How Language Is Learned: Mechanisms and Theories

25 min

Two minutes of nonsense

In 1996 Jenny Saffran, Richard Aslin and Elissa Newport played 8-month-old babies a recording of a made-up language for just two minutes. It consisted of four invented three-syllable words, such as bidaku and padoti, strung together in random order with no pauses, no stress and no change in pitch: …bidakupadotigolabubidaku… The only clue to where one word ended and the next began was statistical: inside a word, bi was always followed by da, but after ku, the end of a word, any of three other words could follow.

Then the babies heard either the “words” again or “part-words” that crossed a boundary, such as kupado. They listened longer to the part-words, a sign that these sounded new. In two minutes, babies who could not yet talk had tracked how often syllables followed one another and used it to find the words.

Figure 1. Transitional probabilities in Saffran’s artificial language.

Worked example: computing a transitional probability

The transitional probability of syllable $Y$ following syllable $X$ is

TP(Y \mid X) = \frac{\text{number of times } XY \text{ occurs}}{\text{number of times } X \text{ occurs}}

Suppose that in the recording ku occurs 45 times and is followed by pa 15 times. Then

TP(pa \mid ku) = \frac{15}{45} = \frac{1}{3} \approx 0.33

If bi occurs 45 times and is followed by da all 45 times, then $TP(da \mid bi) = \frac{45}{45} = 1.0$ . The dip from 1.0 to 0.33 marks a likely word boundary.

Definition

Statistical learning: picking up regularities in the input, such as which sounds or words tend to follow each other, without being taught.

Fast mapping: forming an initial link between a new word and its meaning after only one or a few exposures.

Joint attention: two people attending to the same object or event and knowing that the other is attending to it too.

Recast: an adult’s correct rewording of a child’s error, without explicit correction (child: “He goed”; adult: “Yes, he went to school”).

Zone of proximal development: the gap between what a learner can do alone and what they can do with help (Lev Vygotsky).

1. Mechanisms of language learning

Mechanism	Evidence
Imitation	Children copy words and accents they hear, but also say things no adult says (“goed”).
Reinforcement and feedback	Parents respond more to successful communication; recasts give implicit feedback on form.
Statistical learning	Saffran and colleagues (1996): infants segmented words using transitional probabilities.
Fast mapping	Susan Carey and Elsa Bartlett (1978) used the made-up colour name “chromium” once in a preschool classroom (“bring me the chromium tray, not the red one”); a week later some children still showed partial knowledge of the new word.
Social cues	Dare Baldwin (1993) found that 18-month-olds linked a new word to the object the speaker was looking at, not the one they were looking at themselves.

2. Theories of language acquisition

Theory	Key claim	Strength	Weakness
Behaviourist (Skinner)	Imitation and reinforcement shape language.	Explains why children learn their community’s language and accent.	Cannot explain novel sentences or overregularisation.
Nativist (Chomsky)	An inborn language acquisition device and universal grammar.	Explains speed and universality of acquisition.	Hard to test; downplays input and social use.
Cognitive (Piaget)	Language develops as thinking develops; words map onto concepts already formed.	Explains why words like “gone” appear when children understand object permanence.	Underestimates how language itself shapes thought.
Social interactionist (Vygotsky, Bruner)	Language grows through interaction with more skilled partners; Bruner’s language acquisition support system (LASS) is the adults’ routines, such as book-sharing and naming games, that scaffold the child.	Matches evidence on joint attention and recasts.	Children in cultures where adults talk less directly to infants still acquire language.
Usage-based (Tomasello)	Children learn from use, by reading others’ intentions and finding patterns, building grammar gradually from specific phrases.	Matches early, item-by-item grammar and statistical learning.	Struggles to explain some abstract grammatical knowledge.
Connectionist	Networks of simple units learn by adjusting connections from input; Rumelhart and McClelland’s (1986) network even produced “goed”-type errors.	Shows rules can emerge without being stored as rules.	Early models were criticised for training choices (Pinker and Prince, 1988).

Most researchers now combine these: children bring powerful learning abilities, some perhaps specific to language, and apply them to rich social input.

CUNY Graduate Center: “How Babies Learn Language: A Talk by Jenny Saffran”. Saffran explains her own experiments; listen for how infants’ listening times reveal what they have learned.

Important

Evaluate statistical learning as an explanation. Saffran’s finding has been replicated many times and extended to tones, shapes and even other species, which shows statistical learning is a general, powerful mechanism. But the artificial language was far simpler than real speech, where words vary in length and frequency, and listening time is an indirect measure. Statistics can find words, but critics argue they cannot by themselves explain how children acquire abstract grammar or link words to meanings, which is why social cues and other mechanisms are needed too.

Warning

Common mistake: calling Bruner’s LASS and Chomsky’s LAD the same thing. The LAD is an inborn mental device inside the child; the LASS is the support provided by people around the child.

Tip

Exam technique: in a theories essay, place each theory on a line from “nature” (nativist) to “nurture” (behaviourist), with interactionist and usage-based views in between, then use one study per theory as evidence.

✦

Practice

Define statistical learning and describe how Saffran and colleagues tested it.
In a recording, go occurs 60 times and is followed by la 60 times; bu occurs 60 times and is followed by pa 20 times. Calculate both transitional probabilities and say which pair spans a word boundary.
What did Baldwin (1993) find, and which theory does it support?
Distinguish Chomsky’s LAD from Bruner’s LASS.
A child says “Mummy, I runned fast!” and the mother replies “Yes, you ran very fast!” Name the mother’s response and explain its role.
Give one strength and one weakness of the usage-based theory.
Evaluate the claim that children acquire language mainly through imitation.

Show answersHide answers

Learning regularities in input, such as which syllables follow which, without instruction. Eight-month-olds heard two minutes of a continuous stream of invented words; afterwards they listened longer to part-words than to words, showing they had used transitional probabilities to segment words.
$TP(la \mid go) = \frac{60}{60} = 1.0$ ; $TP(pa \mid bu) = \frac{20}{60} \approx 0.33$ . The lower probability, bu–pa, marks a word boundary.
Eighteen-month-olds linked a new word to the object the speaker was looking at, not the one they were looking at. It supports social interactionist (and usage-based) theories, which emphasise reading others’ intentions.
The LAD is an inborn mental device for grammar inside the child (nativist). The LASS is the support system of adults and routines around the child (interactionist).
A recast: she repeats the idea with the correct form without saying the child was wrong, giving implicit feedback that helps the child notice the correct form.
Strength: matches evidence that early grammar is built item by item and that children track patterns statistically. Weakness: struggles to explain some abstract grammatical knowledge that seems to go beyond the input.
Imitation explains learning the local language and accent, but children produce forms no adult says (“runned”, “goed”), create new sentences, and Nicaraguan deaf children built grammar no one modelled. Statistical learning, social cues and possibly inborn abilities are also needed. Imitation is one mechanism, not the main one.

Sources and credits

Works described: Saffran, J. R., Aslin, R. N. and Newport, E. L. (1996), “Statistical learning by 8-month-old infants”, Science 274; Carey, S. and Bartlett, E. (1978), “Acquiring a single new word”, Papers and Reports on Child Language Development 15; Baldwin, D. A. (1993), Journal of Child Language 20; Bruner, J. (1983), Child’s Talk: Learning to Use Language, Norton; Vygotsky, L. S. (1978), Mind in Society, Harvard University Press; Tomasello, M. (2003), Constructing a Language, Harvard University Press; Rumelhart, D. E. and McClelland, J. L. (1986), “On learning the past tenses of English verbs”, in Parallel Distributed Processing, MIT Press; Pinker, S. and Prince, A. (1988), Cognition 28.
Video: “How Babies Learn Language: A Talk by Jenny Saffran”, CUNY Graduate Center, embedded from YouTube.

Written by Ganzaa: the text, the tables, Figure 1 and the exercises.

1Two minutes of nonsense

Worked example: computing a transitional probability

21. Mechanisms of language learning

32. Theories of language acquisition

Two minutes of nonsense

1. Mechanisms of language learning

2. Theories of language acquisition