Statistical learning in infants and children

At birth

nfants prefer forward going speech, their own language, they can discriminate function words from context words by acoustic sound, they prefer their mother's voice to other women and can discriminate sound contrasts in all languages. 

They are learning about language in the womb.

By 1 year old

 they can only make contrasts in their own language.

Noam chomsky

He was a nativist and believed that aspects of language such as grammar are innate because of “poverty of stimulus”. The input available to the child is insufficient for language to be learned. 

He believed that there is “universal grammar” and that all languages share innate principles of grammar that allow children to learn them by 3 years old. He also argued that language cannot have been learned because there is no punishment involved in learning speech.

Evidence of Naom Chomsky’s claim

researchers working for over 50 years have never found any grammatical structures common to all languages.

How is language learned?

It is now argued that rather than traditional reinforcement and punishment learning, language is instead learned through statistical learning which is learning of patterns.

Predictability of Language

 Infants ability to learn language is from patterns and predictability. Grammatical categories fall in predictable places in sentences, therefore children can learn where to put their verbs and nouns by listening to others.

(Mintz 2003) - these frames occur reliably in child input and help children learn about grammar.

the brain

 Is specialised for statistical learning. It allows identification of a word, however breaks in the speech stream are often found within words. How does a baby know what is a word?

Continuous speech

 8 month olds were played a continuous speech stream of nonsense syllables for a minute. When syllables are coming together you can start to discriminate what a word might be. The only cues as to what is a word is the transitional probabilities between pairs of syllables. After 1 minute when the stream was changed to introduce some new “words” babies preferred the new ones then the ones that had played in the previous minute also because they had already learned those ones. 

This discredited chomsky because it shows the stimulus is not impoverished.

Teinonen et al 2009

newborns were played nonsense syllables while they slept for 15 minutes. When they woke up they were played 45 minutes with some words from the other stream and some new ones. When their brain activity was measured there was less response to the words from the sleep stream.

Even when babies are asleep they use familiarity to group sounds into words.

Theory of mind

Many claim that a theory of mind (understanding of desires, intentions and beliefs) is innate. However this may also be statistical learning, maybe infants simply learn about patterns in behaviour and later about mental states.

Baldwin 2001

Showed 2 videos. One where it stopped mid bending over and one where bending over had occurred and a rug had been grasped. He found that infants would look longer at the mid bending frame, but they needed to see the sequence of actions first though.

Desires and mental states are expressed through actions.

So, there is the argument that infants look longer at the mid action frame because they have knowledge of intention, which supports the theory of mind.

Or there is a statistical argument that infants learn to predict patterns of behavior and that stopping mid sequence is unusual. 

Either of these is plausible. According to the behaviour theory infants can then learn from this about mental states.

Impairment of statistical learning - Specific language impairment (SLI)

Short sentences, small vocab, word finding problems, difficulty learning new words. 50% develop dyslexia. So if statistical learning underpins language learning then people with these impairments and worse language skills should be impaired in a statistical task.

Evans 2009 -

tested 5-14 yr olds. Some with SLI and some typical.

The children then got to draw while either listening to nonsense syllables for 21 or 42 minutes, or they drew while listening to tones with irregular patterns for 42 minutes. The children were asked which of the two combinations they had heard while drawing. 

Typically developing children has above chance guessing for all tasks. 

SLI children did much worse.

It can therefore be concluded that SLI kids had general difficulties in learning about patterns in both syllables and tones.

Statistical learning and autism

Autism is a neural disorder with deficits in social interaction, understanding others, communication and language and effects 1-2 out of a 1000 people. 4x more males than females. 

Do autistic children have difficulty detecting co-occurrence between syllables? They compared brain activity listening to a series of syllables between typical and autistic children. Typical children had an increase in activity in certain brain regions when listening to nonrandom syllables compared to random. In autisitc children they did not see this increase.

The putamen and the inferior parietal lobe

key brain areas for patterns. Less severe autistics have more activity in these areas than more severe ones. This difference in autistic children explains their language and social symptoms and why they can’t predict what people will do.

Grammar

grammatical categories occur in predictable positions in sentences —> frequency frames - these frames occur reliable in child input (Mintz, 2003) —> help children learn about grammar, statistical learning assists language learning.

Word segmentation

S.L.I allow identification of a ‘word’. Breaks in the speech stream are found within words = 8months, continuous, synthesised speech stream. Example= words in a random order for 2 mins only cues to word boundaries —> transitional probabilities between syllable pairs.

• infants prefer novel words (turn head)