Elsevier

Cognitive Psychology

Volume 67, Issue 3, November 2013, Pages 130-150
Cognitive Psychology

Words and possible words in early language acquisition

https://doi.org/10.1016/j.cogpsych.2013.08.001Get rights and content

Highlights

  • 18-month-olds extract words and possible words from the same speech input.

  • 12-month-olds can better extract possible words than statistically coherent units.

  • Sensitivity to possible words is more primitive than the ability to find actual words.

  • Mechanisms for structure extraction are fundamental for language acquisition.

  • The ability to compute statistical relations is fully effective relatively late in development.

Abstract

In order to acquire language, infants must extract its building blocks—words—and master the rules governing their legal combinations from speech. These two problems are not independent, however: words also have internal structure. Thus, infants must extract two kinds of information from the same speech input. They must find the actual words of their language. Furthermore, they must identify its possible words, that is, the sequences of sounds that, being morphologically well formed, could be words. Here, we show that infants’ sensitivity to possible words appears to be more primitive and fundamental than their ability to find actual words. We expose 12- and 18-month-old infants to an artificial language containing a conflict between statistically coherent and structurally coherent items. We show that 18-month-olds can extract possible words when the familiarization stream contains marks of segmentation, but cannot do so when the stream is continuous. Yet, they can find actual words from a continuous stream by computing statistical relationships among syllables. By contrast, 12-month-olds can find possible words when familiarized with a segmented stream, but seem unable to extract statistically coherent items from a continuous stream that contains minimal conflicts between statistical and structural information. These results suggest that sensitivity to word structure is in place earlier than the ability to analyze distributional information. The ability to compute nontrivial statistical relationships becomes fully effective relatively late in development, when infants have already acquired a considerable amount of linguistic knowledge. Thus, mechanisms for structure extraction that do not rely on extensive sampling of the input are likely to have a much larger role in language acquisition than general-purpose statistical abilities.

Introduction

To master a language, a learner must acquire both the extended network of words composing its lexicon and the complex web of structural relationships that make the lexicon productive, allowing the speaker to enjoy the infinite creativity unique to a human language (Chomsky, 1957). All children throughout the world accomplish these tasks in an incredibly short period of time, considering the complexity of a natural language. Their database is made up of samples of speech uttered under many different conditions, by many different speakers, with different intentions, purposes, and means of expression. Yet they converge to a final stable state within a few years. Thus, they must possess efficient means to extract the words of their language and their structural relationships from these speech snippets. What such means are, and what the relative importance of each of them is, are matters of intensive research.

Several layers of complexity make these tasks difficult. Speech contains no obvious marks of segmentation between words. So in order to find words, infants must break continuous speech into segments that will eventually become words in their languages. An important contribution towards the solution of this problem may come from infants’ proven ability to extract some statistical relationships from speech. Notably, at 8 months of age, infants can spot words on the basis of their absolute frequency of occurrence (Jusczyk & Aslin, 1995), an ability that could help them mark some fixed points to break into continuous speech (Lew-Williams, Pelucchi, & Saffran, 2011). At around the same age or earlier (Thiessen & Erickson, 2013), they can compute the conditional frequencies with which syllables immediately follow each other (or their adjacent transitional probability; hereafter, TPs) to begin segmenting an artificially created continuous stream into word-like units (Aslin et al., 1998, Saffran et al., 1996). They can also exploit such distributional information when exposed to real speech (Pelucchi, Hay, & Saffran, 2009), suggesting that TP computations might be involved in real-world language acquisition (Lany & Saffran, 2010). In spite of these suggestive findings, we still do not know how strong infants’ sensitivity to TPs is and whether it is strong enough to exploit real statistical differences in real languages, which are much less marked than those tested in the laboratory (Yang, 2004). Nor do we know whether sensitivity to such measures occurs at the right time in development. Undoubtedly, adults’ finesse at detecting statistical relationships is exquisite (Vouloumanos, 2008). Timing is crucial in development, however: the acute sensitivity to small differences in statistical relationships needed to segment real speech is of little use for language acquisition if it comes when the problem that it should help solve—word segmentation—is already well on its way towards a solution.

Importantly, word segmentation is only a small step towards the construction of a lexicon. Words are not just bare syllable chunks. They are often composed of roots with added prefixes, as the word prefix exemplifies, suffixes, as suffixes exemplifies, or even simultaneous prefixes and suffixes, as downhearted or embolden exemplify. In short, words have internal structure, defined among adjacent and nonadjacent subparts. Their combinations define what is in the lexicon, but also what could be in the lexicon. To put the issue more abstractly, actual words are fundamental for the knowledge of a language, but so is the notion of a possible word—the general scaffolding of lexical items that contributes to language productivity (Pinker, 1991).1

Already at her very first steps into language, the learner faces the problem of finding both actual and possible words inextricably intermingled. However, to date, we know little about how and when infants form the notion of a possible word, nor whether they do it with the same kind of computational resources allegedly exploited to split the continuum into its composing segments. We know that at the very early stages of their production, children show a mastery of relatively abstract syntactic and morphosyntactic properties (Guasti, 2002, Lidz et al., 2003), not only when they generate words that actually exist in their languages, but also when they make mistakes producing nonexistent, but legal words (e.g., Legate & Yang, 2007). We do not know, however, how such productive knowledge of possible words appears nor how it relates to the acquisition of actual words. We do know that very early in development infants can extract rules in different situations, but the role of this ability in the acquisition of linguistic competences in language acquisition is not well understood (Gerken, 2006, Gerken and Bollt, 2008, Gómez and Gerken, 1999, Johnson et al., 2009, Marchetto and Bonatti, in preparation, Marcus et al., 1999).

Much research in language acquisition holds or implies the view that rules in language arise from an inductive process requiring a considerable amount of vocabulary and of variability within vocabulary (e.g., Altmann, 2002, Bates and Elman, 1996, Christiansen et al., 2009, Gómez, 2002, Gómez and Maye, 2005, Onnis et al., 2008, Reali and Christiansen, 2005, Seidenberg, 1997). Infants first acquire many similar words, and from that database they extract information about their structure (Gómez, 2002, Gómez and Maye, 2005, Onnis et al., 2008). For reference, we will call this picture the empiricist model of language acquisition. According to this view, in principle, language acquisition is no different from other processes of statistical learning: learners only need data and some general-purpose mechanism of data analysis in order to acquire language (e.g., Altmann, 2002, Bates and Elman, 1996, Christiansen et al., 2009, Gómez, 2002, Gómez and Maye, 2005, Laakso and Calvo, 2011, Onnis et al., 2008, Reali and Christiansen, 2005, Seidenberg, 1997), perhaps aided by primitive perceptual biases (Aslin & Newport, 2012). As Bates and Elman wrote, the imperfections due to the impoverished nature of linguistic input, often considered the sign of the presence of rich linguistic structures in the learner, “wash out with a large enough sample” (Bates & Elman, 1996, p. 1849).

Despite the attractiveness and simplicity of this model, the induction of structure from a large database is not necessarily a winning strategy for an infant learner. One word is one word, but one rule is thousands of words. In learning a lexicon that in a few months will grow spectacularly, infants would benefit greatly from grasping possible words as early as they can, projecting word structure without waiting for the lexicon to grow beyond their memory and their cognitive limitations. Thus, instead of extensively exploring evidence about words to induce their structure, infants may benefit from being less conservative and project hypotheses about possible words after acquiring few real words, deploying a learning procedure less akin to a rational assessment of available evidence than to guessing from exemplars (Bonatti, 2008). Several studies involving adults and infants suggest that mechanisms not reducible to the tracking of statistical distributions are present in language acquisition. Thus, adults engaged in a word-learning task tend to deploy a one-trial, “fast-mapping” strategy to fixate the meaning of novel words (Medina et al., 2011, Trueswell et al., 2013), a strategy that requires neither extensive encounters with word–object pairings nor statistical sampling of such encounters. Likewise, already at around 18 months, infants quickly learn novel words by means of elementary logical inferences that neither resemble nor require statistical computations (Halberda, 2003, Spiegel and Halberda, 2011). Even in highly idealized speech environments such as artificial language experiments, there is evidence that learners find words and word structure by means of different mechanisms that are sensitive to different features of the stimulus and are computing different functions. Peña, Bonatti, Nespor, and Mehler (2002) showed that adult participants can find words inside a continuous speech stream by computing statistical relationships among nonadjacent syllables, but are unable to extract information about the structures of these words when exposed to the same continuous familiarization, regardless of how long the familiarization is. Yet they are quickly capable of extracting the item structure once even a very short stream contains minimal, even subliminal, segmentation cues.2 Rather than helping, more experience is detrimental to this process of structure extraction. Indeed, a longer exposure to speech streams, which would give the learner more opportunities to extract statistical relations, hinders the detection of generalizations (Endress & Bonatti, 2007). These facts are difficult to account for assuming that the acquisition of lexical knowledge is exclusively based on a single inductive process requiring extensive data gathering (Endress & Bonatti, 2013). Instead, adult learners seem to look for very specific information in order to either extract words or find the possible words of a language. For example, adults can find words inside an artificial stream of consonant–vowel syllables by computing TPs among the consonants, but they cannot find the structural relationship inside these words on the basis of the same computations. Conversely, they can find word structure by computing relationships over the vowels, but cannot use the same relationships to find the words (Bonatti et al., 2005, Toro et al., 2008). Remarkably, even young infants are sensitive to the same differences at the beginning of their word-learning journey (Hochmann et al., 2011, Pons and Toro, 2010). All such considerations suggest that there is more to language acquisition than a cold analysis of the data by means of statistical computations. Words and possible words may be acquired by different learning strategies that respond to different aspects and different doses of the available experience. Recent adult neuroimaging evidence supports this conclusion, suggesting that the circuits involved in the projection of lexical rules and in the acquisition of lexical elements are different (De Diego Balaguer et al., 2007, De Diego-Balaguer et al., 2011, Mueller et al., 2008).

Here, we explore the hypothesis that at its onset language acquisition works differently from the process suggested by the empiricist model: that even before acquiring an extensive repertoire of words, infants try to construct the possible words of their language, so as to build a generative lexicon that extends beyond their experience with real words. If infants try to construct a notion of possible words while building their lexicon, then the question of the relative importance of actual and possible words arises. We propose that the most basic notion in lexical development is not the acquisition of novel words by means of statistical computations over a corpus of speech streams, but the projection of hypotheses about the possible words of a lexicon. We test this hypothesis by studying how 12- and 18 month-old infants—two crucial ages in lexical development—find words and possible words when exposed to artificial language streams. If our hypothesis is correct, then one would expect that infants develop a notion of possible words earlier than the moment at which they possess sufficiently refined statistical tools to fruitfully extract words from speech. By comparing infants at these two ages, we can also explore this prediction.

Artificially synthesized speech is well suited to explore our question, because of the control it offers over many features of the stimuli, such as the possibility to eliminate prosody, to manipulate the speech rate, to create exact statistical relationships among syllables, or to introduce precise elements of segmentation. One way to probe sensitivity to real and possible words in early language acquisition is to create situations in which the same experimental stimuli contain diverging information. Suppose infants listen to artificial speech streams from which they could both identify some snippets on the basis of their statistical coherence and find word-internal structure that cannot be extracted by computing the same statistical measures. The way the conflict is solved will indicate the relative importance that possible words and real words have in early language acquisition. In the following experiments, we created such streams by merging and adapting to infants two paradigms introduced by Peña et al. (2002) and Aslin et al. (1998). Peña et al. (2002) exposed adults to streams of trisyllabic sequences characterized by internal nonadjacent transitional probabilities of 1 and a varying middle syllable. Following standard practice, we call such statistically coherent items words, although obviously this definition has only a limited relationship with real words in a natural language. Participants could use nonadjacent distributional information to identify such words. However, the words also came in families. For example, PULIKI, PUBEKI, and PURAKI were words as identified by nonadjacent distributional information, but also shared the common structure PU_KI. Thus, not only could participants extract the actual words that occurred in the stream, but they could also identify their morphosyntactic structure. We will call languages constructed with these items AXC languages. After exposure to a stream of an AXC language, participants were tested for their preference for words vs. part-words or for rule-words vs. part-words. Rule-words were sequences of three syllables obtained by replacing the middle syllable of words with another syllable. They never occurred in the stream but were structurally similar to words because they shared the same morphological construction. Part-words were sequences of three syllables that occurred in the stream across word boundaries, but had no common constructions. Thus, AXC languages contain both statistical information favoring certain groups of syllables (words or part-words, to different degrees) and examples of words that may induce the projection of a rule potentially applying to unheard, novel words. Because rule-words never occurred in the stream, sensitivity to rule-words signals that the learner identified the structural aspects of word construction. Instead, because either words or part-words always had non-null TPs and frequencies, the ability to extract them from a stream signals that learners identified them by computing some measure of statistical information.

To create a conflict between sensitivity to possible words and statistically extracted words, we created AXC languages by chaining a few words with a common structure (and a nonadjacent TP of 1 between their first and last syllables), but, adopting a method introduced by Aslin et al. (1998) used here with a different logic, we repeated some of them so that some of the transitions between words would also turn out to be statistically highly coherent. Specifically, our streams contained only four words, but two of them were twice as frequent, so that the stream also contained trisyllabic sequences of adjacent syllables, spanning word boundaries, that occurred with a frequency as high as that of words and with high adjacent and nonadjacent TPs (.67). Thus, with LIMUFE and BAGASO twice as frequent as ligafe and bamuso (where uppercase letters indicate highly frequent words, and italic type indicates words with low frequency), the streams could contain snippets such as:

……LIMUFEBAGASOligafeBAGASObamusoligafeBAGASOLIMUFEBAGASO…

Given the arrangement of the words in the stream, some trisyllabic sequences spanning two infrequent words, such as musoli, occurred infrequently. However, some other sequences spanning two frequent words occurred frequently. For example, the sequence LIMUFEBAGASO occurred frequently and, as a consequence, so did the sequence FEBAGA, which is contained within it.3 Indeed, such a sequence occurred even more frequently than a low-frequency word such as bamuso. We call such sequences high frequency (HF) part-words.

In order to convey the logic of our experiments, it is worth commenting on the role of statistical measures in defining the notion of a word in artificial languages and on its relationship with words in natural languages. In natural languages, words are rich entities endowed with many layers of representations, spanning from phonology to meaning. By contrast, in artificial languages what counts as a word depends entirely on the statistical relationships among syllables. It is a presupposition of research using artificial speech that, at some level, such statistical relationships do contribute to creating those groups of sounds that we call words in natural languages. However, given the enormous distance between natural language words and ‘words’ induced by exposing participants to a few minutes of an artificial language, the relationship between the latter ones and the former ones is at most indirect and reduces to the fact that they both enjoy similar statistical properties: they both have TPs and absolute frequencies higher than other syllable sequences in their respective languages. In short, artificial languages can show how certain computations may allow the learner to single out some snippets of speech stream as potential candidates for items that might become ‘real words’ under the appropriate conditions. Going back to our stimuli with these considerations in mind, it is important to remember that words in our AXC languages became potential candidates for segmentation by virtue of their statistical properties, but also HF part-words possess most of the same statistical properties. Therefore, HF part-words may be equally good candidates for word segmentation as words themselves. In some respects, they are even better candidates: infrequent words have higher nonadjacent TPs than HF part-words, but have lower adjacent TPs. Although ours is the first study comparing infants’ ability to compute adjacent and nonadjacent TPs in the same experiments, there is universal agreement that, if anything, for infants (and most likely for adults) adjacent relations should be more salient than nonadjacent relations. Thus, besides containing few words having a common internal structure, our streams also contain some part-words that could be perceived as highly statistically coherent, on a par with or better than some words, and with a level of statistical coherence much higher than any real word in natural languages.

In our experiments, we first familiarized infants with streams that differed in some crucial properties that varied across experiments, but maintained the same statistical structure as explained above. Then, we tested their preferences for rule-words or HF part-words (Table 1). Crucially, in our experiments, structure is present within words, defined as constancy of their first and last elements, but the variation of the middle syllable that could induce infants to create an abstract representation of the words’ structure is minimal. Thus, if infants grasp possible words, they cannot do so by extracting statistical measures of inner variability which have been suggested to be crucial for infants to acquire nonadjacent structural patterns (Gómez, 2002). By contrast, HF part-words are favored by most statistical measures that we know infants can compute, particularly adjacent TPs and absolute frequency. If the representations that infants can form from a speech stream exclusively depend on such computations, then infants should always favor those items that are more strongly marked by the statistical relationships present in the input. If, instead, infants do not only look for actual words, but also try to form an abstract template of word structure from a few examples, then they should also be able to create a representation of the words that are possible, given the input, and hence accept rule-words as legal, although they have 0 frequency and 0 adjacent TPs.

We began our investigation by testing the boldest hypothesis we could formulate: infants extract words and possible words pretty much just as adults do, with the same kinds of stimuli, in the same input conditions, and hence most likely by means of the same processes (Peña et al., 2002). Thus, we tested the hypothesis that encounters with a few words instantiating a morphosyntactic rule embedded in a stream with segmentation indices separating words should foster sensitivity to possible words, favoring rule-words over HF part-words. By contrast, exposure to a continuous stream with the same statistical properties should favor the detection of the most statistically coherent items, favoring HF part-words over rule-words. We tested infants of two age classes: 18 months and 12 months. At 18 months, infants are already firmly engaged with language production. This implies that they have already mastered the perception and production of most sounds of their language and that their lexicon already includes information about word structure (Lidz et al., 2003). So, we expect them to be able to extract either words or possible words once exposed to speech streams with properties that may allow them to trigger either statistical computations over the streams or the projection of generalizations from word examples. In contrast, at 12 months, infants are only at their first stages of speech production. They are still in the process of zooming into their natural language and completing the fixation of crucial language-specific phonological properties (Werker and Tees, 1983, Werker and Tees, 2002)—a precondition for developing an extensive natural language lexicon. Indeed, word learning proper is generally assumed to begin around the first birthday (Bloom, 2000), when infants can also tell the difference between potential content words and language-specific functors (Shi, Werker, & Cutler, 2006). However, evidence that infants of that age can master morphosyntactic relations or can extract nonadjacent relations between or within words, is absent or severely limited. Yet, 12-month-olds are surely capable of computing adjacent TPs under varied conditions (Fiser and Aslin, 2002, Saffran et al., 1999, Saffran et al., 1996). Thus, if the empiricist picture of language acquisition is correct, 12-month-olds should be able to find statistically defined snippets of an AXC stream, but might not be able to form any notion of possible words. If, instead, a prominent pressure driving the acquisition of a lexicon is the urge to find its possible words quickly and early, then we should find sensitivity to word structure even at that age. Therefore, the way 12-month-olds negotiate the conflict between HF part-words and rule-words may allow us to gauge the relative importance of words and possible words at the onset of vocabulary acquisition.

We first test infants with the task that, in principle, is harder at both ages: the ability to form a notion of a possible word from a few word examples. Experiments 1, 2, and 3 probe 18-month-olds’ ability to distinguish rule-words from HF part-words after exposure to a segmented stream while controlling for some important aspects of the material. Experiment 4 will test the same ability in 12-month-olds. Then, we will test whether infants can solve the conflict posed by a continuous stream containing items favored by most statistical measures (HF part-words) and items favored by a common structural description (words). Experiment 5 tests 18-month-olds, and Experiments 6 and 7 tests 12-month-olds. To anticipate our results, we will show that at both ages infants are able to find possible words, but only at the later age do they possess sufficient statistical sophistication to identify statistically coherent items in the streams. We believe that these findings may turn the way we conceive of the process of lexical acquisition upside-down, as they show that possible words have developmental priority over words in the construction of a novel lexicon. They pose a serious challenge to the empirical picture of language learning and call for alternative ways to conceive of language acquisition, which we explore in the General Discussion.

Section snippets

Participants

Sixteen 18-month-old, full-term infants from Italian-speaking families with a minimum APGAR of 8 and no hearing or vision problems were retained for analysis (12 girls; mean age: 18 mo, 20 d; age range: 18 mo, 3 d to 19 mo, 14 d). We used stringent criteria for including participants and trials in our dataset. Infants were considered to be fussy if during familiarization or test they gave signs of discomfort while listening to the stimuli, such as making more than sporadic vocal emissions or

Participants

Sixteen 18-month-old, full-term infants from Italian-speaking families, with a minimum APGAR of 8 and no hearing or vision problems, were retained for analysis (7 girls; mean age: 18 mo, 26 d; age range: 18 mo, 7 d to 19 mo, 10 d). An additional 18 infants participated but were excluded from analysis (16 because of fussiness, and 2 because they exceeded maximum looking time criteria).

Stimuli and procedure

Infants were exposed to the same segmented familiarization stream as in Experiment 1. However, in the test phase the

Participants

Sixteen 18-month-old, full-term infants from Italian-speaking families, with a minimum APGAR of 8 and no hearing or vision problems, were retained for analysis (7 girls; mean age: 18 mo, 7 d; age-range: 17 mo, 29 d to 19 mo, 9 d). An additional 18 infants participated but were excluded from analysis (16 because of fussiness, 2 because they exceeded maximum looking time criteria).

Stimuli and procedure

We synthesized a novel familiarization stream, built with exactly the same constraints as in Experiment 1 except for the

Participants

Sixteen 12-month-old, full-term infants from Italian-speaking families, with a minimum APGAR of 8 and no hearing or vision problems, were retained for analysis (8 girls; mean age: 12 mo, 20 d; age range: 12 mo, 4 d to 13 mo, 0 d). An additional 22 infants participated but were excluded from analysis (18 because of fussiness, 4 because they exceeded maximum looking time criteria).

Stimuli and procedure

Stimuli and procedure were identical to Experiment 1.

Results and discussion

Fig. 1B presents the results of Experiment 4. Like their older peers,

Actual and possible words at the onset of lexical acquisition

Adult experiments show that segmentation marks in a stream—however minuscule—are crucial for grasping word structure. When listening to a continuous stream, adults are insensitive to structure, but they can find statistically coherent units computing distributional information (Endress and Bonatti, 2007, Peña et al., 2002). We now study what information infants can extract when exposed to streams that have the same statistical properties as those of Experiments 1–4, but have no segmentation

Participants

Sixteen 18-month-old, full-term infants from Italian-speaking families, with a minimum APGAR of 8 and no hearing or vision problems, were retained for analysis (12 girls; mean age: 18 mo, 20 d; age range: 18 mo, 3 d to 19 mo, 14 d). An additional 22 infants participated but were excluded from analysis (20 because of fussiness, 2 because they exceeded maximum looking time criteria).

Stimuli and procedure

We synthesized a novel familiarization stream. It was built with the same constraints as in Experiment 1, but any pauses

Participants

Sixteen 12-month-old, full-term infants from Italian-speaking families, with a minimum APGAR of 8 and no hearing or vision problems, were retained for analysis (8 girls; mean age: 12 mo, 20 d; age range: 12 mo, 4 d to 13 mo, 0 d). An additional 22 infants participated but were excluded from analysis (20 because of fussiness, 2 because they exceeded maximum looking time criteria).

Results and discussion

Fig. 1D presents the results of Experiment 6. Twelve-month-olds showed

Participants

Sixteen 12-month-old, full-term infants from Italian-speaking families, with a minimum APGAR of 8 and no hearing or vision problems, were retained for analysis (9 girls; mean age: 12 mo, 21 d; age range: 12 mo, 2 d to 13 mo, 8 d). An additional 19 infants participated but were excluded from analysis (14 because of fussiness, 5 because they exceeded maximum looking time criteria).

Stimuli and procedure

We synthesized a novel continuous familiarization stream. It was built exactly as in Experiment 5, but we repeated the

General discussion

Learning a language requires building a lexicon and fixing the generative system that allows humans to constantly understand, plan, and utter sentences they have never heard before. However, the problems of finding words in a speech stream and of identifying the structural information that makes language productive are not independent. Words are more than simple unanalyzed chunks of syllables. They also contain structural information that a learner must acquire.nstead, adult learners seem to

Acknowledgments

The work was completed despite the attempts to sabotage scientific research operated by the Ministerio de Competitividad y Productividad (headed by Luis de Guindos). It was supported by grants PSI2012-31961 and FVG “PsyScope XL” to L.L.B., and by the Fyssen Foundation Grant to E.M. We thank A. Isaja, L. Uttley, L. Filippin, F. Gandolfo, M. Sjekloca, K. Brink, N. Sebastián Gallés, A. Endress, K. Mehta, J.M. Toro, C. Clifton and three anonymous reviewers for scientific discussions and technical

References (81)

  • L. Onnis et al.

    Learn locally, act globally: Learning language from variation set cues

    Cognition

    (2008)
  • F. Pons et al.

    Structural generalizations over consonants and vowels in 11-month-old infants

    Cognition

    (2010)
  • J.R. Saffran et al.

    Statistical learning of tone sequences by human infants and adults

    Cognition

    (1999)
  • M. Shukla et al.

    An interaction between prosody and statistics in the segmentation of fluent speech

    Cognitive Psychology

    (2007)
  • C. Spiegel et al.

    Rapid fast-mapping abilities in 2-year-olds

    Journal of Experimental Child Psychology

    (2011)
  • J.C. Trueswell et al.

    Propose but verify: Fast mapping meets cross-situational word learning

    Cognitive Psychology

    (2013)
  • M. van Heugten et al.

    Linking infants’ distributional learning abilities to natural language acquisition

    Journal of Memory and Language

    (2010)
  • A. Vouloumanos

    Fine-grained sensitivity to statistical information in adult word learning

    Cognition

    (2008)
  • J.F. Werker et al.

    Cross-language speech perception: Evidence for perceptual reorganization during the first year of life

    Infant Behavior & Development

    (2002)
  • C.D. Yang

    Universal grammar, statistics or both?

    Trends in Cognitive Sciences

    (2004)
  • R.N. Aslin et al.

    Statistical learning: From acquiring specific items to forming general rules

    Current Directions in Psychological Science

    (2012)
  • R.N. Aslin et al.

    Computation of conditional probability statistics by 8-month-old infants

    Psychological Science

    (1998)
  • E. Bates et al.

    Learning rediscovered

    Science

    (1996)
  • P. Bloom

    How children learn the meanings of words

    (2000)
  • L.L. Bonatti

    On pigeons, humans, language and the mind

  • L.L. Bonatti et al.

    How to hit Scylla without avoiding Charybdis: Comment on Perruchet, Tyler, Galland, and Peereman (2004)

    Journal of Experimental Psychology: General

    (2006)
  • L.L. Bonatti et al.

    Linguistic constraints on statistical computations: The role of consonants and vowels in continuous speech processing

    Psychological Science

    (2005)
  • N. Chomsky

    Syntactic structures

    (1957)
  • M.H. Christiansen et al.

    The secret is in the sound: From unsegmented speech to lexical categories

    Developmental Science

    (2009)
  • R. De Diego Balaguer et al.

    Different neurophysiological mechanisms underlying word and rule extraction from speech

    PLoS ONE

    (2007)
  • R. De Diego-Balaguer et al.

    Brain dynamics sustaining rapid rule extraction from speech

    Journal of Cognitive Neuroscience

    (2011)
  • Dutoit, T., Pagel, V., Bataille, F., & Vreken, O. (1996). The MBROLA project: Towards a set of high-quality speech...
  • J.L. Elman

    The emergence of language: A conspiracy theory

  • Endress, A., & Bonatti, L. L. (2013). Some mechanisms of artificial language learning. submitted for...
  • J. Fiser et al.

    Statistical learning of new visual feature combinations by infants

    PNAS – Proceedings of the National Academy of Sciences of the United States of America

    (2002)
  • L. Gerken et al.

    Three exemplars allow at least some linguistic generalizations: Implications for generalization mechanisms and constraints

    Language Learning and Development

    (2008)
  • Y. Gertner et al.

    Learning words and rules: Abstract knowledge of word order in early sentence comprehension

    Psychological Science

    (2006)
  • R. Gómez et al.

    The developmental trajectory of nonadjacent dependency learning

    Infancy

    (2005)
  • R.L. Gómez

    Variability and detection of invariant structure

    Psychological Science

    (2002)
  • M.T. Guasti

    Language acquisition: The growth of grammar

    (2002)
  • Cited by (0)

    View full text