Summary needs 600 words 1.General research question 2.Experiments 1 & 2: methods and findings ( actual data should be included

Isolated words enhance statistical language learning in infancy

Casey Lew-Williams1, Bruna Pelucchi1,2, and Jenny R. Saffran1 1Department of Psychology and Waisman Center, University of Wisconsin-Madison, USA 2Department of Evolutionary Biology, University of Ferrara, Italy

Abstract Infants are adept at tracking statistical regularities to identify word boundaries in pause-free speech. However, researchers have questioned the relevance of statistical learning mechanisms to language acquisition, since previous studies have used simplified artificial languages that ignore the variability of real language input. The experiments reported here embraced a key dimension of variability in infant-directed speech. English-learning infants (8–10 months) listened briefly to natural Italian speech that contained either fluent speech only or a combination of fluent speech and single-word utterances. Listening times revealed successful learning of the statistical properties of target words only when words appeared both in fluent speech and in isolation; brief exposure to fluent speech alone was not sufficient to facilitate detection of the words’ statistical properties. This investigation suggests that statistical learning mechanisms actually benefit from variability in utterance length, and provides the first evidence that isolated words and longer utterances act in concert to support infant word segmentation.

Introduction For 15 years, infant statistical learning has been a focus of research in the area of cognitive development. Many researchers have shown that infants have a prodigious and domain- general ability to track statistical patterns in their environments (e.g. Aslin, Saffran & Newport, 1998; Fiser & Aslin, 2002; Kirkham, Slemmer & Johnson, 2002; Pelucchi, Hay & Saffran, 2009; Saffran, Aslin & Newport, 1996). Given sequential events X and Y, infants are highly sensitive to the probability that event X will transition to event Y (and vice versa). This sensitivity to transitional probabilities (TPs) has been robustly demonstrated in several auditory and visual domains. For example, when exposed to continuous speech with no pause-defined cues to word boundaries, infants can segment words by noting which syllables occur together with the most consistency.

A majority of these studies have assessed learning using highly simplified artificial languages. While the materials used in statistical language studies typically afford a great deal of experimental control, they fail to capture many of the dimensions of variability that characterize infant-directed speech (though see Pelucchi et al., 2009; Singh, Nestor & Bortfeld, 2008; Thiessen, Hill & Saffran, 2005). This has led several researchers to suggest that statistical learning mechanisms cannot handle the complexities of real language, and to question the hypothesis that statistical learning launches infants into language acquisition (Johnson & Tyler, 2010; Yang, 2004). Here, we aimed to capture a significant source of variability in infant-directed speech – the presence of both isolated words and longer

© 2011 Blackwell Publishing Ltd. Address for correspondence: Casey Lew-Williams, Waisman Center, Infant Learning Lab, 1500 Highland Ave., Room 501, University of Wisconsin-Madison, Madison, WI 53705-2280, USA; [email protected]

NIH Public Access Author Manuscript Dev Sci. Author manuscript; available in PMC 2012 February 16.

Published in final edited form as: Dev Sci. 2011 November ; 14(6): 1323–1329. doi:10.1111/j.1467-7687.2011.01079.x.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

utterances – to assess the degree to which statistical learning abilities are well matched to the structure of early language input.

Natural speech is replete with pauses, which in some cases provide valuable cues to the locations of word boundaries. Corpus analyses show that while caregivers often produce long, uninterrupted strings of words, they also produce much shorter utterances, such as words in isolation. These single-word utterances, which by definition have pauses at their edges, are a salient feature of infant-directed speech. Brent and Siskind (2001) found that 9% of caregiver utterances consisted of isolated words, of which 27% occurred two or more times in neighboring utterances (e.g. See the doggie over there? Doggie!). Fernald and Morikawa (1993) found similar proportions: an average of 9% of utterances spoken by American and Japanese mothers were single content words. Other researchers have estimated similar or even higher proportions of isolated words (Aslin, Woodward, LaMendola & Bever, 1996; Siskind, 1996; van de Weijer, 1998). Researchers disagree about whether this is a substantial proportion of early language input. But viewed through a wide lens, the presence of isolated words seems more impressive: by combining estimates of total language exposure by Hart and Risley (1995) and of isolated word exposure by Brent and Siskind (2001), we can estimate that a child may hear up to 4 million isolated word tokens by their fourth birthday.

Given these numbers, isolated words could be expected to play a role in early language acquisition. At the perceptual level, many studies have demonstrated that after hearing words in sentences, infants can recognize them when they appear in isolation (e.g. Jusczyk & Aslin, 1995; Jusczyk & Hohne, 1997; Pelucchi et al., 2009). Conversely, after hearing words in isolation, infants can recognize them when they appear within sentences (e.g. Gout, Christophe & Morgan, 2004; Houston & Jusczyk, 2000; Jusczyk & Aslin, 1995). Of particular interest are findings demonstrating that caregivers’ use of isolated words is linked to later vocabulary development. Brent and Siskind (2001) found that 75% of 18-month- olds’ first words had previously appeared in isolation in mothers’ speech, and moreover, that the frequency of hearing a word in isolation (but not the frequency of hearing a word overall) was a unique predictor of later word use (see also Ninio, 1992).

Given the presence of isolated words in infant-directed speech, what is their role in statistical learning? To date, we do not know whether these acoustically demarcated units have any effect on infants’ computations of statistically coherent syllable sequences, or more broadly speaking, whether statistical learning mechanisms can help infants make sense of natural speech that contains a mix of isolated words and longer utterances. Do isolated words offer a shortcut for word segmentation, such that infants do not need to track TPs in natural speech? Alternatively, perhaps the presence of isolated words facilitates the detection of TPs in fluent speech.

To address this issue, we tested 8- to 10-month-old English-learning infants’ abilities to segment words from natural Italian speech that either did or did not contain isolated words. In Experiment 1, infants were tested in one of two conditions. In the Fluent Speech condition, infants heard a set of Italian sentences abbreviated from the materials used in a previous study of statistical learning (Pelucchi et al., 2009, Experiment 3). Two types of target words were embedded within the sentences: high-TP words (TP = 1.0), where the component syllables never appeared elsewhere in the corpus, and low-TP words (TP = 0.33), where the component syllables recurred within many other words in the corpus. This brief, pause-free familiarization was designed to minimize the likelihood that infants would be able to learn enough about high- vs. low-TP words. In the Fluent Speech + Isolated Words condition, infants heard the same target words within sentences, and also heard the target words spoken in isolation between sentences. Critically, this condition was designed

Lew-Williams et al. Page 2

Dev Sci. Author manuscript; available in PMC 2012 February 16.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

such that successful learning of the words’ statistical properties required infants to integrate their experience hearing fluent speech and isolated words.

After familiarization with the Italian speech, infants were tested on their listening times to high- and low-TP words using the Headturn Preference Procedure. If infants cannot detect the words’ statistical coherence based on a very short exposure to fluent speech alone, then we would expect listening times to be equivalent for both types of target words in the Fluent Speech condition. In the Fluent Speech + Isolated Words condition, the same results would emerge if infants relied on isolated words alone, since all target words appeared in isolation with the same frequency. However, we predicted an interaction between condition and the type of target word. If infants can integrate statistical information gained from hearing words in isolation and in fluent speech, then we would expect listening times to differ for high- and low-TP words in the Fluent Speech + Isolated Words condition.

Experiment 1 Method

Participants—Participants were 40 healthy, full-term infants with a mean age of 9.5 months (range = 8.3–10.9). Infants were from monolingual English-speaking households, and had no prior exposure to Italian or Spanish. Parents reported no pervasive developmental delays and no history of hearing or visual problems. Infants were randomly assigned to the Fluent Speech condition (10 female, 10 male) or the Fluent Speech + Isolated Words condition (10 female, 10 male). Eleven additional infants were tested but not included in analyses due to fussiness (10) or inattentiveness (1).

Speech stimuli—Infants were familiarized with natural, infant-directed Italian speech that either did or did not contain isolated tokens of target words (see Figure 1). In the Fluent Speech condition, infants heard grammatically correct and semantically meaningful sentences adapted from a previous study of statistical learning (Pelucchi et al., 2009, Experiment 3). Sentences were recorded by a female native speaker of Italian at a rate of approximately six syllables per second, presented at approximately 60 dBSPL. Embedded within the sentences were four bisyllabic target words: casa, bici, fuga, and melo (English translations: house, bike, escape, and apple tree). The target words appeared with equal frequency during familiarization, but their internal TPs differed. For half of the infants in this condition, casa and bici were high-TP words: their component syllables (ca, sa, bi, and ci) always appeared in the contexts of the words casa and bici. Thus the internal TPs of these words were 1.0. For these infants, fuga and melo were low-TP words: the initial syllables (fu and me) appeared 33% of the time in the words fuga and melo, but 67% of the time else-where in the speech (always in stressed position). Thus the internal TPs of fuga and melo were 0.33. The other half of the infants heard a different, counterbalanced familiarization, with casa and bici as low-TP words and fuga and melo as high-TP words.

In Pelucchi et al. (2009), infants successfully discriminated high-TP from low-TP words after receiving 18 exposures to each target word. Importantly, in the present experiment, infants received just 12 exposures to each target word. This reduction was designed to decrease the likelihood that infants would successfully segment target words when hearing only fluent speech.

In the Fluent Speech + Isolated Words condition, infants received only six exposures to each target word within sentences, with 12 additional occurrences of the initial syllables of low- TP words. Each target word was also presented three times in isolation, yielding a total of nine exposures to each target word. Additional filler sentences were interspersed in order to match familiarization length across conditions. In order to mimic caregiver use of isolated

Lew-Williams et al. Page 3

Dev Sci. Author manuscript; available in PMC 2012 February 16.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

words (Brent & Siskind, 2001), each isolated word appeared between sentences, bordered by 300-ms pauses, and always followed a sentence that contained the same target word. Test items for both conditions consisted of the four target words (two high-TP, two low-TP); the tokens were identical to the single-word utterances used in the Fluent Speech + Isolated Words condition.

Procedure—The experiment consisted of a familiarization phase (90 sec) and a test phase (approximately 3 min), using the Headturn Preference Procedure. Infants were seated on a caregiver’s lap within a double-walled, sound-attenuated booth, equipped with a center light and two side-lights. Infants heard speech stimuli while parents listened to unrelated speech over headphones. The experimenter coded infant head-turns outside of the booth via a closed-circuit camera using a button-box connected to a computer. During familiarization, infants heard Italian speech from laterally placed speakers. The lights flashed contingently on infants’ looking behavior. Each test trial was initiated when infants looked to a flashing center light, at which point the experimenter signaled the center light to extinguish and one of the side-lights to begin flashing. When infants looked at the side-light, an isolated token of a high- or low-TP word was repeated until infants looked away for 2 s or 20 s had elapsed. Side of stimulus presentation was counterbalanced, and the experimenter was blind to test-item presentation. Infants’ looking time was calculated automatically using custom- designed software. During the test, each high- and low-TP word was repeated on three trials, randomized by block, for a total of 12 test trials.

Results and discussion Looking times were analyzed in a 2 × 2 mixed ANOVA with condition (Fluent Speech, Fluent Speech + Isolated Words) as a between-subjects factor, and test item (high-TP, low- TP) as a within-subjects factor (see Figure 2).1 Looking times revealed a significant main effect of test item, F(1, 38) = 4.3, p < .05, ηp2 = .10, suggesting that infants listened significantly longer to high-TP words than to low-TP words. Analyses also showed a significant interaction between condition and test item, F(1, 38) = 7.8, p < .01, ηp2 = .17. Bonferroni-corrected post-hoc comparisons (p < .025) revealed that infants in the Fluent Speech condition listened equally to high- and low-TP words, t(19) = .5, p = .6, d = .09, but that infants in the Fluent Speech + Isolated Words condition listened significantly longer to high-TP words, t(19) = 3.2, p < .005, d = .53. Given this very brief exposure, infants failed to learn enough about the words’ statistical properties to successfully discriminate the test items in the absence of isolated words. However, the presence of single-word utterances facilitated infants’ detection of the statistical properties of those words in fluent speech.

Infants’ successful discrimination of high- and low-TP words in the Fluent Speech + Isolated Words condition is particularly striking given the frequency with which infants were exposed to target words across conditions. Infants in this condition received just nine exposures to each target word, while infants in the Fluent Speech condition received 12 exposures. Even with fewer opportunities to track statistical properties of Italian speech, infants integrated information gained from hearing words in isolation with statistical information available in fluent speech. Critically, infants heard high- and low-TP words equally often in isolation. If they had attended only to isolated words, we would expect no difference in looking to each word type, since all target words would have an identical TP of 1.0. Thus, attention to the statistical coherence of syllables was required for infants to successfully discriminate high- from low-TP words.

1In both Experiments 1 and 2, performance did not vary by gender and did not differ across the two counterbalanced corpora. Thus, these factors were not included in the main analyses.

Lew-Williams et al. Page 4

Dev Sci. Author manuscript; available in PMC 2012 February 16.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

An alternative explanation for these findings is that the presence of isolated words simply made the speech more engaging. We know that infants prefer listening to speech with variability in pitch (Cooper & Aslin, 1990; Fernald & Kuhl, 1987), and that pitch variability heightens infants’ ability to track sequential statistics (Thiessen et al., 2005). It is thus possible that variability in utterance length heightened infants’ interest in listening to the Italian familiarization. This may have resulted in an increased ability to detect statistical properties of target words from fluent speech alone, regardless of the contribution of isolated words. To test this alternative hypothesis, infants in Experiment 2 listened to a combination of fluent Italian speech and isolated words, but there was no phonological overlap between the words presented in fluent speech and isolation. If variability alone enhances performance, then infants should continue to show successful discrimination in Experiment 2. However, if the crucial factor is the phonological overlap between the isolated words and the fluent speech, then infants should fail to discriminate between the test items in Experiment 2.

Experiment 2 Method

Participants—Participants were 20 healthy, full-term infants (10 female) with a mean age of 9.4 months (range = 8.4–10.8). Six additional infants were tested but not included in analyses due to fussiness (4) or inattentiveness (2).

Speech stimuli and procedure—Materials and procedures were comparable to the Fluent Speech + Isolated Words condition from Experiment 1. In both experiments, there were nine exposures to the target words casa, bici, fuga, and melo during the familiarization (90 s). In the previous experiment, infants heard each target word six times in fluent speech and three times in isolation; however, in Experiment 2, all nine exposures were in fluent speech, and infants heard different Italian words in isolation: pane, dono, tema, and socio. Like the target words, these unrelated isolated words were bisyllabic and conformed to a strong–weak stress pattern. Crucially, however, these isolated words did not overlap with the target words.

Results and discussion Listening times in Experiment 2 were compared to listening times in the Fluent Speech + Isolated Words condition from Experiment 1 in a 2 (condition) × 2 (test item) mixed ANOVA. Analyses revealed a significant interaction, F(1, 38) = 5.6, p < .025, ηp2 = .13, suggesting that listening times to high- and low-TP words differed in the two conditions. A paired-samples t-test showed that infants did not show a significant difference in listening times to high- and low-TP words in Experiment 2, t(19) = .3, p = .8, d = .07 (see Figure 2). The lack of discrimination suggests that isolated words do not simply enhance attention to the speech stream. Rather, caregiver use of isolated words seems to play a role in helping infants detect statistical regularities related to those words in otherwise pause-free speech.

General discussion The usefulness of statistical learning mechanisms in explaining early language learning is dependent upon infants’ abilities to organize the variability inherent in real speech. We focused here on a particularly salient source of variability in caregiver speech – the presence of both isolated words and longer utterances – and provided the first evidence that isolated words and longer utterances act in concert to facilitate infant word segmentation. In two preferential-listening experiments, English-learning infants received brief exposure to Italian. When hearing high- and low-TP target words in a brief corpus of fluent speech

Lew-Williams et al. Page 5

Dev Sci. Author manuscript; available in PMC 2012 February 16.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

alone, infants did not demonstrate that they learned enough about the words’ statistical properties to distinguish between them. However, infants who heard a similarly brief corpus containing target words both in fluent speech and in isolation successfully detected the coherence of their component syllables. This finding indicates that statistical learning mechanisms may in fact benefit from variability in utterance length. The only way that infants could have discriminated high- from low-TP words in this condition was through the interaction of fluent speech and isolated words.

The industry of studies on statistical language learning was built on the premise that pauses do not reliably occur between words, posing a difficult puzzle of word segmentation to young language learners. Previous research shows that infants are highly adept at discovering statistical regularities in pause-free speech (Saffran et al., 1996), and that embedding words in sentences has many other benefits in early language learning, by helping infants recognize words faster (Fernald & Hurtado, 2006), anticipate subsequent words (Lew-Williams & Fernald, 2007), discover grammatical categories (Mintz, 2003), and produce multiword speech (Bannard & Matthews, 2008).

Isolated words – which by definition are acoustically demarcated from other utterances by moments of silence – may also have positive consequences for language learning (Peters, 1983; Pinker, 1984). Using computational or adult experimental data, previous researchers have demonstrated that isolated words facilitate the discovery of word boundaries in fluent speech (Brent & Cartwright, 1996; Conway, Bauerschmidt, Huang & Pisoni, 2009; Dahan & Brent, 1999; van de Weijer, 1998). Descriptions of this process converge on a common premise: words spoken in isolation become perceptually familiar, ‘pop out’ when heard in fluent speech, and provide a means of segmenting adjacent words in fluent speech. Similarly, infant studies suggest that highly familiar words, such as the infant’s name or the family’s word for mother, provide useful cues to the boundaries of the words adjacent to them (Bortfeld, Morgan, Golinkoff & Rathbun, 2005), and even help infants in challenging artificial language tasks (Mersad & Nazzi, 2010). However, this hypothesis has not yet been tested using actual isolated words with infant learners.

The experiments reported here offer a new perspective on how a combination of isolated words and longer utterances may support early word segmentation. Words heard in isolation are relatively familiar and recognizable when heard in fluent speech, and they enhance infants’ abilities to detect the statistical properties of syllable sequences in the global language input. Importantly, isolated words are not likely to be essential for language learning, and they do not simply replace the need for infants to compute the locations of word boundaries in fluent speech. But isolated words are a prevalent feature of caregiver speech, and even a minimal number of exposures to words in isolation may confer advantages in early statistical learning. This research thus complements and extends previous work on the importance of infant-directed speech (Brent & Siskind, 2001; Gleitman, Newport & Gleitman, 1984; Kemler-Nelson, Hirsh-Pasek, Jusczyk & Cassidy, 1989; Singh, Nestor, Parikh & Yull, 2009; Thiessen et al., 2005; Werker, Pons, Dietrich, Kajikawa, Fais & Amano, 2007).

Isolated words appear to facilitate learning about statistical attributes of fluent speech. However, both long sentences and isolated words are very much part of the statistical landscape of speech, and their relationship is likely to be symbiotic: the distributions of syllables in fluent speech could also help infants learn about the isolated words themselves. While infants may treat a novel isolated word like casa as a single, unsegmented unit, they may also oversegment, treating casa as two monosyllabic words. Recent computational models of word segmentation have focused on this issue, examining the initial assumptions that learners could make about the nature of words (Goldwater, Griffiths & Johnson, 2009).

Lew-Williams et al. Page 6

Dev Sci. Author manuscript; available in PMC 2012 February 16.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

Generally, ideal learners tend to favor the most efficient segmentation strategies given the input data, which in this case would be treating casa as a single word. But statistics are needed to guide learners toward this correct interpretation. Over time, infants could gradually determine that the component syllables ca and sa should be more tightly linked by noting the regularity with which those syllables co-occur in the language environment. Thus, while isolated words support statistical learning in long utterances, statistical features of long utterances also help infants learn about the structure of isolated words.

While isolated words are particularly salient in the speech stream, they are just one in a network of useful segmentation cues that may inform each other multidirectionally. It will be important to discover how isolated words work together with other linguistic and domain- general cues to signal the locations of word boundaries, such as lexical stress (Curtin, Mintz & Christiansen, 2005; Johnson & Jusczyk, 2001), phrase-level prosody (Shukla, Nespor & Mehler, 2007), sentence position (Seidl & Johnson, 2006), language-specific phonotactic constraints (Brent & Cartwright, 1996), and concurrently presented visual information (Cunillera, CQmara, Laine & Rodríguez-Fornells, 2010). Future studies should also aim to understand how statistical learning mechanisms operate given other dimensions of variability in infant-directed speech, such as the presence of multiple talkers in complex turn-taking interactions. In some cases, as in the research on utterance length reported here, variability does not pose a problem to the young learner, but in other cases, variability may disrupt learning. Studies of the variability inherent in infant-directed speech will both enhance and constrain accounts of early language acquisition, by allowing us to determine the degree to which infant learning mechanisms are well matched to the structure of early language input.

Acknowledgments We would like to thank the participating families, the members of the Infant Learning Lab, and especially Rosie Maier for her help with data collection. We also thank Michael Frank, Erik Thiessen, and anonymous reviewers for comments on a previous version of this paper. This research was funded by grants from the Eunice Kennedy Shriver National Institute of Child Health and Human Development to CLW (F32HD069094), JRS (R01HD0 37466), and the Waisman Center (P30HD03352). Additional funding was provided by a grant from the James F. McDonnell Foundation to JRS, and an NICHD supplementary grant made possible by the American Recovery and Reinvestment Act of 2009 (R01HD037466-09S1).

References Aslin RN, Saffran JR, Newport EL. Computation of conditional probability statistics by 8-month-old

infants. Psychological Science. 1998; 9:321–324. Aslin, RN.; Woodward, JZ.; LaMendola, NP.; Bever, TG. Models of word segmentation in fluent

maternal speech to infants. In: Demuth, K.; Morgan, JL., editors. Signal to syntax. Mahwah, NJ: Erlbaum; 1996. p. 117-134.

Bannard C, Matthews D. Stored word sequences in language learning: the effect of familiarity on children’s repetition of four-word combinations. Psychological Science. 2008; 19:241–248. [PubMed: 18315796]

Bortfeld H, Morgan JL, Golinkoff RM, Rathbun K. Mommy and me: familiar names help launch babies into speech-stream segmentation. Psychological Science. 2005; 16:298–304. [PubMed: 15828977]

Brent MR, Cartwright TA. Distributional regularity and phonotactic constraints are useful for segmentation. Cognition. 1996; 61:93–125. [PubMed: 8990969]

Brent MR, Siskind JM. The role of exposure to isolated words in early vocabulary development. Cognition. 2001; 81:B33–B44. [PubMed: 11376642]

Conway CM, Bauerschmidt A, Huang SS, Pisoni DB. Implicit statistical learning in language processing: word predictability is the key. Cognition. 2009; 114:356–371. [PubMed: 19922909]

Lew-Williams et al. Page 7

Dev Sci. Author manuscript; available in PMC 2012 February 16.

N IH

-P A

A uthor M

anuscript N

IH -P

A A

uthor M anuscript

N IH

-P A

A uthor M

anuscript

Cooper RP, Aslin RN. Preference for infant-directed speech within the first month after birth. Child Development. 1990; 61:1584–1595. [PubMed: 2245748]

Cunillera T, Cámara E, Laine M, Rodríguez-Fornells A. Speech segmentation is facilitated by visual cues. Quarterly Journal of Experimental Psychology. 2010; 63:260–274.

Curtin S, Mintz TH, Christiansen MH. Stress changes the representational landscape: evidence from word segmentation. Cognition. 2005; 96:233–262. [PubMed: 15996560]

Dahan D, Brent MR. On the discovery of novel wordlike units from utterances: an artificial-language study with implications for …