focusing their attention on learning meaning. Other explanations have emphasized the role of increased lexical competition in learning similar-sounding words (Swingley & Aslin, 2002; Swingley & Aslin, 2007), or suggested that the difficulty might arise from poorly defined phonetic category boundaries at that stage of infant development (Rost & McMurray, 2009) and limited experience with phonological categorization (Yoshida et al., 2009). Regardless of the exact explanation, the consensus is that chil-

dren outgrow this initial difficulty, and by 17–20 months of age succeed at learning new similar-sounding words (Werker et al., 2002). However, despite this acquired sensitivity to minimal dif- This document is copyrighted by the American Psychological Association or one of its allied publishers. ferences between words in the learners’ L1, phonological similar- This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. ity continues to play a role in lexical processing in both older children and adults. This is indicated, for example, by robust and automatic activation of words that sound similar to the target word (e.g., Andruski, Blumstein, & Burton, 1994; Allopenna, Magnuson, & Tanenhaus, 1998; Magnuson, Dixon, Tanenhaus, & Aslin, 2007; Mani & Plunkett, 2011; Swingley & Aslin, 2000; White & Morgan, 2008). Adults are also slower at processing words that have a high neighborhood density (i.e., have a large number of similar-sounding words, generally defined as a onephoneme distance; e.g., Luce & Pisoni, 1998) compared to words in sparse lexical neighborhoods (e.g., Vitevitch & Luce, 1998), and have increasing difficulty distinguishing phonologically nativelike nonsense words as the word similarity increases (Creel &

Dahan, 2010; Creel et al., 2006). All these results suggest gradient effects of phonological similarity, where the encoding and the retrieval of similar-sounding minimal pair words are impaired relative to dissimilar words. (But see, e.g., Storkel, 2004; Storkel, Armbrüster, & Hogan, 2006, for evidence that children and adults learn new dense-neighborhood words in their native language more readily than new sparse-neighborhood words, suggesting that partial phonological overlap with known words may help strengthen newly formed lexical representations.) Thus, both children and adults are known to have difficulty

learning novel similar-sounding words whose phonological form resembles their native language. These results do not, however, answer the question whether adults are affected by phonological similarity during learning of an unfamiliar language, a situation more parallel to the case of 14-month-old infants learning their native language. As we mentioned earlier, it is known that adult L2 learners have

extreme difficulty distinguishing—and therefore also learning— similar-sounding words that involve novel sound contrasts not found in their native language, such as rake versus lake for native speakers of Japanese (e.g., Escudero, Broersma, & Simon, 2013; Escudero, Hayes-Harb, & Mitterer, 2008; Hayes-Harb & Masuda, 2008; Weber & Cutler, 2004). In those cases, L2 learners have to override the L1 phonetic category information that is incompatible with the L2 information (e.g., the acoustic–phonetic range occupied by the English “r” and “l’ roughly corresponds to a single category in Japanese; e.g., Miyawaki et al., 1975; see also, e.g., Escudero, Simon, & Mulak, 2014, for how orthography may help or hinder learning in these cases). Indeed, it is already known that, within a given L1, listeners’ ability to discriminate a nonnative contrast predicts how well they learn words that differ by that contrast (Silbert et al., 2015). It is also known that individual differences in learning may in these cases arise from variability in purely auditory abilities (Kidd, Watson, & Gygi, 2007), as well as variability in phonological short-term memory (Silbert et al., 2015). We know less, however, about listeners’ overall abilities—as a group—to learn similar-sounding words when (a) they can reliably perceive the perceptual contrast and (b) the L1 phonetic– category information does not strongly interfere with L2 perception (cf. Pajak & Levy, 2014). This is the topic of the present study. That is, instead of trying to predict an individual’s ability to learn words from that individual’s ability to discriminate those words, as in prior work (e.g., Silbert et al., 2015), we investigated the relation between perceptual discrimination and word learning in listeners as a function of one of two different native-language backgrounds. We used two native-language populations that we know have complementary expertise in perceptual discrimination: Mandarin speakers, who are sensitive to nonnative sibilant place-of-articulation distinctions, and Korean speakers, who are sensitive to nonnative consonant length distinctions (Pajak & Levy, 2014). In particular, we examined two specific questions. First, is the

mismatch between discrimination and word learning a developmental phenomenon, or is it driven by the information being learned? That is, will adult L2 learners—who, like infants, are concurrently learning the language’s phonological categories, but unlike infants, have vastly greater working memory capacity— show greater difficulty in word learning tasks relative to discrimination tasks when acquiring new L2 vocabulary? Second, how does phonological similarity moderate discrimination versus learning of similar-sounding L2 words? Answers to these questions for the two different L1 populations

we studied can potentially provide a key missing link connecting theories of adult and infant language learning. If adults of particular L1 background—just like 14-month-old infants—are found to have difficulty learning similar-sounding words that they can nevertheless discriminate perceptually, then this would provide evidence in favor of the existence of common language learning mechanisms that operate throughout development and into adulthood. Furthermore, a more detailed examination of the role of phonological similarity in word discrimination and learning, and how similarity interacts with different task demands, can help us shed more light on the nature of those common mechanisms that underlie language learning. Before we continue, however, we first examine factors that might

contribute to the difficulty of learning the correct label/referent pairing. One such factor is that beginner learners, both infant and adult, might have noisy phonetic representations, reflecting low confidence in the fidelity of phonetic encoding of the newly learned words or in the exact location of phonetic category boundaries (Rost & McMurray, 2009; Yoshida et al., 2009). Another factor might involve taskspecific memory limitations, which pose particular challenges for highly similar-sounding words. In the next section, we describe in detail how these factors might lead to potential difficulty in word learning relative to discrimination.

A Conceptual Model of Word–Referent Mapping Difficulties Learners acquiring words need to rely on their memory repre-

sentations of label/referent pairs, where each label can be described as a sequence of sounds sampled from the language’s phonetic categories. Precise encoding of the label’s phonetic form thus requires establishing from which categories the sounds were sampled, a difficult task at the early stages of language learning. Word learning is then likely affected by two sources of noise: (a) noise and uncertainty associated with categorization of each individual sound and (b) noise associated with retrieving a memory trace of the phonetic input and of the label/referent pairing. Thus, one way of thinking about the difference between the discrimination and the word learning tasks is that the quality of phonetic representations for individual input exemplars is lower for word learning than for discrimination due to heavier long-term memory demands in the word learning task: In word learning, listeners have This document is copyrighted by the American Psychological Association or one of its allied publishers. to simultaneously keep track of the referent and try to form This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. phonetic representations, while a discrimination task only requires comparing short-term memory traces of phonetic input, without the need to link them to referents. As more data are obtained over time, successful word learning requires integrating multiple memory traces to arrive at the correct label/referent pairing. This would correspond to narrowing the effective variance around the memory representation of the label’s phonetic form. In cases when the words being learned are composed of highly

dissimilar sequences of sounds—that is, the phonetic distance between sound categories is large relative to the variance of the label’s phonetic form—the learner’s performance should not be impeded because a small number of samples would be sufficient to learn the distinctions among the categories. However, when the

words are highly similar—that is, the phonetic form variance is large relative to the distance between sound categories—it should be much harder for learners to separate the sounds into categories and pinpoint the right label/referent pairings. It is expected that much more data (i.e., more learning instances) would be needed in this case before learners can accumulate a sufficient number of exemplars to learn phonetic category distinctions and form correct label/referent mappings. Furthermore, if the learner is unable to integrate the information accumulated from a number of exemplars of the label/referent pairing, he or she will have great difficulty learning the pairing reliably at all. This conceptual model is consistent with prior suggestions that

additional cognitive load, such as simultaneous presentation of visual stimuli, lowers the resolution of auditory processing of phones (Mattys & Palmer, 2015). Such lower resolution processing may be due to missing some temporal pulses in the auditory signal (Casini, Burle, & Nguyen, 2009) or to reduced cochlear sensitivity (Lukas, 1980; Puel, Bonfils, & Pujol, 1988) during concurrent attention to visual stimuli. Interestingly, perceptual sensitivity seems to linearly decrease as the effort involved in the simultaneous visual task increases (Mattys, Barden, & Samuel, 2014). This type of disruption in auditory processing may be understood as an increased tolerance to imprecise acoustic encoding, and, as a result, to an increased perceptual overlap between similar-sounding phones (Mattys et al., 2014; Mattys & Palmer, 2015).

The Current Study In the current study, we examined how adults learn vocabulary

in a new language that is phonologically unfamiliar (i.e., an L2), but composed of discriminable speech–sound categories. In particular, we compared two populations of participants with differential perceptual sensitivities to certain speech sound contrasts that are due to their different L1 backgrounds. We tested one participant group on a discrimination task and another group on a word learning task, and examined whether the known L1-backgrounddriven differences in sound discrimination would also be observed in the word learning task when participants learned words that differed by those sounds (as described in more detail below). The situation of learning phonologically novel words that are similar sounding, but that adults can nevertheless discriminate perceptually, is analogous to the situation of 14-month-old infants observed in Stager and Werker (1997). This allowed us to assess whether the good-discrimination-without-learning pattern observed in Stager and Werker (1997) reflects a purely developmental phenomenon or reflects general mechanisms of (language) learning. In addition, we included multiple sets of word pairs that differed in their degree of similarity, which let us investigate how phonological similarity modulates discrimination and learning of L2 words. We constructed a miniature language with pairs of words at

three levels of similarity: (a) dissimilar (e.g., [tala]–[kenna]), (b) similar (e.g., [tala]–[taja]), and (c) highly similar, where the words differed either in consonant length (e.g., [taja]–[tajja]) or in place of articulation between alveolopalatal and retroflex sounds (e.g., [got a]–[got a]). We chose the length and the place dimensions because they have been shown to be differentially discriminable by two different L1-speaker populations: L1 Korean and L1 Mandarin (Pajak, 2012; Pajak & Levy, 2014). In particular, Korean speakers have an advantage over Mandarin speakers in discriminating consonant length contrasts, while Mandarin speakers have an advantage over Korean speakers in discriminating alveolopalatal and 1 retroflex consonant contrasts. Therefore, we were able to investigate whether these differential L1-based perceptual advantages on highly similar word pairs occurred not only in discrimination but also in word learning. Note that we were not asking whether performance on the discrimination task predicted performance in the word learning task at the individual level. Rather, we investigated whether between-groups differences in discrimination ability arising from differences in native language would also be reflected in between-groups differences in word learning ability. There might certainly be some individual variation in how well learners take advantage of their L1-based perceptual abilities when learning words, but the group-level comparisons reveal the overall trends in the population as a whole, and this is the question we addressed here. We did not expect differences in word learning performance between L1-Korean and L1-Mandarin participants for dissimilar and similar items because those contrasts were acoustically more salient (relative to the contrasts in highly similar pairs) and there is no reason to believe that participants’ language background would affect their discrimination in a differential way. Crucially, the language was phonologically novel to all participants in that all phonetic properties of the stimuli (e.g., voice onset time, vowel quality, stress) were taken from an unrelated language, Polish. Therefore, this scenario was more comparable to the situation of the 14-month-old infants than many previous studies in which adults learned phonologically native-like words. We had two main sets of predictions. The first concerned the

overall effect of phonological similarity on discrimination versus learning of L2 words that are all discriminable by learners, either because the differences are salient (dissimilar, similar) or because a related distinction is used in the learners’ L1 (highly similar length for Korean, place for Mandarin). Given the gradient acoustic similarity between the different sets of word pairs, we expected discrimination also to be gradient: best for dissimilar words, intermediate for similar words, and poorest for highly similar words. As for word learning, prior work using native-like words has shown gradience in performance as a function of words’ phonological similarity (e.g., Creel & Dahan, 2010; Creel et al., 2006). But how does phonological similarity interact with word learning in the case of learning a new language with overall nonnative phonology? If it works similarly to learning vocabulary in an L1 phonological system, then we would expect learners’ performance to change as a function of similarity between the word pairs, This document is copyrighted by the American Psychological Association or one of its allied publishers. matching the discrimination performance: best when identifying This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. the referent in the context of two dissimilarly named possible referents, intermediate for similarly named referents, and poorest for highly similar pairs. On the other hand, it is possible that learning words with an unfamiliar phonology is not affected by similarity in the same way that discrimination is. That is, we might expect a mismatch between the discrimination versus the word learning task: For example, gradient performance for participants in the discrimination task, but no differences in performance for participants identifying word referents; or more exaggerated gradient effects in one task than the other. Our second set of predictions concerned highly similar words.

These predictions are in some sense a more focused version of the first set of predictions, because both are examining the influence of

perceptual similarity. However, here we focused on the highly similar words whose underlying speech sound categories are acoustically overlapping, thus being most comparable to infants’ nascent speech sound categories. More specifically, we compared two populations of speakers with differential perceptual sensitivities to these word differences: Korean speakers better on length and Mandarin speakers better on place. We made the Koreanversus Mandarin-speaker comparison for two different tasks, word learning and discrimination, where each task was completed by independently recruited subjects. Therefore, we tested word learning versus discrimination of highly similar words that either (a) were relatively easily discriminable because the differences were based on a phonetic dimension informative in L1 (even though the learners did not actually know any words distinguished by some of the specific sounds used, thus resembling the situation of young infants) or (b) were not easily discriminable because the differences were based on a phonetic dimension not informative in L1. The contrasts in (a) were length for Korean speakers and place for Mandarin speakers, while the contrasts in (b) were place for Korean speakers and length for Mandarin speakers. We expected the word learning task to be harder than the

discrimination task, which would be reflected in overall worse performance for the group performing the word learning task. However, overall task differences do not tell us about participants’ use of their native-language-based perceptual abilities in word learning. The goal of our discrimination task was twofold: (a) to replicate prior results (Pajak, 2012; Pajak & Levy, 2014) with materials more comparable to our word learning task materials, and (b) to show how much perceptual advantage Korean speakers have over Mandarin speakers in length contrasts, and how much perceptual advantage Mandarin speakers have over Korean speakers in place contrasts. Comparing the degree of this group-level asymmetry in the discrimination versus the word learning tasks let us assess how much participants use their L1-based perceptual abilities in word learning. Therefore, we expected both Korean and Mandarin speakers to perform worse on highly similar trials in word learning than in discrimination (i.e., overall task differences). However, the critical question was: Is the relative difference between the two L1 populations (Korean better at length; Mandarin better at place) also observed in the word learning task? If it was, then it would suggest that L1-based perceptual abilities are used in word learning (whether or not the overall performance was lower than in discrimination). If it was not, then it would suggest that the word learning task makes it difficult for participants to use their L1-based perceptual abilities. The extreme version of the latter would be no group-level difference between Korean and Mandarin speakers on length or place in the word learning task, showing that the whole relative perceptual advantage observed at the group level has been eliminated during word learning. It is also the latter case (no use of perceptual advantages evident during word learning) that would be most analogous to the results reported for L1- learning 14-month-olds.

Pajak and Levy (2014) argued that these differential perceptual sensitivities follow from the fact that Korean has some length distinctions, while Mandarin has none, whereas Mandarin has some alveolopalatal and retroflex sounds, while Korean only has alveolopalatal but no retroflex sounds (Lin, 2001; Sohn, 1999).

Method

Participants

Ninety undergraduate students at the University of California San Diego participated in the experiment for course credit or payment. Half were speakers of Korean and other half were speakers of Mandarin. We recorded participants’ language background information, including self-reported proficiency in both Korean/Mandarin and English, current language exposure, as well as the scores from the Shipley Vocabulary Test (Shipley, 1967)as a measure of English proficiency. Participants varied in length of residence in the United States: some were born in the United States, while others immigrated at some point after birth or were international students who arrived very recently. Consequently, they varied in English proficiency. However, they all learned Korean or Mandarin from birth, reported high proficiency in those languages, and still used them regularly, predominantly with family. In most cases, they had some high school and/or college exposure to Spanish or French. Some Mandarin speakers were also familiar with Taiwanese Hokkien, mostly through family exposure. No exposure to any other languages was reported, including Polish, the language that provided source material for novel words. All participants reported no history of speech or hearing problems. We collected individual measures of participants’ nonverbal IQ using the Matrices subtest of the Kaufman Brief Intelligence Test (Kaufman & Kaufman, 2004). All participant characteristics are shown in Table 1. More detailed comparisons between participants depending on language background are provided in the Appendix A. To avoid potential carryover effects from one task to the other,

we tested discrimination and word learning in a betweenparticipants design, investigating whether between-L1-population differences in discrimination ability arising from differences in native language would also be reflected in between-L1-population differences in word learning ability. Task cross-contamination in a within-subjects design would be a serious barrier in interpreting the results: whichever order of tasks we might choose, participants would be biased in the second task because (a) their attention would be directed to the tested contrasts and (b) they would have received a great deal of perceptual exposure to those distinctions. In fact, there is ample evidence from the perceptual learning literature that even relatively brief exposure can affect adults’ perception (e.g., Clayards, Tanenhaus, Aslin, & Jacobs, 2008; This document is copyrighted by the American Psychological Association or one of its allied publishers.Kraljic & Samuel, 2005; Norris, McQueen, & Cutler, 2003). This This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.design choice is analogous to infant studies, where the results regarding the dissociation of discrimination and word learning abilities are based on differences between groups. Fifty-four participants were assigned to the word learning task

2 and 36 to the discrimination task. In each group, half were speakers of Korean and the other half were speakers of Mandarin. Comparing participants assigned to the discrimination versus the word learning task revealed no significant differences in any of the measures we collected (see Table 1).

contrasts tested by Pajak, 2012; Pajak & Levy, 2014). The words were constructed such that there were eight minimal pair words differing only in the middle consonant; these were the highly similar word pairs. The complete list of trial types is described in the Procedure section. More specifically, the minimal pairs differed either in length (a short vs. a long middle consonant) or place of articulation (an alveolopalatal vs. a retroflex sibilant consonant 3 in the middle position). The materials were constructed using the sound inventory and other phonological properties of Polish, and were recorded by a phonetically trained Polish native speaker. The inventories of Korean and Mandarin include some sound

materials consisted of 16 bisyllabic consonant–vowel– consonant-vowel (CVCV) nonce words (see Table 2; a subset of

distinctions along the dimensions of length and place, respectively, that are similar but not identical to the distinctions used in the experiment. Korean employs the dimension of length, distinguishing between short and long sounds, but not the dimension of place We found previously that Korean speakers are better than Mandarin speakers at discriminating consonant length contrasts, while Mandarin speakers are better than Korean speakers at discriminating alveolopalatal versus retroflex place contrasts (Pajak, 2012; Pajak & Levy, 2014). This follows from the fact that Korean has some length distinctions, while Mandarin has none; whereas Mandarin has some alveolopalatal and retroflex sounds, while Korean does not (Sohn, 1999; Lin, 2001). More specifically, Korean uses length distinctions mostly on

Materials The

vowels (e.g., [pul] “fire” vs. [pu:l] “blow”), but some long consonants ([ll], [nn], [mm]) arise from phonological assimilation pro-

The difference in the number of participants in the two tasks was due to the fact that we have had more experience with discrimination experiments, and so we had a better sense of how many participants we would need to obtain good statistical power. Studying word learning in this type of task was relatively novel to us, and we expected more between-subjects variability, which is why we decided to collect data from more participants. ³ We chose the middle consonants in our stimuli in such a way that half

of the corresponding sound distinctions had their analog in the listener’s L1, and the other half did not. This was done to compare performance between distinctions that were relatively familiar to our participants from their L1s versus completely unfamiliar distinctions that yet varied along familiar dimensions. In previous work on length and place discrimination by Korean and Mandarin speakers (Pajak, 2012; Pajak & Levy, 2014), Korean speakers outperformed Mandarin speakers on discriminating all length contrasts, whether familiar or not, while Mandarin speakers outperformed Korean speakers on discriminating all place contrasts (note that the stimuli in that study were also based on the Polish length and place contrasts). However, there was a trend in that earlier study for both groups to perform slightly better at the distinctions familiar from their L1s compared to the unfamiliar distinctions that varied along familiar dimensions (e.g., familiar [m]–[mm] unfamiliar [j]–[jj] for Korean speakers). Therefore, we expected that a similar difference might hold in a word learning task: that is, both Korean and Mandarin speakers would be better at learning similar-sounding words that included familiar categories than those that included unfamiliar categories which nevertheless varied along a familiar dimension. (See Table 2 for the list of word contrasts based on (a) familiar categories, where the specific distinction exists in Korean/ Mandarin, and (b) unfamiliar category contrasts, but familiar phonetic dimensions.) To obtain sufficient power for such a comparison, but at the same time keep the total number of words relatively small to ensure their learnability in a single experimental session, we decided to focus our analysis on length words (12 words in total), and included a much smaller number of place words (four words in total). Length was chosen as the dimension of main interest because length contrasts are possible for many more types of segments than the alveolopalatal versus retroflex contrasts (of which Polish only has four). No difference between familiar categories versus unfamiliar category contrasts but familiar dimension was borne out in the current results.

Table 1 Individual Measures: Participants in the Discrimination vs. the Word Learning Task Discrimination task

Note.L1 native language. ^a On a 0 –10 scale (0 none

and 10 perfect). If born in the United States, coded as 0.

cesses (Sohn, 1999), and Korean tense obstruents ([p=], [t=], [k=], [s=], [t =]) have sometimes been analyzed as long (Choi, 1995). In terms of place of articulation, Korean has some alveolopalatal sounds ([ ], [t ]), but no retroflex sounds, thus lacking the place contrast as defined in this article. Mandarin, on the other hand, has both alveolopalatal and retro-

$$ [\mathfrak{s}^{\prime}],[\mathfrak{t}\mathfrak{c}^{\prime}]\big) $$

flex sounds that are distinguished by spectral shape in the frication noise (the place dimension), but does not use the length dimension. In particular, Mandarin has voiceless alveolopalatals ([ ], [t ]) and retroflexes ([ ], [t ]) as allophones of the same phonemic category. In addition, the voiced retroflex fricative ([]) is a betweenspeaker variant of the retroflex approximant ([ ]). Other voiced sibilants are assumed to be absent because Mandarin has obstruent distinctions in aspiration, not in voicing (Lin, 2001). Note, however, that the analogous place distinction in Polish, which we used in the stimuli, is not exactly the same as that in Mandarin, differing somewhat in the placement of the tongue tip. Note that all participants spoke American English, where length

and alveolopalatal versus retroflex place are not used contrastively. While vowel length varies in English, it correlates with other cues (e.g., the tense–lax distinction), and native speakers of English identify vowels relying predominantly on spectral properties (e.g., Hillenbrand, Clark, & Houde, 2000). Long consonants This document is copyrighted by the American Psychological Association or one of its allied publishers.

are sometimes attested, but only at morpheme boundaries (e.g., dissatisfied; Benus, Smorodinsky, & Gafos, 2003), and only produced as long by some speakers (Kaye, 2005). English has neither alveolopalatal nor retroflex obstruents, although some speakers produce the alveolar approximant [ɹ] as retroflex (Ladefoged & Maddieson, 1996; Westbury, Hashi, & Lindstrom, 1998). While it is possible that knowledge of English might affect discrimination of the Polish length and place contrasts (and, in particular, help with the length contrasts), Pajak and Levy (2014) found no evidence to support that hypothesis when testing discrimination of these contrasts by a variety of bilingual listeners (English–Korean, English–Vietnamese, English–Cantonese, and English–Mandarin), whether their dominant language was English or their native language. The materials were recorded in a soundproof booth by a pho-

netically trained native speaker of Polish. There were 10 tokens recorded for each word. For length words, two tokens of each word with long consonants were chosen for the experiment. Subsequently, words with short consonants were created by shortening the tokens with long consonants in a way that, for each word and each recording, the naturally recorded long consonant was reduced to half its duration so to maintain a constant 2:1 duration ratio (cross-linguistically, the long-to-short consonant ratio varies be-

Table 2 Stimuli (in International Phonetic Alphabet) Length words

$$ \ {\mathrm{Y e s\ (})^{\mathrm{a,b}}} $$

$$ \mathrm{Y e s^{b}} $$

$$ \mathrm{Y e s^{c}} $$

$$ \mathrm{Y e s^{b}} $$

general distinction exists, the exact place of articulation in Mandarin differs for both the alveolopalatal and the retroflex, and the distinction is only allophonic.

tween 1.5 and 3; Ladefoged & Maddieson, 1996). For place words, given that the alveolopalatal versus retroflex distinction is intrinsically already very subtle, even in natural speech (Nowak, 2006), we used natural recordings of both alveolopalatals and retroflexes with no additional manipulations. Two tokens each were chosen for the experiment with the goal of maximizing the similarity between the words in minimal pairs with regard to how vowels were pronounced, but at the same time choosing tokens with clearly enunciated sibilants. This was a departure from how the stimuli were constructed by Pajak (2012) and Pajak and Levy (2014), where both alveolopalatals and retroflexes were spliced into an identical word frame. Pajak and Levy’s procedure removed one of the cues to the contrast (the formant transition into the following vowel), thus making it extremely subtle. In the current study, we left this cue intact so that stimuli in the word learning task were not overly difficult. These auditory stimuli were used for both the discrimination and

the word learning task. For the word learning task, each word was paired with a picture of a different kind of mushroom (see two examples in Figure 1), which were chosen to include objects unfamiliar to our participants, but not so unfamiliar that participants would find them bizarre and hard to remember. We selected pictures that varied in shape and color so to maximize visual differences between them. We created four different one-to-one word-to-picture mappings that were counterbalanced between participants to ensure that the results were not driven by any peculiarities in the mappings we chose.

Procedure

Participants sat in front of a computer and responded by using a mouse. They were instructed that, in this experiment, they would be listening to a novel language, and, specifically, either (a) learn to distinguish this language’s sounds (in the discrimination task) or (b) learn the language’s words for different types of mushrooms (in the word learning task). The experiment was completed in a single session, and each participant took part in only one of the tasks. The discrimination and the word learning tasks were made equal in the total auditory exposure to each stimulus in order to keep them as parallel as possible.

Discrimination Task Discrimination was tested in an ABX task. In each trial, three

words were presented auditorily through headphones: A [500 ms] This document is copyrighted by the American Psychological Association or one of its allied publishers.

1 1 2 whether X sounded more like A or more like B. As indicated by subscripts, the X word was always acoustically different (i.e., a physically different recording) from both A and B words to ensure that the matching of X onto A or B was not based on pure acoustical identity of two tokens. This procedure differed from that in Pajak (2012) and Pajak and Levy (2014), where an AX task was used. In this study, however, we wanted to maintain a close parallel between the discrimination and the word learning tasks, which was achieved with the ABX procedure. The AB word order was counterbalanced, and the trial order was randomized for each participant. There were four blocks, each with 64 trials and lasting about 5 min. Note that this meant that each block included exposure to 192 words (64 trials 3 words per trial). Blocks were separated by self-terminated breaks. There were four types of trials depending on the AB contrast, as illustrated in Table 3: (a) dis- similar word pairs (e.g., [tala]–[kenna]), which differed in all sounds but the last vowel (16 trials; 8 AB word pairs 2 trials: one trial with X A and one trial with X B), (b) similar word pairs (e.g., [tala]–[taja]), which shared the initial CV sequence, but the middle consonants differed along multiple phonetic dimensions (16 trials; 8 AB word pairs 2 trials: one trial with X A and one trial with X B), (c– d) highly similar word pairs (e.g., length: [tala]–[talla] or place: [got a]–[got a]), where the initial CV sequence was identical and the middle consonants differed minimally, either in length or in place (32 trials per block: 8 AB word pairs 4 trials: two trials with X A and two trials with X B). Word Learning Task

In the word learning

task, participants learned to associate auditorily presented words (one word per trial) with pictures of mushrooms. There were four training blocks (each with 128 trials, about 10 –15 min long) and four testing blocks (each with 64 trials, about 5 min long), interleaved. Thus, each train plus test combination contained 128 training trials 64 test trials 192 auditory exposures to the words, the same number of exposures as a block of the discrimination task (64 trials 3 words per trial 192). Blocks were separated by self-terminated breaks. In each trial, two pictures were presented on a computer screen (see Figure 1), and after a delay of 500 ms, a word was played through headphones. Participants were asked to click on the picture that they thought went with the word. In training, feedback was provided after the response in the form of the correct picture staying on the screen. A This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Figure 1. Example of a screen shot from the word learning task. See the online article for the color version of this figure.

Table 3 Trial Types in Discrimination (AB Words Presented Auditorily) and Testing in Word Learning (AB Labels for Visually Presented Pictures) Dissimilar Similar Highly similar

Note. For each pair, both orders of presentation were tested.

mouse click triggered the start of the next trial. Because participants were learning via feedback presented after each response, early responses were necessarily random. Participants were told to guess at first, and that through feedback they would eventually learn the correct word-to-picture mappings. In testing, no feedback was provided. The training trial types consisted of picture pairs that were

always associated with dissimilar word pairs (e.g., [taja]–[diwa], [got a]–[kemma]; see Table 3) so that participants were not directly alerted to the distinctions of interest. Each word was played eight times per training block (8 16 words 128 total), and each time it was accompanied by a different set of two pictures. None of the training picture pairings appeared in later testing. The Testing trial types always differed from the training trials,

and were completely analogous (in form and number) to trials in the discrimination task, as illustrated in Table 3. Specifically, each picture pair in the word learning task test was an analog of an AB word pair in the discrimination task, and the auditorily presented word in the word learning task corresponded to the X word in the discrimination task. This meant that there were the following trial types: (a) dissimilar picture pairs (e.g., picture of [tala] and picture of [kenna]) (16 trials; 8 picture pairs 2 trials: one trial where the auditorily presented word corresponded to the picture on the left, and one trial where the auditorily presented word corresponded to the picture on the right), (b) similar word pairs (e.g., picture of [tala] and picture of [taja]) (16 trials; 8 picture pairs 2 trials: one trial where the auditorily presented word corresponded to the picture on the left, and one trial where the auditorily presented This document is copyrighted by the American Psychological Association or one of its allied publishers. word corresponded to the picture on the right), (c– d) highly similar This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. (e.g., length: picture of [tala] and picture of [talla]; or place picture of [got a] and picture of [got a]) (32 trials per block; 8 picture pairs 4 trials: two trials where the auditorily presented word corresponded to the picture on the left, and two trials where the auditorily presented word corresponded to the picture on the right). Picture position was counterbalanced for both training and test-

ing trials. Trial order was pseudorandomized: we created four randomized lists and then altered them manually so that the same word was never repeated in two consecutive trials. Furthermore, minimal pair trials were always separated by at least two other trials. Each participant heard each list once, with a different list for each block. Block order was counterbalanced across participants.

Results

We investigated two main questions in our data analysis, based

on the two sets of predictions that guided our study’s design. First, how does phonological similarity moderate discrimination versus learning of similar-sounding words in an L2 with clearly nonnative phonology? Second, when testing phonologically highly similar word pairs, is an L1 background that confers good discrimination sufficient for good learning of similar-sounding words in adult L2 learners, or is there a disconnect between discrimination and word learning like that observed in infant L1 learners (Stager & Werker, 1997)? To answer these questions, we analyzed accuracy scores from

both discrimination and testing in word learning with mixedeffects logit models (Jaeger, 2008). Following Barr, Levy, Scheepers, and Tily (2013), who recommended maximal random-effects structures for mixed-effects models as best practice, we included random intercepts for participants and items, and random slopes for participants and items for all effects of interest (including interaction effects) that were respectively manipulated within par- 4 ticipants or within items. There was no stepwise model selection. We controlled for participants’ nonverbal IQ and L1 proficiency and use—through a combined score of (a) proficiency of L1 in speaking, (b) proficiency in L1 understanding, and (c) the percentage of time of current L1 exposure— by including them as fixed 5 effects in the models. All the binary and continuous predictor variables were centered; three-level variables were coded using successive differences contrast coding. The reported p values came from the z statistic. For highly similar trials, the difference between familiar categories and familiar dimensions only (as shown in Table 2 and discussed in footnote 5) was not significant in either task, so we did not report it in the analysis.

Gradient Phonological Competition in L2 Learning? We begin by addressing the second question concerning the

effects of phonological similarity on discrimination and learning of L2 words. We analyzed trials from both discrimination and word learning tasks for word pairs that were all expected to be discrim- inable by learners due to their familiarity with the tested sound contrasts from their L1s. That is, the analysis included the following trials: dissimilar (e.g., [tala]–[kenna]), similar (e.g., [tala]– [taja]), and a subset of highly similar pairs, depending on participants’ L1: length (e.g., [tala–talla]) for Korean speakers and place (e.g., [got a]–[got a]) for Mandarin speakers. Crucially, we withheld from this analysis the following highly similar pairs: place for Korean speakers and length for Mandarin speakers, which, based on prior work, we knew would not be easily discriminable by the learners due to their native language backgrounds. These trials were analyzed separately, and the results are reported in the next section, where we address the main question regarding the discrimination versus word learning performance.

analysis of variance procedures. ⁵ We checked the fit of our two main models (Task Trial Type

Language and Task Feature Type and the L1 proficiency/use factors. InLanguage) with and without the IQ both cases, adding these factors significantly improved the model fit (ps.05); the other effects remained unaffected across the models.

This document is copyrighted by the American Psychological Association or one of its allied publishers.

Based on prior work, we expected discrimination to be gradient: best for dissimilar words, intermediate for similar words, and poorest for highly similar words. Of most interest was the comparison between discrimination and word learning in order to evaluate how phonological similarity moderates performance across different tasks. The results are illustrated in Figure 2 (see Appendix Figure B1

for the results by block). We began the analysis by evaluating a model with fixed effects of Task (discrimination, word learning Trial Type (dissimilar, similar, highly similar), and Language (Korean, Mandarin). Trial Type was coded with similar trials as the reference level so that we would be able to directly compare similar and dissimilar trials, as well as similar and highly similar trials. As expected, there was a significant effect of Trial Type in that accuracy, pooled across the discrimination and the word learning tasks, varied in accordance to the similarity between words: the responses on dissimilar trials were significantly higher than on the similar trials (p.001), which in turn were higher than highly similar trials (p.001). Furthermore, there were significant Task Trial Type interac-

tions (ps.001; note that there were two interaction terms due to the contrast coding of Trial Type), suggesting that accuracy on each type of trial was moderated by the task: discrimination versus word learning. To examine this further, we directly compared performance in discrimination and word learning separately for each Trial Type using models with the fixed effect of Task (dis- crimination, word learning). For dissimilar trials, there were no main effects of task (p.59), suggesting that performance did not differ across tasks (although it is possible that an underlying difference between tasks was masked by ceiling effects, given that overall performance was above 95%). However, for both similar and highly similar trials, we found significant main effects of Task (ps.001): higher overall performance in discrimination than in the word learning task. No other effects in the full model were significant, including

effects or interactions involving Language (ps.2). This suggests that the two L1 populations had similar overall response patterns on the word pairs that were predicted to be relatively well discriminated by all participants (i.e., all word pairs excluding place trials for Korean speakers and length for Mandarin speakers).

In sum, these results suggest that there is a gradient effect of phonological similarity in both word discrimination and word learning: performance decreases as similarity grows. However, this effect is moderated by the specific task: relative to discrimination, performance in word learning suffers substantially more as similarity increases.

Discrimination Versus Word Learning: Are L2 Learners Like Infants?

Addressing whether L2 learners show a discrimination–word

learning asymmetry like that observed in L1-learning infants (Stager & Werker, 1997) entails a specific comparison between the two different highly similar trial types—length and place—for both L1 populations and across the two different tasks. Given previous studies with similar stimuli (Pajak, 2012; Pajak & Levy, 2014), we expected differential discrimination of length and place contrasts by the two L1 populations: Korean speakers more accurate than Mandarin speakers on discriminating length trials, and Mandarin speakers more accurate than Korean speakers on discriminating place trials (a Feature Type Language interaction). The question was whether this interaction would also extend to the word learning data. The results are illustrated in Figure 3. We analyzed these data in

a model with fixed effects of Task (discrimination, word learning), Feature Type (length, place), and Language (Korean, Mandarin). (See Appendix B for additional analyses that include Block as a fixed effect, demonstrating that the main result of a Feature Type Language interaction was consistent throughout the experiment.) First, there was a significant main effect of Task (p .001): performance was overall higher in discrimination than in word learning, indicating that the latter task was more difficult. Furthermore, there was a significant Feature Type Language interaction (p.001): as expected, Korean speakers performed better on length trials and Mandarin speakers performed better on place trials. Critically, however, there was also a significant threeway Task Feature Type Language interaction (p.001), indicating that performance as a function of Feature Type differed across the two tasks. To interpret this interaction, we analyzed each task separately.

Figure 2. Results for dissimilar, similar, and highly similar (length for Korean, place for Mandarin) trials. Accuracy scores indicate proportion of correct responses and error bars are standard errors.

Figure 3. Results for all highly similar trials: length and place. Accuracy scores indicate proportion of correct responses and error bars are standard errors.

For the discrimination task (see Figure 3, left), there was a significant Feature Type Language interaction (p.001): as predicted, Korean listeners performed better on length trials, while Mandarin listeners performed better on place trials. For the word learning task (see Figure 3, right), however, there was no interaction between the two variables (p.21; nor any main effects, ps.5), suggesting that the respective group-level perceptual advantages of Korean and Mandarin speakers did not translate into an advantage during word learning. Therefore, the three-way interaction in the main model indicated

$$ (p<.00) $$

$$ p{\mathrm{s}},>.,55) $$

that good discrimination does not necessarily yield better learning. Overall, these results suggest that learners did not take full advantage of their L1-based perceptual abilities in a word learning task. This pattern is highly similar to 14-month old infants learning their native language (Stager & Werker, 1997), suggesting that discordant performance between discrimination and mapping is not a developmental phenomenon, but a more general feature of learning words in a new phonology.

Do Individual Differences Affect Use of Perceptual Advantages in Word Learning? Word learning is a complex task that involves a combination of

cognitive abilities and attention. Therefore, it is possible that there is a high degree of individual differences in how efficiently learners make use of their L1-based perceptual abilities in the word This document is copyrighted by the American Psychological Association or one of its allied publishers. learning task. To answer that question, we examined the word This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. learning data separately for two groups of participants: higher and lower performers. We split participants into higher and lower performer groups

based on their accuracy scores on dissimilar and similar trials. These trials consisted of more salient distinctions (see Table 3), and were independent of the variables of interest. Median accuracy 6 on all these trials combined was 94.5%. Seven participants scored exactly at 94.5%. We performed two separate analyses of the data, where the seven participants were either all included in the higher performer group or in the lower performer group, later referred to as Split 1 and Split 2, respectively. The results were equivalent in both cases, but, for simplicity reasons, we only illustrate the Split 1 results. The distribution of participants was fairly equal across

language background (see Table 4). The table also shows scores on dissimilar and similar trials for both higher and lower performers. Both groups were highly accurate on these trials, but there was much more variability among lower performers, as indicated by the higher standard deviations. Figure 4 illustrates the word learning results for length and place

trials split into higher and lower performers. By visual inspection alone, it can be seen that participants in the higher performer group were clearly learning the minimal pair words, as indicated by their much higher levels of accuracy. In the lower performer group, on the other hand, participants’ responses were close to chance. We analyzed these results with models with fixed effects of

Feature Type (length, place), Language (Korean, Mandarin), and Performer Type (high, low), separately for Split 1 and Split 2. In both cases, we found a significant effect of Performer Type (ps .001), reflecting higher accuracy in the higher performer group than in the lower performer group. Critically, there were also significant three-way Feature Type Language Performer Type interactions (in both Split 1 and Split 2; ps.05), indicating distinct response patterns for Korean versus Mandarin speakers on length and place trials depending on their overall success rate in learning, as measured by their accuracy on dissimilar and similar trials. To assess the nature of the three-way interaction, we analyzed

the effects of Feature Type and Language separately for each Performer Type. Higher performers showed a pattern more consistent with taking advantage of their perceptual biases: Korean speakers were more accurate on length trials than Mandarin speakers, but the reverse was true on place trials, as indicated by significant Feature Type Language interactions (in both Split 1

A reviewer pointed out that a median level of performance of 94.5% suggests ceiling effects, such that distinguishing good learners from poor learners ceases to be very meaningful. We agree, in principle, that this is a valid concern, but because this median split revealed a significant and interpretable interaction, we did not think it was a concern in practice.

Dissimilar and Similar Trial Scores (Standard Deviations) From the Word Learning Task Split by Higher and Lower Performers Higher performers Lower performers

and Split 2; ps.05). The lower performers, on the other hand, showed no significant Feature Type Language interactions (and no other significant effects). Therefore, analyzing the word learning data split by accuracy on

$$ 2; p s < . 0 5). ^ {7} $$

more salient distinctions revealed a difference between higher and lower performers, with the former group taking more advantage of their L1-based perceptual abilities than the latter group. What might underlie such individual differences? One possibility is that the lower performers in the word learning task simply had lower discrimination abilities. Due to the between-subjects design, we did not have any direct evidence about how discrimination related to word learning at the level of individual participants. However, we were able to examine a similar split of participants in the discrimination task to see whether the difference between Korean and Mandarin speakers would disappear for lower performers. We found that, with a similar split of participants (see Figure 5), the Feature Type Language interactions were significant for both higher performer and lower performer groups (ps.01). Therefore, while we cannot rule it out, we did not find any evidence to support the hypothesis that overall lower performers might have lower discrimination abilities. Instead, these results suggest that the higher performer advantage in the word learning task was driven by something other than (or perhaps in addition to) better discrimination abilities. We outline possible explanations in the Discussion section. To summarize, in the overall results, there was no effect (inter-

$$ (p\ <<<11) $$

action) of language background in the word learning task, suggesting that discrimination is necessary but not sufficient for successful learning of similar-sounding words. While we found some evidence of language background– based learning of highly similar words in participants who scored better on the learning task overall, the effect was confined to better word learners and far less This document is copyrighted by the American Psychological Association or one of its allied publishers.robust than the effect we saw in the discrimination task. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Discussion

In this article, we asked two specific questions: (1) Do adult L2 learners have difficulty learning similar-sounding words that they can nevertheless discriminate, just like 14-month-old infants?; and (2) How does phonological similarity moderate discrimination versus learning of similar-sounding L2 words? To answer these questions, we asked participants of two different language backgrounds (Korean and Mandarin) either to discriminate similarsounding words in a new language or to map similar-sounding words onto novel referents. That is, we investigated the relation between perceptual discrimination and word learning across speaker populations varying in native-language background. We

used three levels of word similarity: dissimilar word pairs (differing in all sound segments but the last vowel), similar word pairs (same first syllable, but salient differences in the middle consonant), and highly similar word pairs (same first syllable and only subtle differences in the middle consonant). Furthermore, the highly similar word pairs consisted of two distinction types to which the two participant populations had differential perceptual sensitivities (as shown by previous work, and confirmed by our discrimination task). In particular, some distinctions varied along dimensions familiar to participants from their L1, and others did not. Comparing the three levels of word pairs (dissimilar, similar, highly similar) allowed us to assess the role of phonological similarity in discrimination versus word learning (Question 2), while investigating the differences between L1 populations on the two types of highly similar word pairs allowed us to answer the question about the potential group-level mismatch between discrimination and learning of similar-sounding words in adult L2 learning (Question 1). Regarding the effects of phonological similarity on discrimina-

tion versus word learning, we found—as in prior work on learning phonologically native-like words— gradient performance. In both discrimination and word learning tasks, performance was best on dissimilar word pairs, intermediate for similar word pairs, and worst on highly similar word pairs. However, we also found a mismatch between discrimination and learning: the gradient effects were more exaggerated for participants in the word learning task than for participants in the discrimination task, with performance dropping disproportionally as word similarity increased. Therefore, our results suggest that word learning is particularly difficult when words are similar-sounding, even when the differences are fairly salient (as in our similar word pairs). The results from highly similar word pairs revealed that partic-

ipants in the word learning task were unable to use their L1-based perceptual abilities effectively during word learning: Despite Korean- versus Mandarin-speaker differences in discrimination of highly similar word pairs, there were no analogous Korean- versus Mandarin-speaker differences for participants in the word learning task. However, there was some evidence that higher performers (as independently assessed by performance on unrelated trials) were

One might wonder if these results only showed up after better learners were alerted to the nature of the task. However, the interaction was already numerically present (marginal on Split 1, p.08; Split 2, p.11) in the first block of testing. This suggests that the effect of Feature Type was present before test trials “tipped off” participants to the presence of the length and place minimal pairs, because the pictures corresponding to these minimal pairs were never shown together in training.

This document is copyrighted by the American Psychological Association or one of its allied publishers.

Figure 4. Word learning task results (length and place trials) split by higher and lower performers (Split 1; results for Split 2 were equivalent). Accuracy scores indicate proportion of correct responses and error bars are standard errors.

somewhat more successful. This was in sharp contrast to the discrimination results, where L1-based perceptual advantages were observed for all participants. This result thus reveals an intermediate effect between failure to learn similar-sounding words (as observed for 14-month-old infants) and a full ability to use existing L1-based perceptual abilities in learning (which should mimic the discrimination data). What factors led to better performance in learning highly

similar-sounding words? It is possible that the observed differences in the word learning task performance for higher versus lower performers simply arise from better discrimination abilities of the higher performers; that is, higher performers might be better at detecting subtle differences, especially when given the chance to directly compare the highly similar words during test trials. However, our data did not provide any evidence that the lower performers were actually perceptually worse. In the conceptual model we discussed in the Introduction, the difference between higher and lower performers can be explained by higher performers being able to keep track of and/or integrate evidence from more exem-

plars relative to the lower performers. In addition, higher performers might be able to store exemplars better (i.e., with higher phonetic fidelity) than the lower performers from the outset of language exposure, and thus be faster at forming new phonetic categories. This would not hurt lower performers when asked to identify dissimilar words, but would impact them more substantially when identifying label-referent pairings of similar-sounding words. This account means that the observed differences between higher and lower performers can, at least in part, be attributed to individual differences between learners. To investigate this further, we examined more closely the individual measures collected from our participants, as shown in Table 5 (Split 1 only; minor differences between Split 1 and Split 2 are reported in table footnotes). A series of t tests revealed that the higher performer and the

lower performer groups did not significantly differ in nonverbal IQ or L1 proficiency. However, they did differ in English proficiency: higher performers reported higher overall proficiency than lower performers. One possibility, thus, is that better knowledge of English benefitted the higher performers in this task. However, as

Figure 5. Discrimination task results (length and place trials) split by higher and lower performers. Accuracy scores indicate proportion of correct responses and error bars are standard errors.

Table 5 Higher vs. Lower Performers in the Word Learning Task Higher performers

Note. Significant differences in boldface. L1 native language. ^a b On a 0 –10 scale (0 none and 10 perfect). If ^born in the United States, coded as 0.

Effect only marginal in Split 2: t(51.4) 1.74, p ^d .087. Effect not significant in Split 2: t(51.8) 1.36, p.181.

outlined earlier in describing the stimuli, it is not entirely clear what properties of English would produce this type of benefit. It is true that English uses segmental length as a secondary cue to some vowel contrasts (as well as coda voicing), but our stimuli only included length differences for consonants, and prior work (Pajak & Levy, 2014) found no benefit of higher English proficiency on discriminating consonant length contrasts. Another possibility is that the better performance of some participants in the word learning task was an effect of early bilingualism: more balanced bilinguals might be better equipped to use their L1-based perceptual abilities when learning novel words. Indeed, higher English proficiency of the higher performers in our experiment was likely the consequence of their earlier age of arrival to the United States and a longer length of residence relative to the lower performer group. Therefore, the higher performer group might have been composed of early bilinguals who were perhaps more balanced in both languages (i.e., fluent in both Korean/Mandarin and English), while the lower performer group included more speakers strongly dominant in their L1. This is consistent with other findings suggesting that early simultaneous bilingual adults have a general cognitive advantage over monolinguals for novel word learning, whether phonologically familiar or unfamiliar, and independently of phonological memory capacity (Kaushanskaya, 2012; Kaushanskaya & Marian, 2009). Yet another alternative explanation is This document is copyrighted by the American Psychological Association or one of its allied publishers.that higher English proficiency of some participants might be an This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.indicator (more precise than IQ) of better language-learning skills that led to superior performance in our word learning task. However, this explanation seems unlikely given that the differences in self-reported English proficiency between the higher and lower performer groups disappeared after controlling for age of arrival or length of residence in the United States.

Why Is Word Learning Difficult? Irrespective of individual differences, the results we have re-

ported suggest that there is something inherently hard about the early stage of word learning that precludes attention to fine phonetic detail that is otherwise available during phonetic processing—this is true for adult L2 learners in the same way it seems to

be true for 14-month-old L1 learners. Together with the results of White et al. (2013), who showed that infants and adults are similarly affected by lexical familiarity during word learning, these findings provide evidence that there might be common learning mechanisms operating throughout development. But what are these mechanisms, and what is the source of difficulty in learning similar-sounding words? One answer is that learning novel words is simply a highly

complex task that leads to information processing overload. This explanation was originally proposed by Stager and Werker (1997) and Werker et al. (2002), who argued that young infants struggle with accessing sufficient phonetic detail in their lexical representations to successfully differentiate between words that are phonetically highly similar. This means that—relative to task difficulty— only individuals with better attentional or general cognitive abilities might effectively manage concurrent information at multiple levels of processing. This would explain why older, 17- to 20-month-old infants outperform 14-montholds, as well as why adults with some general cognitive advantages (such as early simultaneous bilinguals; e.g., Bialystok, 1999) might outperform other adults, as shown in this study. This would also suggest that encoding and using phonetic detail when learning similar-sounding words would be improved by decreasing memory demands in the word learning task (consistent with prior work; e.g., Fennell & Werker, 2003), as well as providing additional support for storing the phonetic detail of word exemplars and forming new phonetic categories. The latter idea is consistent with recent findings that auditory training on difficult phonetic distinctions improves subsequent word learning, and is particularly beneficial for learners with low pretraining auditory sensitivity (Cooper & Wang, 2013; Ingvalson, Barr, & Wong, 2013). Note, however, that it does not seem to be the case that word

learning is simply overall harder than discrimination— our results indicate that, for highly dissimilar word pairs, performance is equally high in both tasks (although an underlying difference might be hidden due to ceiling effects). Instead, performance in word learning drops disproportionately as the word pairs get phonologically more similar, suggesting that the task is especially taxing when detailed phonetic representations are needed to distinguish between words, even when the words are perceptually easily discriminable (as our similar word pairs). This is consistent with other work showing that confusability between newly learned— but phonologically native-like—words is modulated by the phonetic distance between the sounds that differentiate between them (e.g., Creel & Dahan, 2010; Creel et al., 2006; White et al., 2013). Thus, the overall difficulty of word learning in general lies most likely not just in the task itself, but also in pinning down the correct label/referent pairings, as outlined in the conceptual model we proposed in the Introduction.

Implications for L2 Acquisition The results we have presented contribute to our understand-

ing of L2 learning, as well as provide more practical implications for second-language teaching. The current view is that L2 phonetic category learning is largely hindered due to perceptual difficulties arising from prior acquisition of L1 phonology (e.g., Best, 1995; Flege, 1995; Kuhl & Iverson, 1995). While we do not dispute the importance of L1 influence on L2 phonological acquisition, our results suggest that at least some of the learners’ difficulties in distinguishing among novel L2 sounds might be due to their introduction in the context of highly similarsounding lexical items. In such a context, when attention is directed at forming label-referent mappings, learners might be unable to properly separate similar-sounding categories. In fact, it has been shown that, when discrimination between some L2 sounds is initially present but fragile, the mere act of learning the word–referent mappings for similar-sounding words that differ by those sounds makes their perceptual discrimination even worse (Dobel, Lagemann, & Zwitserlood, 2009). It thus seems that the introduction of highly similar L2 phonetic categories might be more effective either outside of a word learning task (e.g., as a focused nonnative sound discrimination practice) or when learning words that are overall fairly dissimilar. The latter conclusion is independently supported by the results from infant studies, which have shown that prior experience with sounds in nonminimal pair lexical contexts improves infants’ ability to later learn minimal pair words distinguished by those sounds (Thiessen, 2007). A similar conclusion can be drawn from both behavioral and computational modeling work, which has shown that nonminimal pair lexical contexts improve distributional learning of overlapping sound categories This document is copyrighted by the American Psychological Association or one of its allied publishers. for both infants and adults (Feldman, Griffiths, Goldwater, & This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. Morgan, 2013; Feldman, Myers, White, Griffiths, & Morgan, 2013).

Conclusion The results presented in this article show that adults, just like

young infants, have difficulty learning similar-sounding words that they can nevertheless distinguish perceptually, demonstrating that discrimination is necessary but not sufficient for successful learning of similar-sounding words. This parallel between infants and adults points to a common mechanism underlying the initial stage of lexical acquisition throughout development, whether in the native language or any additional language acquired in adulthood.

Together with recent results that have shown other parallels between infant and adult lexical acquisition (White et al., 2013), our findings highlight the necessity for greater interaction between infant and adult language learning literatures that would investigate the commonalities and differences between native-language development in infancy and second-language learning in adulthood, thus shedding more light on the degree of developmental continuity in language learning.

the time course of spoken word recognition using eye movements: Evidence for continuous mapping models. Journal of Memory and Language, 38, 419 – 439. http://dx.doi.org/10.1006/jmla.1997.2558 Andruski, J. E., Blumstein, S. E., & Burton, M. (1994). The effect of subphonetic differences on lexical access. Cognition, 52, 163–187. http://dx.doi.org/10.1016/0010-0277(94)90042-6 Ballem, K. D., & Plunkett, K. (2005). Phonological specificity in children at 1;2. Journal of Child Language, 32, 159 –173. http://dx.doi.org/10 .1017/S0305000904006567 Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68, 255–278. http://dx.doi.org/10.1016/j.jml .2012.11.001 Benus, S., Smorodinsky, I., & Gafos, A. (2003). Gestural coordination and the distribution of English ‘geminates’. University of Pennsylvania Working Papers in Linguistics, 10(1), 4. http://repository.upenn.edu/ pwpl/vol10/iss1/4 Best, C. T. (1995). A direct realist view of cross-language speech perception. In W. Strange (Ed.), Speech perception and linguistic experience: Issues in cross-language research (pp. 171–204). Timonium, MD: York Press. Bialystok, E. (1999). Cognitive complexity and attentional control in the bilingual mind. Child Development, 70, 636 – 644. http://dx.doi.org/10 .1111/1467-8624.00046 Birdsong, D. (2009). Age and the end state of second language acquisition. In W. Ritchie & T. Bhatia (Eds.), The new handbook of second language acquisition (pp. 401– 424). Amsterdam, the Netherlands: Elsevier. Birdsong, D., & Molis, M. (2001). On the evidence for maturational constraints in second-language acquisition. Journal of Memory and Language, 44, 235–249. http://dx.doi.org/10.1006/jmla.2000.2750 Casini, L., Burle, B., & Nguyen, N. (2009). Speech perception engages a general timer: Evidence from a divided attention word identification task. Cognition, 112, 318 –322. http://dx.doi.org/10.1016/j.cognition .2009.04.005 Choi, D.-I. (1995). Korean “tense” consonants as geminates. Kansas Work- ing Papers in Linguistics, 20, 25–38. Clayards, M., Tanenhaus, M. K., Aslin, R. N., & Jacobs, R. A. (2008). Perception of speech reflects optimal use of probabilistic speech cues. *Cognition, 108,*804 – 809. http://dx.doi.org/10.1016/j.cognition.2008.04 .004 Cooper, A., & Wang, Y. (2013). Effects of tone training on Cantonese tone-word learning. The Journal of the Acoustical Society of America, 134(2), EL133–EL139. http://dx.doi.org/10.1121/1.4812435 Creel, S. C., Aslin, R. N., & Tanenhaus, M. K. (2006). Acquiring an artificial lexicon: Segment type and order information in early lexical entries. Journal of Memory and Language, 54, 1–19. http://dx.doi.org/ 10.1016/j.jml.2005.09.003 Creel, S. C., & Dahan, D. (2010). The effect of the temporal structure of spoken words on paired-associate learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36, 110 –122. http://dx .doi.org/10.1037/a0017527

References Allopenna, P. D., Magnuson, J. S., & Tanenhaus, M. K. (1998). Tracking

http://dx.doi.org/10.3758/PBR.16.4.705 .doi.org/10.3758/PBR.16.4.705

http://dx.doi.org/10.1080/01690965.2012.662279 01690965.2012.662279

418 – 435. http://dx.doi.org/10.1121/1.2743154 http://dx.doi.org/10.1121/1.2743154

nium, MD: York Press. http://dx.doi.org/10.1017/S1366728913000436 913000436

http://dx.doi.org/10.1016/j.cognition.2013.02.007 j.cognition.2013.02.007

http://dx.doi.org/10.1080/03640210709336987 03640210709336987

Dobel, C., Lagemann, L., & Zwitserlood, P. (2009). Non-native phonemes in adult word learning: Evidence from the N400m. Philosophical Trans- actions of the Royal Society of London Series B, Biological Sciences, 364, 3697–3709. http://dx.doi.org/10.1098/rstb.2009.0158 Escudero, P., Broersma, M., & Simon, E. (2013). Learning words in a third language: Effects of vowel inventory and language proficiency. Lan- *guage and Cognitive Processes, 28,*746 –761. http://dx.doi.org/10.1080/ 01690965.2012.662279 Escudero, P., Hayes-Harb, R., & Mitterer, H. (2008). Novel secondlanguage words and asymmetric lexical access. Journal of Phonetics, 36, 345–360. http://dx.doi.org/10.1016/j.wocn.2007.11.002 Escudero, P., Simon, E., & Mulak, K. E. (2014). Learning words in a new language: Orthography doesn’t always help. Bilingualism: Language and Cognition, 17, 384 –395. http://dx.doi.org/10.1017/S1366728 913000436 Feldman, N. H., Griffiths, T. L., Goldwater, S., & Morgan, J. L. (2013). A role for the developing lexicon in phonetic category acquisition. Psy- chological Review, 120, 751–778. http://dx.doi.org/10.1037/a0034245 Feldman, N. H., Myers, E. B., White, K. S., Griffiths, T. L., & Morgan, J. L. (2013). Word-level information influences phonetic learning in adults and infants. Cognition, 127, 427– 438. http://dx.doi.org/10.1016/ j.cognition.2013.02.007 Fennell, C. T., Waxman, S. R., & Weisleder, A. (2007). With referential cues, infants successfully use phonetic detail in word learning. In H. Caunt-Nulton, S. Kulatilake, & I. Woo (Eds.), Proceedings of the 31st Boston University Conference on Language Development (pp. 206 – 217). Somerville, MA: Cascadilla Press. Fennell, C. T., & Werker, J. F. (2003). Early word learners’ ability to access phonetic detail in well-known words. Language and Speech, 46, 245–264. http://dx.doi.org/10.1177/00238309030460020901 Flege, J. E. (1995). Second-language speech learning: Theory, findings and problems. In W. Strange (Ed.), Speech perception and linguistic expe- rience: Issues in cross-language research (pp. 229 –273). Timonium, MD: York Press. Hakuta, K., Bialystok, E., & Wiley, E. (2003). Critical evidence: A test of the critical-period hypothesis for second-language acquisition. Psycho- logical Science, 14, 31–38. http://dx.doi.org/10.1111/1467-9280.01415 Hayes-Harb, R., & Masuda, K. (2008). Development of the ability to lexically encode novel second language phonemic contrasts. Second Language Research, 24, 5–33. http://dx.doi.org/10.1177/0267658 307082980 Hillenbrand, J. M., Clark, M. J., & Houde, R. A. (2000). Some effects of duration on vowel recognition. Journal of the Acoustical Society of America, 108, 3013–3022. http://dx.doi.org/10.1121/1.1323463 Ingvalson, E. M., Barr, A. M., & Wong, P. C. M. (2013). Poorer phonetic perceivers show greater benefit in phonetic-phonological speech learning. Journal of Speech, Language, and Hearing Research, 56, 1045– This document is copyrighted by the American Psychological Association or one of its allied publishers. 1050. http://dx.doi.org/10.1044/1092-4388(2012/12-0024) This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. Jaeger, T. F. (2008). Categorical data analysis: Away from ANOVAs (transformation or not) and towards logit mixed models. Journal of Memory and Language, 59, 434 – 446. http://dx.doi.org/10.1016/j.jml .2007.11.007 Johnson, J. S., & Newport, E. L. (1989). Critical period effects in second language learning: The influence of maturational state on the acquisition of English as a second language. Cognitive Psychology, 21, 60 –99. http://dx.doi.org/10.1016/0010-0285(89)90003-0 Kaufman, A. S., & Kaufman, N. L. (2004). Kaufman Brief Intelligence Test, Second Edition. Circle Pines, MN: American Guidance Service. Kaushanskaya, M. (2012). Cognitive mechanisms of word learning in bilingual and monolingual adults: The role of phonological memory. *Bilingualism: Language and Cognition, 15,*470 – 489. http://dx.doi.org/ 10.1017/S1366728911000472

word learning. Psychonomic Bulletin & Review, 16, 705–710. http://dx .doi.org/10.3758/PBR.16.4.705 Kaye, A. (2005). Gemination in English. *English Today, 21,*43–55. http:// dx.doi.org/10.1017/S0266078405002063 Kidd, G. R., Watson, C. S., & Gygi, B. (2007). Individual differences in auditory abilities. The Journal of the Acoustical Society of America, 122, 418 – 435. http://dx.doi.org/10.1121/1.2743154 Kraljic, T., & Samuel, A. G. (2005). Perceptual learning for speech: Is there a return to normal? Cognitive Psychology, 51, 141–178. http://dx .doi.org/10.1016/j.cogpsych.2005.05.001 Kuhl, P. K., & Iverson, P. (1995). Linguistic experience and the “perceptual magnet effect.” In W. Strange (Ed.), Speech perception and linguis- tic experience: Issues in cross-language research (pp. 121–154). Timonium, MD: York Press. Ladefoged, P., & Maddieson, I. (1996). The sounds of the world’s lan- guages. Cambridge, MA: Blackwell. Lenneberg, E. (1967). Biological foundations of language. New York, NY: Wiley. Lin, H. (2001). A grammar of Mandarin Chinese. München, Germany: Lincom Europa. Luce, P. A., & Pisoni, D. B. (1998). Recognizing spoken words: The neighborhood activation model. Ear and Hearing, 19, 1–36. http://dx .doi.org/10.1097/00003446-199802000-00001 Lukas, J. H. (1980). Human auditory attention: The olivocochlear bundle may function as a peripheral filter. Psychophysiology, 17, 444 – 452. http://dx.doi.org/10.1111/j.1469-8986.1980.tb00181.x Magnuson, J. S., Dixon, J. A., Tanenhaus, M. K., & Aslin, R. N. (2007). The dynamics of lexical competition during spoken word recognition. Cognitive Science, 31, 133–156. http://dx.doi.org/10.1080/ 03640210709336987 Mani, N., & Plunkett, K. (2011). Does size matter? Subsegmental cues to vowel mispronunciation detection. *Journal of Child Language, 38,*606 – 627. http://dx.doi.org/10.1017/S0305000910000243 Mattys, S. L., Barden, K., & Samuel, A. G. (2014). Extrinsic cognitive load impairs low-level speech perception. Psychonomic Bulletin & Review, 21, 748 –754. http://dx.doi.org/10.3758/s13423-013-0544-7 Mattys, S. L., & Palmer, S. D. (2015). Divided attention disrupts perceptual encoding during speech recognition. The Journal of the Acoustical Society of America, 137, 1464 –1472. http://dx.doi.org/10.1121/1 .4913507 Maye, J., & Gerken, L. (2000). Learning phonemes without minimal pairs. In S. C. Howell, S. A. Fish, & T. Keith-Lucas (Eds.), Proceedings of the 24th Annual Boston University Conference on Language Development (pp. 522–533). Somerville, MA: Cascadilla Press. Maye, J., Weiss, D. J., & Aslin, R. N. (2008). Statistical phonetic learning in infants: Facilitation and feature generalization. Developmental Sci- ence, 11, 122–134. http://dx.doi.org/10.1111/j.1467-7687.2007.00653.x Maye, J., Werker, J. F., & Gerken, L. (2002). Infant sensitivity to distributional information can affect phonetic discrimination. Cognition, 82, B101–B111. http://dx.doi.org/10.1016/S0010-0277(01)00157-3 Miyawaki, K., Jenkins, J. J., Strange, W., Liberman, A. M., Verbrugge, R., & Fujimura, O. (1975). An effect of linguistic experience: The discrimination of and [I] by native speakers of Japanese and English. Perception & Psychophysics, 18, 331–340. http://dx.doi.org/10.3758/BF03211209 Norris, D., McQueen, J. M., & Cutler, A. (2003). Perceptual learning in speech. Cognitive Psychology, 47, 204 –238. http://dx.doi.org/10.1016/ S0010-0285(03)00006-9 Nowak, P. M. (2006). The role of vowel transitions and frication noise in the perception of Polish sibilants. Journal of Phonetics, 34, 139 –152. http://dx.doi.org/10.1016/j.wocn.2005.03.001 Pajak, B. (2012). Inductive inference in non-native speech processing and learning (Doctoral dissertation, University of California, San Diego). Retrieved from http://escholarship.org/uc/item/6bw4949m#page-1

http://dx.doi.org/10.1121/1.4913507 .4913507

http://dx.doi.org/10.1177/0267658307082980 307082980

http://dx.doi.org/10.1016/j.jml.2007.11.007 .2007.11.007

http://dx.doi.org/10.1016/S0010-0285%2803%2900006-9 S0010-0285(03)00006-9

http://dx.doi.org/10.1017/S1366728911000472 10.1017/S1366728911000472

This document is copyrighted by the American Psychological Association or one of its allied publishers.

http://dx.doi.org/10.1016/j.jml.2008.03.001 dx.doi.org/10.1016/j.jml.2008.03.001

http://dx.doi.org/10.1038/41102 http://dx.doi.org/10.1038/41102

http://dx.doi.org/10.1016/j.jml.2013.01.003 .2013.01.003

http://dx.doi.org/10.1017/S0142716404001109 16404001109

http://dx.doi.org/10.1111/j.1467-7687.2008.00789.x .00789.x

Learning additional languages as hierarchical probabilistic inference: Insights from L1 processing. Language Learning. Pajak, B., & Levy, R. (2011). Phonological generalization from distributional evidence. In L. Carlson, C. Hölscher, & T. Shipley (Eds.), Pro- ceedings of the 33rd Annual Conference of the Cognitive Science Society (pp. 2673–2678). Austin, TX: Cognitive Science Society. Pajak, B., & Levy, R. (2014). The role of abstraction in non-native speech perception. Journal of Phonetics, 46, 147–160. http://dx.doi.org/10 .1016/j.wocn.2014.07.001 Pater, J., Stager, C., & Werker, J. (2004). The perceptual acquisition of phonological contrasts. Language, 80, 384 – 402. http://dx.doi.org/10 .1353/lan.2004.0141 Puel, J. L., Bonfils, P., & Pujol, R. (1988). Selective attention modifies the active micromechanical properties of the cochlea. Brain Research, 447, 380 –383. http://dx.doi.org/10.1016/0006-8993(88)91144-4 Rost, G. C., & McMurray, B. (2009). Speaker variability augments phonological processing in early word learning. Developmental Science, 12, 339 –349. http://dx.doi.org/10.1111/j.1467-7687.2008.00786.x Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274, 1926 –1928. http://dx.doi.org/10 .1126/science.274.5294.1926 Saffran, J. R., Newport, E. L., & Aslin, R. N. (1996). Word segmentation: The role of distributional cues. Journal of Memory and Language, 35, 606 – 621. http://dx.doi.org/10.1006/jmla.1996.0032 Shipley, W. C. (1967). Shipley Institute of Living Scale: Vocabulary Test and Abstraction Test. Los Angeles, CA: Western Psychological Services. Silbert, N. H., Smith, B. K., Jackson, S. R., Campbell, S. G., Hughes, M. M., & Tare, M. (2015). Non-native phonemic discrimination, phonological short term memory, and word learning. Journal of Phonetics, 50, 99 –119. http://dx.doi.org/10.1016/j.wocn.2015.03.001 Sohn, H.-M. (1999). The Korean language. New York, NY: Cambridge University Press. Stager, C. L., & Werker, J. F. (1997). Infants listen for more phonetic detail in speech perception than in word-learning tasks. *Nature, 388,*381–382. http://dx.doi.org/10.1038/41102 Storkel, H. L. (2004). Do children acquire dense neighborhoods? An investigation of similarity neighborhoods in lexical acquisition. Applied Psycholinguistics, 25, 201–221. http://dx.doi.org/10.1017/S01427 16404001109 Storkel, H. L., Armbrüster, J., & Hogan, T. P. (2006). Differentiating phonotactic probability and neighborhood density in adult word learn-

http://dx.doi.org/10.1207/S15327078IN0301_1 S15327078IN0301_1

http://dx.doi.org/10.1002/dev.20060 .1002/dev.20060

http://dx.doi.org/10.1016/S0167-6393%2898%2900058-2 6393(98)00058-2

Pajak, B., Fine, A. B., Kleinschmidt, D. F., & Jaeger, T. F. (in press).

http://dx.doi.org/10.1016/j.wocn.2014.07.001 .1016/j.wocn.2014.07.001

http://dx.doi.org/10.1353/lan.2004.0141 .1353/lan.2004.0141

http://dx.doi.org/10.1126/science.274.5294.1926 .1126/science.274.5294.1926

1192. http://dx.doi.org/10.1044/1092-4388(2006/085) Swingley, D., & Aslin, R. N. (2000). Spoken word recognition and lexical representation in very young children. Cognition, 76, 147–166. http:// dx.doi.org/10.1016/S0010-0277(00)00081-0 Swingley, D., & Aslin, R. N. (2002). Lexical neighborhoods and the word-form representations of 14-month-olds. Psychological Science, 13, 480 – 484. http://dx.doi.org/10.1111/1467-9280.00485 Swingley, D., & Aslin, R. N. (2007). Lexical competition in young children’s word learning. Cognitive Psychology, 54, 99 –132. http://dx.doi .org/10.1016/j.cogpsych.2006.05.001 Thiessen, E. D. (2007). The effect of distributional information on children’s use of phonemic contrasts. Journal of Memory and Language, 56, 16 –34. http://dx.doi.org/10.1016/j.jml.2006.07.002 Vitevitch, M. S., & Luce, P. A. (1998). When words compete: Levels of processing in perception of spoken words. Psychological Science, 9, 325–329. http://dx.doi.org/10.1111/1467-9280.00064 Weber, A., & Cutler, A. (2004). Lexical competition in nonnative spoken word recognition. Journal of Memory and Language, 50, 1–25. http:// dx.doi.org/10.1016/S0749-596X(03)00105-0 Werker, J. F., Fennell, C. T., Corcoran, K. M., & Stager, C. L. (2002). Infants’ ability to learn phonetically similar words: Effects of age and vocabulary size. Infancy, 3, 1–30. http://dx.doi.org/10.1207/ S15327078IN0301_1 Werker, J. F., & Tees, R. C. (2005). Speech perception as a window for understanding plasticity and commitment in language systems of the brain. Developmental Psychobiology, 46, 233–251. http://dx.doi.org/10 .1002/dev.20060 Westbury, J. R., Hashi, M., & Lindstrom, M. J. (1998). Differences among speakers in lingual articulation for American English/ɹ/. Speech Com- munication, 26, 203–226. http://dx.doi.org/10.1016/S0167- 6393(98)00058-2 White, K. S., & Morgan, J. L. (2008). Sub-segmental detail in early lexical representations. *Journal of Memory and Language, 59,*114 –132. http:// dx.doi.org/10.1016/j.jml.2008.03.001 White, K. S., Yee, E., Blumstein, S. E., & Morgan, J. L. (2013). Adults show less sensitivity to phonetic detail in unfamiliar words, too. Journal of Memory and Language, 68, 362–378. http://dx.doi.org/10.1016/j.jml .2013.01.003 Yoshida, K. A., Fennell, C. T., Swingley, D., & Werker, J. F. (2009). Fourteen-month-old infants learn similar-sounding words. Developmen- tal Science, 12, 412– 418. http://dx.doi.org/10.1111/j.1467-7687.2008 .00789.x

This document is copyrighted by the American Psychological Association or one of its allied publishers.

LEARNING SIMILAR-SOUNDING WORDS

Appendix A Individual Measures

Appendix A includes a more detailed version of the individual measures collected from participants that were provided in Table 1. Tables A1 and A2 compare participants of the discrimination

and the word learning tasks separately for each language background (see Table A1 for Korean speakers; see Table A2 for Mandarin speakers). Just as reported in Table 1, there were no significant differences between participants within each L1 population, with one minor exception: Korean speakers in the word learning task reported slightly higher proficiency in understanding English than Korean speakers in the discrimination task.

Tables A3 and A4 compare Korean and Mandarin speakers separately for each task (see Table A3 for discrimination; see Table A4 for word learning). There were no significant differences between Korean and Mandarin speakers assigned to the discrimination task. There were, however, some differences between participants assigned to the word learning task. Namely, relative to speakers of Mandarin, Korean speakers on average immigrated to the United States earlier, had a longer length of residence in the United States, and reported higher proficiency in English and slightly lower proficiency in their L1. At the same time, they did not differ on an objective measure of English proficiency (the

Table A1 Individual Measures: Korean Speakers by Task

^a On a 0-10 scale (0 = none and 10 = perfect). ^b If born in the United States, coded as 0.

(Appendices continue

Table A2 Individual Measures: Mandarin Speakers by Task

Note.L1 native language. ^a On a 0 –10 scale (0 none

and 10 perfect). If born in the United States, coded as 0.

Table A3

Individual Measures: Participants in the Discrimination Task by Language Background L1 Korean L1 Mandarin

Note.L1 native language. ^a On a 0 –10 scale (0 none

and 10 perfect). If born in the United States, coded as 0.

Shipley Vocabulary Test), and all were fluent speakers of both English and Korean/Mandarin, with similar amounts of current exposure to each language. Although it is possible that these differences might have affected our results, prior work with sub- This document is copyrighted by the American Psychological Association or one of its allied publishers.jects from the same populations has shown that, when subjects This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

use both languages regularly, their relative dominance in English versus Korean/Mandarin does not affect the discrimination of the contrasts tested in this study (Pajak & Levy, 2014). Furthermore, and most critically, within each L1 population, participants did not differ across the two tasks, as shown in Tables A1 and A2.

This document is copyrighted by the American Psychological Association or one of its allied publishers.

Table A4 Individual Measures: Participants in the Word Learning Task by Language Background L1 Korean L1 Mandarin

Appendix B

Learning During the Task

In both discrimination and word learning, it is reasonable to expect that performance could improve with exposure to the materials and greater experience in the test task; hence, we analyzed the results including test block as an additional variable (see Figures B1–B4). Throughout the analysis, Block was treated as a mean-centered continuous covariate, which is a simple way of imposing the constraint that temporal changes should be monotonic. Word learning

First, we examined the word learning results for dissimilar similar, and highly similar (Korean: length, Mandarin: place) trials to test whether there was differential improvement over the course of the experiment depending on the word similarity. In particular, we expected relatively stable, good performance on dissimilar trials throughout the experiment, with most dramatic improvements observed for highly similar trials. We tested this by evalu- Word learning

Figure B1. Results for dissimilar, similar, and highly similar (length for Korean, place for Mandarin) trials by block. Accuracy scores indicate proportion of correct responses and error bars are standard errors.

This document is copyrighted by the American Psychological Association or one of its allied publishers.

ating a model with fixed effects of Trial Type (dissimilar, similar highly similar), Language (Korean, Mandarin), and Block (as a continuous predictor). These results are illustrated in Figure B1. Just as in the main analysis, Trial Type was coded with similar trials as the reference level, and there was a significant effect of Trial Type in that accuracy varied in accordance with the similarity between words: The responses on dissimilar trials were signifi-

cantly higher than on similar trials (p.001), which in turn were higher than on highly similar trials (p.001). We found significant main effect of Block (p.001), indicating that participants improved throughout the experiment. There was also significant Trial Type Block interaction (p.05) in that the relative difference between dissimilar and highly similar trials varied as a function of block: The difference was largest at the

$$ (p<.00) $$

$$ (p<.05) $$

Discrimination

Word learning

Figure B2. Results for all highly similar trials by block: length and place. Accuracy scores indicate proportion of correct responses and error bars are standard errors.

This document is copyrighted by the American Psychological Association or one of its allied publishers.

beginning of the experiment, and it gradually decreased with time (no such interaction was found for the difference between similar and highly similar trials). This was in line with our prediction that learners’ performance should not be impeded in cases of learning very dissimilar words;

however, learning was expected to be slower for highly similar words because more data would be needed before learners could accumulate a sufficient number of exemplars to learn phonetic category distinctions and form correct label/referent mappings.

Higher performers

Higher performers

Lower performers

Lower performers

Figure B3. Word learning task results by block (length and place trials) split by higher and lower performers (Split 1; results for Split 2 were equivalent). Accuracy scores indicate proportion of correct responses and error bars are standard errors.

PAJAK, CREEL, AND LEVY

Higher performers Higher performers

length feature place feature

1.0 1.0

Lower performers Lower performers

length feature place feature

1.0 1.0 0.9 0.9 0.8 0.8 Accuracy Accuracy

0.7 0.7 Korean Korean

0.6 0.6 Mandarin Mandarin

0.5 0.5 Test1 Test2 Test3 Test4 Te s t 1 Te s t 2 Te s t 3 Te s t 4 Block Block Figure B4. Discrimination task results by block (length and place trials) split by higher and lower performers (Split 1; results for Split 2 were equivalent). Accuracy scores indicate proportion of correct responses and error bars are standard errors.

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. (Appendices continue)

Next, we analyzed both discrimination and word learning sets of highly similar trials (see Figure B2). For both tasks, we found significant main effects of Block (discrimination: p.05; word learning: p.001), indicating that participants improved throughout the experiment. There were also significant Feature Type Block interactions (discrimination: p.05; word learning: p .01); in word learning, the improvement was more prominent for the length trials than for the place trials; the opposite seemed to be the case for discrimination—more improvement on place than on length trials. Crucially, adding the block information revealed that

$$ p<.05; $$

$$ p<.001) $$

$$ p<.05. $$

$$ p< $$

the main result—a difference between Korean and Mandarin speakers in discrimination, but not in word learning—was consistent throughout the experiment. Figures B3 and B4 illustrate the by-block results split by higher and lower performers in the word learning and the discrimination tasks, respectively.

Accepted December 22, 2015

Received July 3, 2014 Revision received November 28, 2015

Members of Underrepresented Groups: Reviewers for Journal Manuscripts Wanted interested in reviewing manuscripts for APA journals, the

If you are APA Publications and Communications Board would like to invite your participation. Manuscript reviewers are vital to the publications process. As a reviewer, you would gain valuable experience in publishing. The P&C Board is particularly interested in encouraging members of underrepresented groups to participate more in this process. If you are interested in reviewing manuscripts, please write APA Journals at Reviewers@apa.org.

Please note the following important points: • To be selected as a reviewer, you must have published articles in peer-reviewed journals. The

experience of publishing provides a reviewer with the basis for preparing a thorough, objective review. • To be selected, it is critical to be a regular reader of the five to six empirical journals that are most

central to the area or journal for which you would like to review. Current knowledge of recently published research provides a reviewer with the knowledge base to evaluate a new submission within the context of existing research. • To select the appropriate reviewers for each manuscript, the editor needs detailed information.

Please include with your letter your vita. In the letter, please identify which APA journal(s) you are interested in, and describe your area of expertise. Be as specific as possible. For example, “social psychology” is not sufficient—you would need to specify “social cognition” or “attitude change” as well. • Reviewing a manuscript takes time (1– 4 hours per manuscript reviewed). If you are selected to

This document is copyrighted by the American Psychological Association or one of its allied publishers. review a manuscript, be prepared to invest the necessary time to evaluate the manuscript This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. thoroughly. APA now has an online video course that provides guidance in reviewing manuscripts. To learn

more about the course and to access the video, visit http://www.apa.org/pubs/authors/reviewmanuscript-ce-video.aspx.