Thiessen and Saffran (2003)

Stimuli: Acoustics. All stimuli were generated by the MacInTalkspeech synthesizer running on a PowerMac 5300
(Apple Computer, Inc..Cupertino, CA). Two artificial languages were synthesized for use duringfamiliarization, one
iambic and one trochaic, each consisting of the samefour bisyllabic nonsense words: dapu, dobi, bugo, and diti. Each
languageconsisted of the same words spoken in the same randomized order. Therewere no pauses between words, and
all syllables were fully coarticulatedThe synthesizer produced syllables with a monotonic FO (fundamentalfrequency)
of 200 Hz. We used synthesized speech because it allowedbetter control over the acoustics of stress. Pitch contour can
have unnaturalvariation when it occurs across spliced natural utterances. Therefore. eventhough synthesized speech is
less like what an infant hears in the naturalenvironment, it allows tighter control over the parameters of stress.
Stimuli. The familiarization language had the same word order and statistical structure as the languages used in
Experiments 1 and 2. However, no stress cues were present. All words were synthesized monotonically with the
MacInTalk speech synthesis program. Average syllable length was 223.8 ms. The duration of the language was 2
min. The test items were identical to those used in Experiments 1 and 2.

Johnson & Tyler (2010)

In this study, we begin to ask whether infants’ statistical learning abilities are robust enough to handle the complexity
and variation that is characteristic of natural language. We address this question not by using natural language, but by
using an artificial language that contains slightly more variation, and is thus slightly more natural, than the artificial
languages used in earlier infant studies. There are many regularities in artificial languages, not present in natural
language, that we could have focused on in the current study (e.g. syllable structure, word length, and token
variability), but we chose to carry out the manipulation that we believed would result in the least extreme increase in
task complexity: uniformity of word length. Infants have been shown to readily segment an artificial language
consisting solely of trisyllabic words (Saffran et al., 1996; Johnson & Jusczyk, 2001) or disyllabic words (Thiessen &
Saffran, 2003). How would infants perform with an artificial language containing a mixture of disyllabic and
trisyllabic words? If infants’ ability to track transitional probabilities between syllables provides the bootstapping
required for learning additional language-specific word boundary cues, then one would expect this ability to be fairly
robust. If removing the word-length regularity from artificial languages negatively impacts infants’ ability to learn the
statistical structure of the language, then this would indicate that infants can rapidly extract the statistical structure of
an artificial language only if other aspects of that language are simplified to the point of being predictable. This in turn
would raise the question of whether infants’ solution to the word segmentation problem for artificial languages is
qualitatively different from their solution for natural language.

As an additional side question, we also ask whether 5.5-month-olds are as skilled in tracking transitional probabilities
between syllables as 8-month-olds. This question is important since infants begin segmenting words from speech by 6
to 7.5 months of age (Bortfeld, Morgan, Golinkoff & Rathbun, 2005; Jusczyk & Aslin, 1995), and presumably
infants would have to be tracking transitional probabilities in the ambient language for some time before they would
be able to use this information to infer word boundaries. The 5.5-month-olds we test here are 1 month younger than
the youngest age group that has ever been tested in an artificial language segmentation task. If such young infants
were to succeed at segmenting an artificial language using transitional probabilities alone, this would be an impressive
demonstration of infants’ early sensitivity to distributional regularities in the environment.

Infants’ ability to segment words from an artificial language containing no cues to word boundaries other than the
transitional probabilities between syllables is often cited as support for the theory that infants use transitional
probabilities to solve the word segmentation problem. For example, it has been suggested that infants learn which
acoustic characteristics define word boundaries only after they begin segmenting words from speech based on
transitional probabilities between syllables (Thiessen & Saffran, 2003, 2007). We do not deny the importance of
distributional learning in language acquisition; however, we question the notion that transitional probabilities between
syllables necessarily enable infants to initially begin segmenting words from speech. In the introduction, we cited two
major weaknesses in the statistical bootstrapping theory for word segmentation. First, evidence that transitional
probabilities define word boundaries in an infant-directed corpus does not demonstrate that 5- to 6-month-olds have
the phonological knowledge or processing capabilities necessary to extract these regularities. Second, it is not clear
that the statistical learning abilities infants demonstrate when faced with an artificial language can scale up to handle
real language input. In this paper, we begin to address both issues. Our results indicate that both 5.5- and 8-month-olds
have the processing capacity to extract statistical regularities from a highly simplified language; however, this ability
appears to break down when the input language is even slightly more complex than that used in earlier studies. Thus,
our results suggest that infants’ ability to track transitional probabilities between syllables is much more fragile than
earlier studies have suggested. This in turn suggests that an alternative, or at least revised, explanation may be needed
to describe how infants initially solve the word segmentation problem.
To reiterate, our findings do not refute the ability of infants to track transitional probabilities. Indeed, we know of no
other explanation for how infants succeeded in our UWL condition. Our point is simply that natural languages vary a
great deal from artificial languages, and tracking transitional probabilities might be a segmentation solution better
suited for artificial than natural languages. We do not see our results as indicating that infants are incapable of
segmenting words from the MWL. If we had lengthened the familiarization phase, infants may have succeeded.
Likewise, if we had replaced the word length regularity with another type of regularity (e.g. all words start with a
labial consonant), we may have observed successful segmentation of the mixed word length language.
Indeed, Thiessen, Hill and Saffran (2005) report that 6.5- to 7.5-month-olds learn an artificial language containing
words of varying length when the language is produced with infant- but not adult-directed prosody. This finding is
perfectly consistent with the current findings. As an artificial language becomes more complicated (e.g. containing
words of varying as opposed to uniform length), other natural speech cues such as exaggerated prosody are needed to
make word segmentation possible. Here we removed one regularity, unrelated to transitional probabilities, that is
typically present in the artificial languages used with infants and they no longer showed evidence of segmentation.
This raises questions as to how strongly infants rely on this information when segmenting natural language.

The goal of the current study was not to downplay the impressive nature of infants’ statistical learning abilities.
Clearly, infants are highly skilled at detecting probabilistic patterns in their input. Rather, we view this work as a small
step towards beginning to critically examine the psychological plausibility of statistical bootstrapping theories of word
segmentation. One might make the criticism that we have complicated the segmentation task, and yet we still expect
infants to extract words in a very brief period of time. But, in fact, we are still presenting infants with a language that
is highly simplified with respect to natural language; the transitional probabilities defining within- versus between-
word sequences are as exaggerated as those in earlier artificial language studies, and listening to the same four words
repeating one after another at a constant speech rate continuously for 2.5 minutes would never happen in a real
language learning setting (see Johnson & Seidl, 2009, for discussion). The fact that infants were successful at
segmenting only the uniform word length language suggests that the rhythmic word length regularity of this language
may have made it easier to learn. This factor was not considered in earlier artificial language studies of this type. Note
that the major contribution of Saffran et al. was to show that infants could segment an artificial language containing
no cues to word boundaries other than the transitional probabilities between syllables. What our results suggest is that
this may not be the case. The infants tested by Saffran et al. may have succeeded due to the presence of not only
transitional probability cues, but also word length regularities. 2 These regularities do not exist in natural language. To
justify claims that infants initially solve the word segmentation problem in natural language by tracking transitional
probabilities, further studies examining these types of issues are required. This sort of research is integral to
understanding the role transitional probabilities between syllables play in natural language acquisition.

(a) The intake component identifies the features of the information to which learners are sensitive; (b) the UG
component identifies the class of representations that shape the nature of human grammatical systems; and (c)
the inference component identifies how learners combine the intake with UG in selecting a particular
grammar.

deconstructing a language acquisition device into separate components for input sensitivity, a rich space of
representations and rational statistical inference can explain many key results from both traditions of research in
language acquisition.

The model includes three components: the intake component for identifying information to which children are
innately sensitive, the UG component for recognising the category of representations that influence how human
grammatical systems are formed, and the inference component for integrating the input and UG to learn grammar.

Inference mechanisms connect UG with the intake to determine the appropriate mapping from abstract representations
to surface form.

Brand (2000)
The developmental cast of this model makes it imperative to study the origins of word learning as well as the
transformation that takes place in the 2nd year of life. Now with a method that can investigate development of
language in 12- and 24-month-old children, data needed to be collected that assessed the three main hypotheses of the
model: (a) that children detect and utilize multiple cues for word learning, (b) that their reliance on these cues changes
over the course of development, and (c) that these principles for word learning emerge from the word learning
experience.
 Experiment 1 was designed to assess the utility of the Interactive intermodal preferential looking paradigm
and to examine whether, and under what conditions, children at these three ages can learn a novel word. Experiment 2
was a control experiment designed to explore the magnitude of the word learning effects found in Experiment 1.
Experiment 3 offered a further control and a critical test of whether 12-month-olds had really learned novel words in
Experiment 1.

Experiments 1, 2, and 3 addressed three hypotheses: that children are sensitive to multiple cues in word
learning, that the relative weightings of the cues for word learning shift with word learning experience, and that
children move from an immature to a mature principle of reference. Evidence was presented in favour of each of these
claims. Experiments 1 and 2 presented evidence, in the salience phases, that children of all ages were sensitive to
perceptual salience. Indeed, even the 24month-olds were attracted by perceptual salience, although they did not rely
on it in the conflict condition when assigning a word to a referent in the test phase. There also is evidence that children
at each age were sensitive to social cues. The 19- and 24-month-olds were able to follow social eye gaze and
determine word reference, as evidenced by the condition effect observed in the training and test phases of Experiment
1. Furthermore, the 12-month-olds’ sensitivity to social eye gaze in the training phases of Experiments 2 and 3, as well
as the trend toward a condition effect in the test phases of Experiment 3, reinforces the claim that they are sensitive to
multiple cues in the word learning situation, although they might not make full use of these.
Interestingly, the data also suggest that, at the earliest stages of word learning, the cues not only must be
present but also must be in alignment for word learning to occur. The failure of 12-month-olds to show any evidence
of word learning in the conflict condition of Experiment 3 suggests that the cue of perceptual salience and the cue of
social eye gaze must both “point” to the same object. When they do not, word learning becomes very difficult for
these infants, who are unsure of which cues are more reliable in the word learning situation. Similarly, Dunham et al.
concluded, “in the presence of competing cues, [infants] may adopt the strategy of simply aborting the encoding
process rather than risk the linguistic implications of a mapping error” (p. 831).
Experiments 1 and 2 reveal that at the ages of 12 and 24 months, children weight perceptual and social cues
very differently. This differential weighting helps explain the changing character of word learning over time. Taken
together, the first three experiments suggest that 12-montholds rely less on the cue of social eye gaze and relatively
more on perceptual salience, whereas 24-month-olds rely more on social cues. This confirms our second hypothesis—
the weights these cues are given for word learning shift over developmental time. As Table 2 shows for Experiment 1,
perceptual salience is weighted heavily at first, then declines (but does not disappear) by 24 months. The weight given
to social eye gaze clearly increases, as evidenced by the significant condition effect (seen in the percentage change
row for the 19- and 24-month-olds).
Finally, it can be argued that children at 12 months have but an immature principle of reference, as our third
hypothesis presumed, approaching the word learning task in a qualitatively different way than the 19- and 24-month-
olds. The latter take the perspective of the speaker into account in linking a name to a referent; the former, while
sensitive to conflicting cues, only seem to learn words that correspond to their own perspective.

(a) that children detect and utilize multiple cues for word learning, (b) that their reliance on these cues changes over
the course of development, and (c) that these principles for word learning emerge from the word learning experience.

Experiments 1, 2, and 3 addressed three hypotheses: that children are sensitive to multiple cues in word learning, that
the relative weightings of the cues for word learning shift with word learning experience, and that children move from
an immature to a mature principle of reference.