L26 - Constructionist approaches to language development

Question 1

What is the general constructionist sentiment of language development?

Answer 1

• Clearly humans are endowed genetically with the ability to learn and create language.
• But that is also true of baseball. Does anyone think there is innate Universal Baseball?
• Do we have a language/syntax specific endowment? Or does language emerge from the interaction of more domain-general cognitive mechanisms? - constructivists = we have general over specific

Question 2

What are the operating assumptions of constructionist approaches?

Answer 2

• We do not start out with any endowment of pre-specified abstract grammatical categories and phrase structure rules
• Input is extremely rich/ambiguous, despite claims of the “poverty of the stimulus,”
  
  • And thus the inability of the child to learn from it
• More powerful learning mechanisms have been explored since Chomsky refuted Skinner & Piaget such as
  1. Statistical driven learning via the structure of neural networks (as in Cohen/Smith)
  2. Analogical/relational abstraction (as in Gentner).
  3. Also social reasoning, see ToM lectures

Question 3

What is statistical learning?

Answer 3

• Hypothesis: Infants and children will track the transitional probabilities between sounds, words, and phrases, and how they are distributed more globally to learn grammatical categories and phrase structure rules - e.g. which words follow other words
• First test case: infant word segmentation and transitional probabilities.

Question 4

What is a transitional probability and how does it show statistical learning?

Answer 4

• The probability that, after a given syllable, a particular syllable will occur
  
  • Or whatever relevant unit, not just syllable
• If, after pre, the next syllable is tty 50% of the time, the TP of pre-tty is 0.5
• Helps the learner as if you hear a string such as ****prettybaby**** , you can use TPs to find the words
  
  • If you keep track of transitional probability - you know what sounds are most likely to make one word or separate words

Question 5

What did Saffran et al. (1996) do to test if children actually learn language this way?

Answer 5

• Test with artificial grammar in 8 month olds
• Only cue to word boundaries is the transitional probability information: High between syllables within a “word” (e.g. bi → da); low between syllable across a “word” (Ku → Pa) - No gaps in the stream
• Infants listened to this continuous stream of speech sounds for 2 minutes. Then their listening time was measured for:
  
  • A) Repetitions of words: “bidaku, bidaku, bidaku”
  • B) Repetitions of non-word “kupado, kupado, kupado”
• As defined by the transitional probabilities with a ½ second gap between each
• They showed a novelty preference for B

Question 6

What did Mintz (2003) find about grammatical categories and child directed speech?

Answer 6

• Grammatical categories are revealed through their distributions, analysing corpora of child-directed speech
• e.g., the ___ (is usually a noun), is__ing (usually a verb), and the “the ___” is often followed by “is __ing”
• ## Showed infants are sensitive to forming categories of words through shared distribution in artificial grammar

Question 7

What do constructionist’s claim about statistical learning and dependence?

Answer 7

• The constructionist claim is that the learning of grammatical structures can be driven via these sorts of mechanisms, tracking transitional probabilities and the distributions of words and phrases.
• Can this perspective re-explain previous findings interpreted as evidence for UG, such as the Crain’s work on children’s questions?
• Reminder:
  
  • given “The man who is smoking is crazy” children always turn it into a question by respecting phrase structure, as in “Is the man who is smoking crazy?” and not “Is the man who smoking is crazy?”
• BUT
• Whether or not children have innate knowledge of SD, such errors would not be expected to occur given (implicit) knowledge of the transitional probabilities of particular lexical items or categories thereof
• That is, no ever says anything like “who smoking” in any context, question or otherwise (besides auxiliary verb dropping dialects…).
• If children are building up their knowledge of syntax by tracking the statistics/distributions of words and categories of words, they also shouldn’t make such errors

Question 8

What is the constructionist approach to the development of question formation?

Answer 8

• Unlike for is questions, for can questions, a SD error results in a possible word (category) pair
  
  • “The boy who can smoke can drive”
  • “Can the boy who can smoke drive?
• SD error “Can the boy who smoke can drive?”
• “Who smoke” unlike “who smoking” is a grammatical pair of words in other contexts. e.g., “people who smoke die young”
• If children construct rules from complex sentences from simpler smaller units, they might make such errors.
• Ambridge et al. (2008) used an elicited production method similar to Crain & Nakayama, 1987
  
  • gave situations, made 7 year olds ask questions about them
• Unlike with is questions, where there are 0 SD errors, with can questions, they found 7% SD errors, with children ranging from 0% SD errors up to 43%!
• UG would not predict such word-specific effects. Structural dependence is a universal constraint that applies equally to all words.

Question 9

What is gradual abstraction (Tomasello,2003) ?

Answer 9

• Abstract grammatical categories and structures allow for free production of words, independent of how individual words have been used before
• However, children start by only using words in phrases they hear them are used in the input
• Slowly children start noticing commonalities in patterns of words use and generalize individual words to novel contexts
  
  • Hypothesized mechanism: Gentner’s analogical learning
• Children’s earliest grammatical structures are based on individual words, and more abstract adult-like grammars are built from there
• e.g. word that comes after “i kick” is an object
• Process of schematization
  
  • Break utterances into component parts and generalize across them to form partially-productive, lexically specific schema
    
    • Kick it + Kick ball + Kick Mummy = Kick X
  • This creates verb specific schemas.
  • Then across verb specific schemas, find the commonalities in these schema/utterances to form abstract constructions (X kick Y, P kiss Q  SUBJ, VERB, OBJ)

Question 10

What’s the evidence for this verb-specific to verb-general pattern?

Answer 10

• Schematization experiments
  
  • Novel verb studies: Teach children a novel (made up) verb (e.g.: meeking, tamming) in one construction, and see if they will use it in another
  • If so, this is evidence that they have a VERB-general rule (e.g., SUBJECT VERB OBJECT)
• Artificial word experiment
  
  • If kids have syntactic construction then they can see it in the right way that tamming is a verb
    • Verb island as not connected to others
  • When learn new nouns, will happily learn new constructions
  • Small scale schemas and eventually find all the connections between the verbs

Question 11

What is a summary of the constructionist approach?

Answer 11

• Claim 1: Children can learn syntactic categories through the distributional/statistical patterns of words in their input.
• Claim 2: Children’s initial syntactic knowledge is item- based (based on individual words) and abstract constructions are formed slowly by generalizing across such item-based constructions.

Question 12

What are early abstraction accounts?

Answer 12

• Many psychologists have abandoned theorizing that much of the rich specific structure (e.g., that VP = V +NP) is part of a language-specific genetic endowment, but they also reject the constructionist notion that early language knowledge is all item-specific - rejects Tomasello
• Specifically, verb-general notions of agents and patients, and general notions of how they are linked to nouns constrains language learning from the start
  
  • That is, each semantic role needs a noun, and each noun needs a semantic role
• From there abstract knowledge of word-order sequences is built
  
  • While word order needs to be learned, they are utilizing abstract verb-general representations from the beginning

Question 13

What are early abstraction accounts and verb learning?

Answer 13

• English has systematic mappings between syntax and semantics.
• Generally, agents become before the verb, and patients come afterwards.
  
  • The boy broke the vase I threw the ball etc.
• If children can use such verb-general cues to assist them in learning verbs, evidence against item-specific syntactic knowledge
• Comprehension task may be better tasks that production tests

Question 14

How does constructionism respond to early abstraction accounts?

Answer 14

• Dittmar, Abbot-Smith, Lieven, & Tomasello (2008)
  
  • Noticed that in Gertner et al., children were warmed up to the task with familiar verbs and similar kinds of causal events, using transitive syntax and the exact same nouns used at test (e.g., “frog” and “bear”)o “The frog is washing the bear”
• Dittmar et al. replicated their procedure and added a condition that simply did not use the nouns or transitive syntax as part of the warm up.o “This is called washing”
• Then all children had the same test, as in Gertner et al.
• In the latter condition, children did not succeed at test! They required a relevant prior exposure - no abstract syntax

Question 15

Is there a constructivist explanation for sensitive periods?

Answer 15

E.g. If one switch has a higher chance of turning on the light bulb, you start using the one with the highest percentage the most - relates to slow development of cognitive function

Question 16

What do large language models (e.g., ChatGPT) mean for theories of language development?

Answer 16

• On the one hand, old nativist arguments said that human syntax can not be learned just from example sentences, without innate constraints
• LLM’s seem to show that is wrong, but do they learn like children?- Images via Mike Frank
• The number of words chatGPT is trained on far more words than what children learn on - potential active social learning within children allows for more rapid development