Language acquisition: acquiring some aspects of syntax. Anne Christophe and Jeff Lidz Laboratoire de Sciences Cognitives et Psycholinguistique
Language: a productive system the unit of meaning is the word (or morpheme) the meaning of a word is arbitrary coquecigrue 50 000 à 100 000 the meaning of a sentence is computed on the basis of the meaning of its words Le livre racontait l histoire d un chat grincheux qui avait mordu un facteur. infinite
complex rules...... shared by all speakers of a language... Pierre told Paul that he would leave tomorrow He told Paul that Jacques would leave tomorrow
complex rules...... shared by all speakers of a language... Pierre told Paul that he would leave tomorrow He told Paul that Jacques would leave tomorrow The candidate that he saw yesterday didn t call Pierre C principle: * A c-commands B if the first XP that dominates A contains B; * a pronoun cannot c-command its antecedant
and acquired very early by infants. He ate the pancakes when Big Bird was in the kitchen When he ate the pancakes, Big Bird was in the kitchen => some properties of language are innate (poverty of the stimulus argument)
The poverty of the stimulus argument Syntactic computations are complex and in a large measure arbitrary; adults speaking the same language share common intuitions; in order to learn these computations, children would have to have access to rich and detailed information (exact semantic representation of the meaning of the sentence); which is probably not the case; children, as young as you can test them, share adults intuitions; => some properties of language are innate
Other facts: language is specific to humans... All human beings speak; no other species possesses a productive communication system (e.g. bees language, etc..); nobody managed to teach a human language to another species (e.g. sign language to chimpanzees).
Language is not correlated with intelligence You can function normally but have a specific deficit for language (dysphasic children; dyslexics) you can suffer from mental retardation but speak almost normally (Williams syndrome)
Reinventing language: creoles pidgin : used by adults from different mother tongues to communicate; no articles, conjugation, etc.. créole : invented by children who learn a pidgin as their mother tongue; add conjugation, articles, etc... (e.g. hawaian créole, 1900) an instance of creolisation : sign languages: in France, 1780, abbé de l Epée in Nicaragua, 1980
Language is innate...... but what is innate? Consensus: there is a pre-disposition to acquire human languages; Debate: what is the nature of this predisposition? (Elman et al. 1996) at most : can be innate only what is shared by all the languages of the world => Universal Grammar all the rest has to be learnt: words, language-specific phonological properties, language-specific syntactic properties.
A predisposition to acquire a human language How to specify this predisposition? Linguistic principles are extremely abstract and specific (e.g. c-command); it s unclear how to represent that in genes... What is pre-determined? Among universal properties, not all are necessarily predetermined. Are pre-determined properties specific to language, or do they come from other aspects of human cognition? (e.g.: ability to compute recursively).
Do children build a hierarchical syntactic structure? Jeff Lidz Test: study the acquisition of quantifiers; Referential NP: pick out individuals Pierre, Paul, Jacques, etc... Quantificational NP: express generalizations about quantities of individuals every syntactician, no psychologist, some professor, etc...
The Grammar of Quantification Scopal Interaction: The Smurf didn t catch two birds Two interpretations: two > not : 2 birds are free $ 2 birds such that [the S did not catch them] not > two : nb birds caught 2 it is not true that [the S caught 2 birds]
The Grammar of Quantification Linguistic analysis: [the Smurf [ two birds didn t [catch two birds]]] wide scope narrow scope
Experimental Methodology Dynamic Truth Value Judgment Task (Crain & Thornton, 1998; Musolino et al., 2000) (1) Short stories are acted out in front of subjects (2) A puppet makes a statement about what happened in the story (3) The subject tells the puppet whether he s right or wrong (and explains why)
Experiment The Smurf didn t catch two birds 2 X 2 between subject: Age: 4-year-olds / Adults Conditions: W T /N F vs W F /N T
Experimental design: condition 1 Wide-true / Narrow-false The Smurf didn t catch two birds two > not : 2 birds are free $ 2 birds such that [the S did not catch them] not > two: nb birds caught 2 it is not true that [the S caught 2 birds] Make two > not TRUE There exist two birds that the Smurf did not catch Make not > two FALSE The Smurf did catch two birds
Experimental design: condition 1 two > not = TRUE not > two= FALSE
Experimental design: condition 2 Wide-false / Narrow-true Design: The Smurf didn t catch two birds two > not : 2 birds are free $ 2 birds such that [the S did not catch them] not > two: nb birds caught 2 it is not true that [the S caught 2 birds] Make two > not FALSE There is only one bird that the Smurf did not catch Make not > two TRUE The Smurf caught only one bird
Dessin expérimental: condition 2 two > not = FALSE not > two = TRUE
Experimental design Subjects: 24 4-year-olds between the ages of 3;11 and 4;11 (mean 4;4) and 24 adults Stimulus: warm up + test items below + controls The Smurf didn t catch two birds The caveman didn t ride two horses Cookie Monster didn t eat two slices of pizza The boy didn t pet two animals
One of the stories (WtNf) The end of the story The Smurf didn t catch two birds, am I right?
One of the stories (WfNt) The end of the story The Smurf didn t catch two birds, am I right?
Results: Adults NP didn't V two N: adult data 1 Proportion of YES responses 0.8 0.6 0.4 0.2 0 WtNf WfNt
Results: Children NP didn't V two N: child data 1 Proportion of YES responses 0.8 0.6 0.4 0.2 0 WtNf WfNt
Children s justifications (WtNf) Child: You re wrong, she did catch two! The Smurf didn t catch two birds, am I right?
Children s justifications (WfNt) Child: You re right! She only caught one The Smurf didn t catch two birds, am I right?
Descriptive Generalization THE OBSERVATION OF ISOMORPHISM Children, unlike adults, interpret sentences containing negation and Quantified NPs on the basis of the position these elements occupy in the surface string.
Questions Why do children differ from adults? - Linear order (precedence)? - Hierarchical relations (c-command)? Is this language specific?
Questions A is in the scope of B could mean: A occurs to the right of B (B precedes A) A occurs lower in the structure than B (B c-commands A) In the experiments described so far, these two possibilities are confounded.
c-command A B D E C F G H I x c-commands y iff the first branching node dominating x also dominates y
Questions IP NP I The Smurf I didn t VP catch two birds
Why do children differ from adults? - Linear order (precedence)? - Hierarchical relations (c-command)? These two interpretations are confounded in English; => to separate them, use a left-recursive language : Kannada
Kannada Approximately 40 million speakers in Karnataka, south-western India.
Kannada - Kannada shows the same kinds of scope ambiguities as English - In Kannada linear order and hierarchical relations are not confounded
Kannada vs. English Kannada: IP right headed English: IP left headed NP I NP I VP Infl Infl VP NP V 2 book read NEG didn t V read NP 2 books 2 book precedes neg neg precedes 2 book neg higher than 2 book neg higher than 2 book
Assessing Kannada adults knowledge 'NP two Ns Ved not: adults" Proportion of 'Yes' responses 1 0.8 0.6 0.4 0.2 0 0.87 0.85 Wide-true / Narrow-false Wide-false / Narrow-true
Kannada children s knowledge 'NP two Ns Ved not: 4-year-olds' 1 Proportion of 'Yes' responses 0.8 0.6 0.4 0.2 0 Wide-true / Narrow-false Wide-false-Narrow-true
Children s knowledge English Kannada NP didn't V two N: child data 'NP two Ns Ved not: 4-year-olds' 1 1 Proportion of YES responses 0.8 0.6 0.4 0.2 Proportion of 'Yes' responses 0.8 0.6 0.4 0.2 0 WtNf WfNt 0 Wide-true / Narrow-false Wide-false-Narrow-true
Questions: What underlies children s interpretations? - Linear order? - Hierarchical relations? Children do build hierarchical structure
General conclusions Children s failure to behave like adults reveal the structural nature of their linguistic representations. They cannot have learnt to interpret sentences in this way by observing adults, because adults behave differently
A personal comment That may seem like an extremely complex way of testing children s internal representations problem: there is no simple way. Linguists try to infer adults internal representations by asking them about their intuitions regarding the grammaticality of different sentences; the truth judgment task is another way of doing that with young children
Syntactic Bootstrapping How does verb-learning happen? Syntactic bootstrapping (Landau and Gleitman 1985) because verbs with similar meanings share syntactic distribution, the learner can use the syntactic distribution to deduce verb meaning.
Syntactic Bootstrapping act-out: 2 & 3-yr-olds extend the meaning of a known verb consistent with a new frame (Naigles, et al., 1992, 1993; Lidz 1998) the zebra falls the giraffe (frame compliant; causative act-out) the zebra brings to the table (frame compliant; noncausative act-out)
Takes on Syntactic Bootstrapping Universalist: verb-meaning to syntax mappings are principled; they (largely) don t have to be learned Emergentist: verb-meaning to syntax mappings are accidental; they must be learned by observation
Background: Universal act-out: 3-yr-olds meaning extensions limited by constraints on possible verb meaning (Lidz 1998) the zebra falls the giraffe (frame compliant; causative act-out) the zebra falls that the giraffe jumps (not frame compliant; not mental state act-out)
Background: Emergent Elicited production of novel verbs: (Tomasello and Brooks, 1998; Brooks and Tomasello 1999; Akhtar 1999) 2-year-olds saw transitive actions w/ either: a) Jim is tamming the ball b) The ball is dacking shown a new scene and asked what is AGENT doing? if they heard (a) they produced a transitive sentence if they heard (b) they didn t
The Question Are syntactic bootstrapping effects due to Universal or Emergent properties of language? Universal = a consequence of antecedently known mappings between syntax and lexical semantics (Gleitman) Emergent = a consequence of distributional analysis; item-based learning (Tomasello)
The Question: What is the source of syntactic bootstrapping effects? Universal mappings: e.g., Causativity is reliably expressed as transitivity universally, as a matter of principle. Language particular mappings emergent from distribution: e.g., Causativity is reliably expressed as transitivity in English.
The Question How can we tell? Test a language in which argument number is not the most reliable cue for causativity. Universalist: argument number predicts causative interpretation in children learning this language Emergentist: argument number is a poor predictor of causative interpretation in children learning this language
Kannada
Kannada: Causativity Why Kannada? Syntactic transitivity is not the best cue for semantic causation Morphological causativity is a better cue
Kannada: Causativity Causation is expressed morphologically on many verbs: chitte eer-utt-ade butterfly rise-npst-3sn The butterfly rises. * mosale chitte-yannu eer-utt-ade alligator butterfly-acc rise-npst-3sn The alligator raises the butterfly. mosale chitte-yannu eer-is-utt-ade alligator butterfly-acc rise-caus The alligator raises the butterfly. caus-npst-3sn
Kannada: Causativity Some verbs are underlyingly causative: * chitte ett-utt-ade butterfly lift-npst-3sn The butterfly lifts. mosale chitte-yannu ett-utt-ade alligator butterfly-acc lift-npst-3sn The alligator lifts the butterfly.
Kannada: Causativity Many transitive verbs are non-causative: mosale chitte-yannu nood-utt-ade alligator butterfly-acc see-npst-3sn The alligator sees the butterfly. huduga motte-yannu tin-utt-aane boy egg-acc eat-npst-3sm The boy eats an egg.
Kannada: Causativity Causative morpheme is a better predictor of causation than transitivity is causative morpheme Æ causative meaning transitive syntax Æ causative meaning
Kannada Predictions: Universalist: argument number predicts causative interpretation in young children Emergentist: causative morphology predicts causative interpretation in young children
Kannada Task: Act-out Subjects: 24 3-year-old native speakers of Kannada Stimuli: Grammatical and Ungrammatical sentences Argument # x Valency x Morphology
Kannada Design: 2 x 2 x 2 argument number x morphology x valency intransitive-v transitive-v 1 arg bare ok * causative * * 2 arg bare * ok causative ok *
Kannada: Causativity Crucial comparisons: Argument # vs. Morphology * chitte eer-is-utt-ade butterfly rise-caus-npst-3sn The butterfly rises. * mosale chitte-yannu eer-utt-ade alligator butterfly-acc rise-npst-3sn The alligator raises the butterfly.
Kannada: Causativity Crucial comparisons: Argument # vs. Valency * chitte ett-utt-ade butterfly lift-npst-3sn The butterfly lifts. * mosale chitte-yannu eer-utt-ade alligator butterfly-acc rise-npst-3sn The alligator raises the butterfly.
Results Main effect of Argument Number (p <.0001) No effect of Morphology (p >.83) Valency (p >.10) No Interactions
Results Kannada results: 3-yrs-old Proportion Causative Act-out 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 bare 1arg cause 1arg bare 2arg cause 2arg intrans trans intrans 0.122 0.191 0.773 0.825 trans 0.335 0.205 0.809 0.853
Results Argument # by Morphology 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 Bare Cause 0.2 0.1 0 1-Arg 2-Arg
Results Argument # by Valency 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1-Arg 2-Arg Intrans Trans
Kannada conclusion: for Kannada speaking children, argument number is a better predictor of causative interpretation than causative morphology supports universalist position question: is this a fact about children or about Kannada speakers?
Kannada Prediction for adults: argument #, morphology and valency should all play a role in determining causative interpretation.
Results Kannada: Adults Proportion Causative Act-outs 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 bare 1arg cause 1arg bare 2arg cause 2arg intrans trans intrans 0.071 0.602 0.665 0.966 trans 0.542 0.52 0.983 0.782
Results Argument Number 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1-arg 2-arg kids * adults *
Results Valency 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 intrans trans kids adults *
Results Morphology 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 bare cause kids adults *
Conclusions Verb learning is more sensitive to universal/principled properties of the syntax/lexical-semantics mapping than to language particular properties. Hence, early stages of verb learning are driven by architectural properties of UG and not by distribution alone.