Week 5: Language

Poverty of Input

Theory that linguistic input (speech/language) kids here is insufficient (impoverished) to explain complex grammar they eventually master. This indicates innate linguistic capacity for language

Critical Period

A developmental threshold where language acquisition is most effective, often believed to occur during early childhood.

Why is language hard?

Homophones (words that sound same but have different meanings), idioms, word order, and synonyms.

The “textbook child”: typical developmental linguistic milestones

1 month: Cooing

6-9 months: Babbling

6-10 months: First signs of words comprehension, communicative gestures

10 months: Babbling becomes word-like; with rhythm

11-12 months: First words

16-18 months: Vocabulary grows; words are combined

24-28 months: Grammatical function words

30-36 months: Productive language: overgeneralizing grammer rules (ie: “goed”)

DeCasper and Spence 1986

Prenatal Story:

Question: Do babies learn from speech they hear before birth?

Method: Pregnant mothers read a story (The Cat in the Hat) aloud during pregnancy. After birth, newborns do high-amplitude sucking: sucking pattern could control whether the familiar story or an unfamiliar story was played.

Result: Newborns showed a preference for the familiar story (modified sucking to hear that story)

Claim: Language learning begins before birth

Mehler 1988

Question: Can young infants distinguish their native language from another foreign language?

Method: French newborns heard French and Russian speech. DV: sucking rate (high-amplitude sucking) IV: language heard

Result: Newborns sucked more to hear their native language and could tell the two apart

Claim: Newborns already learned/familiar with the rhythm of their native language before birth

Protoconversation

Definition: pre-verbal interaction between an infant and caregiver that mimics the structure of a real conversation before the child learns to speak

Types:

1. Turn taking:

   1. nfants understand back and forth interactions in conversations. Caregiver and infant alternate vocalizing or gesturing in a back and forth pattern without talking over each other

2. Pointing:

   1. Emerges around 6 months; basic form = “I want”; advanced form = “Look” (informing)

3. Babbling

   1. Infants transition from simple coos to complex babbling; word-like babbling with rhythm

   2. ** Deaf infants learn sing language babbling with their hands

      1. THEME: resilience of development

Infant-directed speech (IDS)

Higher and more variable pitch, slower rate, clearer pronunciation. Gives infants more time to process and highlights words

Cooper and Aslin 1990

Preference for IDS

Methods: 2 visual stimulus - one plays IDS ad one plays adult-directed speech; preferential looking was used

Result: newborns and 1 month-olds showed preferential looking for visual stimuli that produced IDS

Claim: Infants prefer IDS

Fernald and Kuhl 1987

Method: Isolated core prosodic components of IDS (pitch, amplitude, and rhythm) to test which feature of IDS is preferred

Result: Preference for IDS is driven by pitch

ManyBabies 1, 2020

Multi-laboratory project that tested infants preference for IDS over ADS. Scope: 69 labs, 16 countries

Question: Do infants reliably prefer IDS

Result: IDS preference is real, preference increases in older infants, higher preference when IDS speech matched native language, higher preference when using head-turn preference procedure

Theme: developmental psych is an active research area

Case study: Genie

What: Genie was deprived of language until 13 y. o. A linguist worked with her for years.

Result: She never learned typical grammar or became fluent

Claim: Early language exposure is necessary; there is a critical period for learning a first language

Theme: fragility of development

Case study: Simon (Singleton and Newport, 2004)

Creating language

Method: Simon, deaf child who’s parents learned ASL late and signed with errors. They were his only signing input

Result: Simon signed better than his parents; close to native fluency; turned inconsistent errors into consistent rules

Claim: Children don’t just copy input, they have a bias to create rules; supports Chomsky’s nativism and rejects Skinner’s behaviorist view that children learn language through imitation;

Theme: resilience of development

Three speech perception problems infants must solve

1. The segmentation problem

2. Invariance problem

3. Tuning problem

Segmentation problem

Problem: speech is a continuous stream. How do infants find where words begin and end?

Solution:

1. IDS: Prosody (pitch and rhythm), emphasis and spacing mark both mark word boundaries

2. Statistical learning: infants track statistical relations between syllables/words

Invariance Problem

Problem: how infants make sense of constantly changing sounds of human speech: phonemes are the smallest sound units that can change a word’s meaning (ie: “cat” vs “cut”); furthermore, phonemes and pronunciation vary by speaker, context, etc

Solution:

1. Formants:

   1. Formants 1 & 2 (identity of sound): what vowel you are saying, as identified by shape of mouth (tongue and rounding of lips)

   2. Formants 3 & 4 (identity of speaker): “color” and vocal identity of person’s voice, as identified by size and anatomy of vocal track

2. Vocal Onset Time (VOT)

   1. Brain stopwatch that tracks time between when lips open to let out air and when vocal cords vibrate

      1. Short VOT: vocal cords buzz immediately after lip opening —> creates “ba” sound

      2. Long VOT: longer time between lips opening and vibration —> creates “pa” sound

      3. Categorical perception: VOT can vary, but our brains will never hear a hybrid “b-pa” sound

3. Kuhl 2004 (context)

   1. Problem

      1. If speech were perfect and invariant, every time someone said "ah," it would land on the exact same coordinate on the graph. Instead, plotting multiple people saying "ah," "ee," or "oo" creates scattered clouds or clusters.

      2. This means the sound wave alone is mathematically ambiguous. If you isolated just that one vowel sound into a 1-second audio clip, a computer or a human brain would physically not be able to tell which vowel it was.

   2. Solution: Context

      1. Brain uses top down processing (grammar, meaning, and context to guess ambiguous sounds)

Eimas, 1971

Experimental proof for categorical perception in infants

Method: high-amplitude sucking with infants; infants are habituated to a sound and as they habituate they get bored and sucking rate slows; VOT either shifts across acoustic categories or stays within acoustic category and sucking is measured

Result: across-category shift was recognized and sucking rates increased bc they liked novelty; within-category rates stayed same or dropped

Claim: infants perceive speech categorically from early on

**Note: the within-category experimental group still had an equivalent physical shift as across-category, but the shift stayed within the same acoustic category

Tuning Problem

Problem: each language has unique subset of sounds it uses/cares about; infants have to “tune” their auditory system to pay attention to acoustic boundaries of their language —> but babies don’t know what language they are born into

2 competing theories for solutions:

1. Growth model:

   1. Babies start as a blank slate and are bad at perceiving all speech contrasts —> they slowly learn to distinguish ones from their native language overtime

2. Pruning/perceptual narrowing model:

   1. Babies are born with knowledge and ability to hear acoustic differences between all human languages —> overtime they prune filters they don’t use and sharpen ones they do use

   2. **evidence supports this one (Werker and Tees, 1984)

Werker and Tees, 1984

Proving perceptual narrowing

Method: conditioned head turn procedure; english learning infants vs hindi learning infants; infants listend to stream of repeating sounds (ie: “ba-ba-ba”); the computer would randomly switch to a sound with a non-english contrast (so only hindi learning infants would hear) paired with a light up toy animal on the side; idea —> if babies can hear the difference they would look to see the toy before it even lit up

Results: 6-8 month olds could tell non-native sounds apart; 8-10 month olds = only some babies could distinguish; 10-12 month olds = almost none of english babies could hear contrast, while Hindi-learning infants could.

Claim: human brain undergoes perceptual reorganization between 6-12 months. Supports perceptual narrowing —> proved that babies lose their “universal ear”

Theme: development does not equate to improvement.

Baker, Golinkoff, Petito 2006

Question: Does perceptual narrowing also happen for sign language?

Results: Hearing infants with no sign exposure could tell apart sign language hand shapes at 4 months, but not at 14 months

Claim: Narrowing happens for visual language as well

Kuhl, Tsao, and Liu 2003

Escaping perceptual narrowing —> builds on/kind of opposes Werker and Tees 1984

Question: Can exposure to a foreign language prevent perceptual narrowing, and does the type of exposure matter?

Method:

• 9-month-old American babies got 12 short sessions of Mandarin: live speaker, same speaker on video, audio only, or live English (control).

• Test (conditioned head-turn): babies hear one Mandarin sound repeat; when it switches, a toy lights up. Babies who can hear the switch learn to turn toward the toy when it happens.

Result: Only the live Mandarin group turned at the right moments (could hear the difference).Only the live Mandarin group kept the ability to hear the Mandarin contrast.

Claim: Even a small amount of live, responsive interaction keeps non-native contrasts; passive exposure does not work, needs social context and cues (ie: eye contact, pointing, shared attention)

The mapping problem

Word learning

Problem: learning a word means mapping it onto a concept (nouns, verbs, adj); each new word could refer to almost anything

Solution:

1. Joint attention and ostensive labeling

   1. Infants look and where others look and parents label what they are looking at

   2. Ostensive labeling: defining or explaining a word by pointing to it

2. Short arms

   1. Infants have short arms which helps because as they hold objects close, fewer things are in view, lowering the possibility of things the word said could label

3. Cross-situational learning

   1. across many ambiguous situations, infants track which words and objects keep appearing together

Markman and Wachtel, 1988

Mutual Exclusivity

Question: How do children decide what a new word means?

Finding: children assume each each object has only one label; so if a new word is used for an object that already has a name, they assume the new word describes a property of the already-labeled object

• Ie: if child already has a cup labeled as “cup”, if someone points to a cup and says “pewter”, the children will assign “pewter” as an adjective for the cup; (rather than thinking that cups are called “pewters"

Claim: Mutual exclusivity lets children more quickly map new words to new things

MacArthur-Bates CDI

For measuring vocabulary

What:

• A checklist of several hundred common words. Parents mark which words their child understands and which the child says. The score is compared with norms from thousands of children, so each child gets a percentile relative to same-age peers.

• Two versions:

  • "Words and Gestures" (about 8 to 18 months) asks what the child understands and says, plus gestures like pointing.

  • "Words and Sentences" (about 16 to 30 months) asks what the child says, plus early grammar.

• Strengths: cheap and fast; covers far more words than a lab test could; draws on everything the parent sees across the child's whole day; available in many languages.

• Weaknesses: relies on parents' memory and judgment, which can be biased (some parents over-report, some under-report). Judging what a baby understands, rather than says, is especially hard for parents.

Marchman and Fernald et al. 2008

Looking-while-listening longitudinal study

Method: The child sits on a parent's lap facing two pictures of familiar things, such as a puppy and a baby. A recorded voice says "Where's the puppy?" A camera films the child's eyes, and researchers code where the child looks, frame by frame, in fractions of a second.

• 2 measures:

  • accuracy: how much of the time the child looks at the named picture after hearing the word

  • reaction time: when the child started on the wrong picture, how fast they switch to the right one —> reaction time is a measure of processing speed

Result: Word recognition speed at 2 years predicts language and IQ at 8 years; Its reaction-time measure is the "processing speed" that predicts later vocabulary and IQ. Faster processing helps them keep up with speech.

Theme: reducing demands on behavior

Deb Roy's Speechome (MIT Media Lab)

Method: Cameras and microphones recorded a child's home for his first 3 years.

Result: "Wordscapes" mapped where in the house each word was said. Words used in more distinctive contexts were learned earlier.

Claim: Technology lets researchers study word learning in real life at a scale that wasn't possible before.

Syntax

Definition: how words combine into sentences. Each word belongs to a category (noun, verb, determiner), and syntax says how categories combine into bigger units.

• Chomsky: syntax (grammar) and semantics (meaning) are separate. "Colorless green ideas sleep furiously" is grammatical but meaningless.

• How children learn it: fast, by observing and talking, without being taught.

• Errors show generalization: overgeneralizing ("goed," "mouses") shows children extract rules instead of just copying what they hear.

Ferry, Hespos and Waxman 2010

Words create categories

Question: Do words help babies form categories?

Method: Infants were shown a series of pictures from a single object category (such as different dinosaurs). One group heard spoken labeling phrases (e.g., "Look at the modi!") while another group heard matched tone sequences. Their looking times were measured during test trials with novel and familiar

Result: The word group looked at the new fish and the non-fish differently; the beep group didn't.

• The "Word Advantage": Infants who heard labeling phrases successfully formed object categories (3-month-olds showed a familiarity preference, and 4-month-olds showed a novelty preference).

• Infants who heard tone sequences showed no evidence of categorization and looked at chance levels.

Claim: Words (not just any sound) help babies form categories. "Words are invitations to form categories."

• Human speech provides a unique, privileged cognitive boost for object categorization in young infants

Language and number

We have and exact number system and an approximate number system. Language adds a 3rd system: counting system; lets us track exact large numebrs

Spaepen, Coppola, Carey and Goldin-Meadow 2011

Question: Without number words, can people track exact large numbers?

Method: Tested deaf adults in Nicaragua who never learned a sign language. They use self-made gestures ("homesign"), which have no counting system.

Result: They couldn't represent exact large numbers.

Claim: Language (counting words) lets us be precise about number. Same conclusion as Frank 2008 (Piraha).

Linguistic determinism (Sapir-Whorf hypothesis)

Claim: The language you speak limits the thoughts you can have.

• Classic example: words for colors, shape which color differences people can notice.

• Current status: Mostly faded. Evidence shows language influences thought, but doesn't strictly determine it.

LLMs vs. Chomsky

Chomsky (poverty of the stimulus): children hear too little language to learn grammar from input alone, so grammar must be innate.

How LLMs learn: next-word prediction on trillions of words, plus attention (which context matters), human feedback (alignment), and images (multimodality).

Challenge to Chomsky: LLMs learn grammar, even tricky sentences, with NO built-in grammar.

• But LLMs ≠ children: - Data: LLMs see trillions of words; a 4-year-old has heard ~10–50 million - Social cues: children learn through joint attention; LLMs don't - Feedback: children rarely get told what's right or wrong; LLMs are trained on feedback

Frank et al. 2008

The Piraha

Question: Can people track exact large numbers without number words?

Method: The Piraha (Amazon) have no words for exact numbers, not even "one." They were asked to match quantities, e.g. put out as many balloons as there were spools.

Result: When the objects stayed in view (lining up the same number of balloons as spools), they matched even large numbers exactly. When the task needed memory (objects hidden or shown one at a time), their answers became approximate.

Claim: Number words are a "cognitive technology." They don't change our basic sense of number, but they let us remember exact large amounts.