Lecture 8: Language acquisition

Question 1

Give the right age periods for the use of the words “baby/newborn”, “infant” and “toddler”.

Answer 1

Baby/newborn = 0-2 months
Infant = 3-12 months
Toddler = 12-36 months

Question 2

What is High Amplitude Sucking (HAS)? Describe it.

Answer 2

HAS is an experimental method appropriate for testing infants from birth to age 4 months.

A non-nutritive pacifier is put in a baby’s mouth that is connected to the computer via a transistor. The sucking of the baby is recorded by an amplifier.
○ When a baby gets excited, it sucks harder and faster
○ If a baby is bored, the speed and strength decreases

Question 3

What is the headturn paradigm? Describe it.

Answer 3

Experimental method appropriate for testing infants from the age of 4- 12 months; they should be able to hold their heads and move their head to one side and the other. At around 12 months, babies become more mobile, hence it’s hard to keep them still.

The experimenter records the time children look at a flashlight to attract their attention located in front of the child or on the sides. In addition, a speaker behind the flashlights produces sounds until the infant gets bored and turns the head away.

Result:

• Infants 4-7/8 months look at the side they already know for a longer time
• Onwards from 7/8 months, this preference switched: they look more towards the new side

Question 4

What are the three important cognitive mechanisms of language acquisition? Are any of them exclusive to humans?

Answer 4

• General acoustic perception (listening to speech)
• Computational abilities (statistically analyze what is important, and what is not)
• Social interaction

Neither of those mechanisms are exclusive to humans, but we are the only ones in the animal kingdom that has language as complex as we have

Question 5

Are tamarins (monkeys) able to perceive different languages based on different rhythmic properties? Explain.

Answer 5

Yes. Experimenters trained tamarins with the head turn paradigm with languages that differ in rhythmic properties as stimuli and the monkeys could successfully distinguish between them.

Question 6

How good are chinchillas in differentiating a /d/ from a /t/ ?

Answer 6

As good as humans

Question 7

What are phonemes?

Answer 7

minimal sound unit in language

Question 8

What are Phones/ Phonemic Units?

Answer 8

Variations of a phoneme (like differences between American English and British English)

Question 9

Do babies pick up on auditory stimuli before birth? Explain?

Answer 9

Yes. From 30 weeks of pregnancy, the babies heart accelerates and/or the baby moves when they hear sounds.

Question 10

Do infants or adults have a more general acoustic perception?

Answer 10

Infants have a more general acoustic perception, while that of adults is much more focussed on their own language.

Question 11

2-day-old babies were presented with their native language or a foreign language and tested via high-amplitude sucking. What happened?

Answer 11

Babies continued sucking high hearing their native languages compared to when they were exposed to the foreign language

Question 12

5-day-old babies were presented first with English and later with Japanese or two times in a row with English and tested via high-amplitude sucking. What happened?
What happens if the same experiment is conducted with English - Dutch?

Answer 12

• Babies who got the same language after were less excited to suck
• Babies who were presented with two different languages were more excited to suck

As English and Dutch have the same rhythmic features, the sucking rate didn’t change here.

Question 13

How does Hebbian learning connect with language acquisition?

Answer 13

Cells that often fire at the same time will create pathways; the more often a word is repeated, the more certain a child becomes of its meaning

Question 14

What is statistical learning useful for in language acquisition?

Answer 14

• Categorical learning: Assigning words/concepts to groups
• Learning transitional probability: The probability of a syllable occuring
• Identifying stress patterns

Question 15

What’s the magnet effect

Answer 15

The exact sound of a vowel can depend on gender, age, dialect, phonetic contexts, etc.
The perceptual magnet effect describes an increased generalization capability for the perception of those vowels.

Question 16

In which age period do children start language specialization? How could that be tested?

Answer 16

Between 6-12 months.

Letting American and Japanese infants try to distinguish /l/ and /r/.

Question 17

What is “native-language neural commitment” (NLNC)?

Answer 17

The use of statistical and prosodic information will help shape our universal ability to a specific native language.

Question 18

What are some important components that will affect the linguistic proficiency of bilingual children? Can NLNC be “tricked” by learning a second language?

Answer 18

• If L1 & L2 are similar, L2 tends to be easier to learn
• If you are younger, learning a second language is easier. However, you will be less proficient in both languages than a monolingual
• Motivation affects how you learn a second language

This suggests that bilingual children develop a slightly different NLNC, but the proficiency of both native languages is going to be slightly worse than in monolinguals.

Question 19

What are the two steps infants need to do before trying to understand the meaning of a sentence?

Answer 19

1. They recognize phonetic units: Abigmonkeyiseatingaredapple
2. They identify word boundaries: A big monkey is eating a red apple

Identifying word boundaries is difficult because there are no consistent audio markers for word beginnings and endings

Question 20

How are word boundaries identified (by infants and computational models)?

Answer 20

+ Isolated words: Words that are said without any words before it. This doesn’t happen often and only makes up 9% of the utterances (for English learning infants).

+ Utterance boundaries: The beginning and end of an utterance can guide word segmentation
-> The beginning of an utterance is the beginning of a word, the end of an utterance is the end of a word

+ Phonological cues

• Phonotactics (What combinations of phonemes are or aren’t used)
• Allophonic variation (Multiple variations of the same phonetic unit)
• Prosodic cues

+ Statistical cues: Regularities in syllable sequence found in speech

Question 21

What is “word segmentation”?

Answer 21

Identifying word boundaries in a sentence.

Question 22

What is “ transitional probability”?

Answer 22

The probability (in a certain language) that a given phoneme is followed by another phoneme.

Example: big ripe apple or bi gripe apple

• On a word level: what is more likely to come before apple? Ripe of gripe
• > P(apple|ripe) > P(apple|gripe)
• On a syllable level: how often is syllable x followed by syllable y?
• > P(ripe|big) > P(gripe|bi)

Question 23

Are the transitional probabilities within words more predictable or those across word boundaries?

Answer 23

Those within words are more predictable than those across word boundaries.

Question 24

What is the general strategies of word segmentation? Give some examples of statistical metrics that can be used to calculate the probabilities.

Answer 24

• Calculate the transitional probabilities between phonemes/syllables
• Insert a word boundary at each local minimum

Examples of statistical metrics that can be used to calculate the probabilities:
▪ Mutual information
▪ n-gram frequencies
▪ Boundary entropy

Question 25

Another algorithm that can be used to determine word segmentation for a string of phonemes:
Case Study: Harris (1955). Turn over to see how the algorithm works.

Answer 25

• Measure the number of (possible) sucessors of a subsequence within the phoneme string
• At what point can a subsequence be follow by a high number of phoneme types?
• This point is the word boundary; a word can be followed by many other words, but the probabilities within the word boundaries are more limited

Question 26

What do Models of Grounded Language Learning have in addition when compared to usual computational models for language acquisition? How can that be modeled?

Answer 26

In addition to phonemes in a language, human beings usually have visual information that give us hints about the context and
meaning of the utterance.

We can model this by combining an image model and a speech model, and the outcome is stored in a joint semantic space.

Question 27

One connectionist model studied in class is the one by Elman (1990)
The input was an artificial string of letters
○ This string is taken from an artificial language, existing of three words: ba, dii and guuu
○ These words can follow each other in a random order

The model is trained to predict the next character
○ The model is good at predicting vowels, but bad at predicting consonants. WHY?

Answer 27

All 3 words start with consonants, and as the model can only guess the next character, the first letter is going to be chosen randomly. Predicting vowels is way easier, as the computational model knows what follows after being sure about the first character of the word. He simply has to use his knowledge base.