Bloom, Paul;
How Children Learn the Meaning of Words
MIT Press, Cambridge, MA, 2000
ISBN
topics: | cognitive | development
1: FIRST WORDS It looks simple. A 14-month old runs after the family dog, smacking it whenever it gets close. The dog wearily moves under the table. "Dog," the child's mother tells her, "You're chasing the dog. That's the dog." The child stops, points a pudgy hand at the dog, and shrieks, "Daw!" The mother smiles: "Yes, dog." - Bloom 262 The first word you learn can be - a specific person - Fred, - an object kind, e.g. dog, - a property, e.g. hot - an action e.g. hitting but can't be - determiner - quantifier - modal, - conjunct - preposition - a verb like dreaming - adj such as former - noun such as mortgate The latter require language in order to be learned - we hear them in the context of sentences and use this linguistic context to figure out what they mean. [one can also learn words related to internal states - e.g. "ouchie"... (pain)... can it be the first word? also - how does one learn internal experiences - e.g. "dream".] Word LEARNING in children - ostensive naming / carefully designed age-appropriate examples. [Miller and Gildea 87] Beyond that, a lot of learning occurs in linguistic contexts. WORD LEARNING AND THEORY OF MIND (Chapter 3) - AMR Word learning is a social act. When children learn meanings of words they are learning something about the referential intent of others, they have a theory of other minds. Bloom reviews the evidence showing that word learning in children draws extensively upon their understanding of the thoughts of others, on their theory of mind. Theory of mind underlies how children learn entities to which words refer, how they intuit how words relate to one another and understand how words can serve as communicative signs. The Associative infant: Central aspect of word learning - which objects specific words refer to. Problems: some words refer to nothing (of. The); others refer to things that are not objects (joke, number) and those that do are beset with the gavagai problem. Associationism: a popular solution to the problem - in Psy and phil, defended by JS Mill, David Hume, J Locke - the mechanism for word learning is a sensitivity to covariation, one rooted in general principles of learning. If two thoughts occur at the same time, they become associated, one gives rise to the other. Children learn the meaning of the word dog, then, because, the word is used when they are observing or thinking about dogs. As such the word and the thought become associated, and the child can be claimed to have learnt the meaning of the word. Associative learning is what is at work in connectionist models of word learning - network trained to associate labels with objects/images. Word learning is best explained through connectionist architecture that is sensitive to statistical regularities in the environment. The CLA fact consistent with associative learning: Early words are perceptually grounded. Generalised, universal psychology. Problems that Bloom points out: input noise: associationist procedure requires that right correlations be present in the environment. In case of word learning - it entails that the words are presented at the same time while children are attending to the referent objects/actions. Even in the best case, object names, this is not the case. In parent children interactions, 50 percent of the time a word is used, children are not attending to the referent object (results from an empirical study). The word cookie - face of an adult - this association almost never happens. Solution: an associative correlation does not need a perfect correlation - a statistically reliable one would do. But the proposal predicts frequent mapping errors. Do children wait till a hypothesis has received some level of statistical support? No. The fact that object-name acquisition is generally both fast and error free suggests that it may not be a form of statistical learning. Cultural variations in how object names are present in the input - explicit labeling, negative evidence in some cultures, but absent in others - word learning does not suffer as s a consequence. Abstractions : one might restrict associative learning to early periods in acquisition, but even in these periods the measure of words referring to abstract concepts tends to be quite high. Nelson et al studied 45 twenty-month olds and found that only half of children's nominals referred to the basic-level object kinds, the rest referred to members of other conceptual categories such as locations (kitchen, pool), actions (kiss, nap), natural phenomena (rain, sky ), social roles (doctor, brother), temporal entities (morning, day). Furthermore, despite their impoverished perceptual experience, blind children learn words, often at the same rate as the sighted children (Landau and Gleitman). Nonhuman primates, richly endowed with motor and perceptual systems and associative mechanisms, are abysmal with word learning. A rejection of associative mechanism doesn't imply a rejection of connectionism. Augustinian infant: children learn meanings of words through their understanding of the referential intentions of others. Word learning is a species of intentional reference or mind-reading. Augustine - out of currency, Wittgenstein's criticism - begins Philosophical Investigations ridiculing Augustine - that children know objects and people before they know language, the mistake is that the child could already think, only not yet speak. The idea is taken seriously by Bloom. Mind-reading,, social cognition, pragmatic understanding, theory of mind - varios ways of referring to this notion. (false belief task - children less than three do not have the ability to reason about false beliefs - keeping this out) Autism believed to be a disorder which affects other-minds ability - affects lg also. What understanding do children have about the minds of others? Sensitivity to what the adult is attending to. By 9 mths children are already following the caretaker line, and the pointing gestures, monitor parental emotional reactions to potentially dangerous situations and react accordingly. Raise the Q - what goes on when the human child follows the gaze of an adult? Is it that it has an implicit assumption that the adult is attending to something, or thinking something, or reacting to something? Gaze following could then be a reflection of a theory of mind. Alternatively, it might be a product of an automatic orienting procedure, innate or learned that is initiated by exposure to certain stimuli, such as eyes and faces, and has nothing to do with intentional attribution. Theory of minds does not preclude the importance of perceptual exp, what it does is to account for how the perceptual info is used. Baldwin expts: the babies were given one object to play with while another object was put into a bucket in front of the expter. While the baby was looking at the object in front of him the expter lookes at the object in the bucket and uttered a new word --- this is a "modi". 18 mth olds do not take Modi as the object they are playing with, they direct their gaze to the expter, and redirect their attention to the object in the bucket that the expeter is looking at and later when presented with both the objects and asked to find Modi they pick out the expter object rather than their own. Even more complex versions of the experiment produced similar results - Conclusion - contiguity between word and percept is not necessary for word learning. Further work suggests that contiguity is also not sufficient. In one version of the expt, the child picks up a toy and hears a disembodied voice - Dawnoo, there's a dawnoo. The children did not learn the word. Thast is even under perfect contiguity between hearing a word and seeing an object, the child makes the connection only if he has some warrant to believe that it is an act of naming, and that happens only when the speaker is present. Adults in a similar situation would learn the word, but not because they are more associationist than the child, but because they infer the disembodied verbal behaviour to be an act of naming. Tomasello and colleagues have found that older children show a more subtle appreciation of intentional cues. 24 mth old, lets find the toma, where is toma. Have five objects in buckets, Under no search condition bring out an object and hold it up with an excited look and say: Ah, and hand the object to the child. Under the search condition, take out one, scowl, put it back, take out another and do the same and then on taking out the third hand it over to the child. Then rest of the objects are also examined and later on, when shown all the five objects and asked to find toma, children perform equally well in both the conditions, picking out the object that the expter seemed happy with - success in this task has to with children's sensitivity to what the adult goal is and when it it is satisfied, as indicated by cues such as expter's excitement and handing over o f the object to the child. Another set of studies has to do with verbs - I'm going to plonk the big bird, the expter would perform an action and say `there!' and then perform another action and say `whoops!' - intended to give the impression that the first action was intentional and second accidental (the order reversed in another expt). The two year olds were sensitive to the emotional cue of intent. Later when asked to plonk big bird, they tended to imitate the intentional action. *Word learning with and without the theory of mind*: Theory of mind as the criterion for word learning makes predictions about how certain disorders should affect the course of word learning. Autism: a cluster of deficits, impairment in socialization and communication, could be accounted for in terms of an impaired, delayed, or nonexistent theory of mind (Baron-cohen, Leslie and Frith 1985). Happe 94, the lg impairments in autistic people are due not to linguistic deficits, but the impaired theory-of-mind abilities which are crucial to normal lg learning. Autistic individuals differ profoundly in their lg abilituies. On one extreme are those with no lg - about 30% of those labeled as autistic fall into this class. Appear deaf initially, since they fail to orient to speech, produce odd vocalizations sometimes not resembling speech or babbling. On the other extreme are those who come to talk, Like `little profs' have perfect syntax, prosody might be bizarre (monotone or sing-song) and so is their pragmatics - a highly literal lg. But their vocabularies are claimed to be normal (very liberal criteria for normal). Falling between the two extremes are a majority of autistic children who have some limited language skills: echo back words and phrases, immediately or after a delay; show pronominal reversal using I for you, or the opposite, and use entire phrases in a parroted way, like do you want a biscuit to mean I want a biscuit. Words and phrases are used in a simple associative way so that Apple might mean give me an apple - associative learning mechanisms at work. Baldwin expt on autistic children - the discrepant looking paradigm - a purely associative learning mechanism would lead the children to map the new word to the object they are looking at, whereas a theory of mind based learning process would lead them to map the word what the experimenter is looking at. Autistic children made associatively based mapping errors while normal children and children with other brain disorders did not. Supports the view that these children's ling deficits are due to their impaired theory-of-mind; they lack the inferential capacities that come naturally to normal children younger than two. The above is for the middle group of autistic people. What about the extremes - some show no lg and some show surprising success with lg -- autobiography of Temple Grandin, an autistic. (Holmes) The extent of lg deficit might be an outcome of the extent of other minds deficit - a sufficiently mild theory of mind deficit might leave the lg ability unimpaired, although such people may still have problems with non-literal modes of interpretation. An alternative theory of mind view which is consistent with modularity thesis: Sperber97 - assuming that children's ability to understand the communicative intentions of others underlies word learning, such ability may not be a general cog ability, but something very specialized, specific to lg. Week view - theory of mind in ling communication is the same ability which is at work in nonling communicative situations. If a thery of communication were distinct from a theory of mind, then one would expect autistic children to have problems with understanding people's intentions but perform well with lg. - such cases do not exist. William's syndrome children provide an interesting contrast: severe deficits in cognitive skills like number and spatial cognition, but lg is relatively spared - vocabulary size of of WS children is typically closer to their chronological age than to mental age. Karmiloff97 - development diffs between WS and normal children, but the outcomes are similar, and very different from autistics. WS - highly social, with a fully functional theory of mind. If autistic children show how an impaired theory of mind can lead to lg deficits, WS show how a preserved theory of mind can lead to full fledged lg, despite other dysfunctional systems. Nonhuman Primates: What they can and cannot learn is a controversial subject, but some facts appear to be clear: The signal systems of primates in the wild such as vervet monkeys consist of small fixed number of signs with determinate meanings, individuals do not create new words and do not teach words to their offspring (Cheney and Seyfarth 90). And no matter how supportive the environment, they do not learn words as human children do. Even by the most enthusiastic claims, chimps who are trained to use lg come to acquire no more than 200-300 signs after years of intensive training (Savage-Rambaugh93). Why are chimps bad? Dedicated lg faculty could be one answer - they have not evolved thids faculty. But if we assume no special faculty for word learning, then how to answer this. The other answer - theory-of-mind. Tomasello98: they lack an instinctive understanding of the referential intent of others. Tomasello sums up the species difference between children and chimps: `Children use symbols, whereas other primates use signals.' Under this view, chimps are somewhere in the position of the extreme autistic - it never knows how words work. However chimps do have ceratin theory-of-mind abilities - social structure for example, but T's hypothesis is interesting in a couple of ways: 1)Chimps have fine perceptual and motor skills, they are fine with associative memory, theory of mind is one area, other than lg itself, in which they are manifestly lesser than children. 2) As T notes, Chimps communicate to regulate dyadic interactions, such as play and mating, but they do not communicate about other entities - they do not refer. This limitation in theory of mind extends to non communicative situations as well: for example, they do not imitate the goal directed behaviour of others - very diff from a human without lg, a child or an aphasic adult for example. This may be why it can never fully engage in the process of word learning. Is theory of mind the whole story? Is word learning only about intentional inference? Perhaps not. Acquisition of a concept associated with the word is needed. What theory of mind does is to enable the child to establish the mapping between a word and a concept. But this presupposes the availability of the concept. And there is still the gavagai problem. Pointing and gaze will not help here. Sensitivity to syntactic cues might be a way out - is gavagai a proper name or a common name? Default expectations - whole object bias - but this may be mistaken. So what does theory of mind accomplish really in an account of word learning? One way to look at the issue is to say that inferences about the referential intentions of others enable the child to create arrows or pointers from a word to the world. The child hears rabbit and the direction of the speaker's gaze tells him that it points to the rabbit. The arrow is going from rabbit to rabbit. This is the first prerequisite for acquiring word meaning. But the arrow has touched an object in the world. It hasn't touched a concept or a meaning. The child still has to figure out whether the word rabbit refers to the kind, individual or property. Even with a theory of mind the child's problems have only begun. --- LEARNING NUMBER WORDS (Chapter 9) Racoons can learn to choose a box containing a certain number of items (e.g. three), even when non-numerical factors such as size and density are controlled (Davis 84). Irene Pepperburg (1987) reports that Alex, an African grey parrot, can be presented with arrays of upto five items and when asked "How many?" will usually squawk the correct number word in reply. A chimpanzee who was taught to read Arabic numerals was able to determine the sum of two sets of oranges and pick out the numeral it corresponded to (Boysen 89). Human babies have similar capacities. Seven-month -olds who are repeatedly presented pictures containing three items will show increased interest when shown a new picture with two items, and vice-versa (Starkey and Cooper 1980). Six-month-olds can distinguish two jumps of a puppet from three (Winn 96), two sounds from three (Starkey 90), and two moving b&w grids from three (von Loosbroek 90). Five month-olds will look shown one object removed or added to another, will look longer, indicating surprise, when the results is not correct. (Wynn 92). ... children might think of numbers as demonstrative pronouns, akin to this or that: now two refers to to the cup; an hour ago it was the og, and yesterday it was one of my fingers. [Despite their propensity to treat words as object names, somehow] children never make this mistake. ... Two year-olds can count upto three or more (Gelman 78), and once they can count objects, they can also count actions and sounds (Wynn 90). By the time children are about four years of age, they begin to appreciate some of teh generative rules underlying the language of numbers (Fuson 82, Siegler 82). [interesting overgeneralizations in number grammar - e.g. 99 as "tenny". (Ginsberg 77) ] [(Wynn 92a) - basis of this theory] 2.5 year to 3yo showed pairs of pic cards, some with a few pics, some with one. Even the youngest children could point correctly when one of the pics had only one item and the other several. After abt 3.5 they could tell four from five. ] So for about one year, they knew that two is a number word, but that it does not refer to three. John Stuart Mill (1843): Children see a scene, perceive its numerosity (such as twoness), hear a word used to refer to the set (such as two), and come to link this perception with what they hear. ... [Frege (1893) pointed out an objection] - individuation - "How many what?" two cats - but also four eyes, eight paws... [Another problem, which Mill himself realized - was with larger numbers. Also - how does the Mill theory] account for the lengthy stage when children can clearly distinguish between two and three entities, know that two and three are number words, and yet do not know precisely what two and three mean. [Rachel Gelman and colleagues (78, 86) - children possess an innate counting principle - one-to-one correspondence; stable-order principle, and cardinality principle.] Three refers to three-ness because it is the final word in a count of three items. (Wynn 95) expt: Child asked to give Big bird five animals. Gives a handful. Can you count and make sure there are five. Child: (counting perfectly): One, two, three - that's five! Experimenter. No. I don't think that's five. How can we fix it so there are five? Child: (Switches the positions of the three items). There. Now there's five. Linguistic cues: - number words ordering to rel to other words in sentence - closed-class morphemes they co-occur with - count mass status of the nouns they modify - they never occur with modifiers [The last is claimed to be particularly strong.] "Given the high frequeny of mass noun usage, it is reasonable to infer from the fact that number words are never used with mass nouns that they cannot be used with them." [STRONGLY PROBABILISTIC STATEMENT; strongly in the language-by-learning model. Yet it is also Chomsky 1981's "indirect negative evidence" stance.] Meck and Church 83: neural Accumulator model - neurons accumulate pulses. [later - p. 25 - Stanislaus Dehaene (97): approximate accumulator - as in a rat - precise for 1, 2 and 3.] But beyond that point, the number line vanishes into a thickening fog.... This fuzziness would befuddle any attempt at a monetary system, much of scientific knowledge, indeed human society as we know it. There is nothing that babies possess relating to number that is not present in other primates, even in rats and birds. [Chomsky: humans (children) innately know that you can keep adding numbers indefinitely - concept of infinite] [But children do not have the innate knowledge of generativity (counting) in numbers.] How then do humans come to possess a generative understanding of number? One proposal is that the generative nature of human numerical cognition develops only as a result of children acquiring the linguistic counting system of their culture. [Not all cultures have a language for numbers]. When children are exposed to the language of number, it causes a dramatic restructuring of their numerical knowledge. ... Peter Gordon(93) suggests that members of the Piraha tribe, which uses only a "1-2-many" number system, have problems with nonlinguistic tasks involving numerical concepts greater than three. If this is true, it would be striking evidence for the role of language in numerical thought. [But how well was it tested without using language?] Some Australian languages have a base 2 number system: 1-4 = gamar, burla, burla-gamar, burla-burla). Aztec / Gaelic languages have base 20 (e.g. French quatre-vingt). [IDEA: Possibly points to a fundamental non-generative origin for all language, rather than the more complex explanation that generative structures develop. The fuzziness of the accumulator is also an important pointer to the non-discrete, qualitative nature of number that we carry with us all the time, and use also in expressions such as "The program starts at 6:30"] -- LINGUISTIC CONTEXT: Chapter 8 A few months ago, I learned the word hobbledehoy. I came across the word in a passage where a disapproving father of the 1930s tells his son why he is not to meet a schoolmate (Amis 94): "Anyway, to me at least, in one way and another he looks a bit of a hobbledehoy. Also, to me just now over the telephone, he sounded a good deal of a hobbledehoy. Do I make myself clear?" Syntactically, hobbledehoy is a garden-variety count-noun - so from a grammatical viewpoint it could refer to anything from pickles to phonemes. The precise understanding of the word's meaning emerges instead through a sensitivity to the meaning of the passage as a whole. One can tell, for instance, that hobbledehoy refers to a kind of person that it is a bad thing to be. Children learn better from ostensive naming and from carefully designed age-appropriate sets of examples [GA Miller and PM Gildea: How children Learn Words, Sci Am 1987, v.257 p.86-91]. But not all childre get this sort of teaching, and even those who do learn at most a few words a week from it (Nagy and Herman 87: Breadth and deepth of vocab acq]. When it comes to explaining how children end up learning thousands of words each year, the only possibility is that they learn most of them through linguistic context (Sternberg 87: Most vocab is learned from context). [To build an adult vocabulary of 60,000 words, children must learn an average of 10 to 20 words per day between the ages of 18 months and 18 years. - Miller MM 369-371] Nagy and Herman: even students who read relatively little, and only during the school year, will read about half a million words a year and be exposed to about 10K words a year that they do not know. This is many times more than what they would be exposed through conversation since conversation is often between children the same age who have roughly the same vocabularies. Sternberg 87: Three processes for learning from reading context: A. Selective encoding - distinguish relevant and irrelevant information - e.g. focusing on hobbledehoy, relevant: father is disapproving; irrelevant - "over the telephone". B. Selective combination: combining the cues into a workable meaning C. Selective comparison: relate new information to background knowledge. e.g. some idea of what qualities parents disapprove of. [NOTE: This "background knowledge" is the same cloth from which any other word meanings are made.] Sternberg: Strong correlation between vocab and IQ because the process of inferring the meaning of new words are the same as those general learning abilities tested by psychometric IQ tests. [Beals and Tabors 95]: Children are often exposed to rare words, e.g. at mealtime. Mother: You have to wait a little while so you don't get cramps. George (age 4): What's cramps? Mother: Cramps are when your stomach feels tight, and it hurts 'cause you have food in it. Doug (age 4): Can I have an ice cream sandwich please, Mm? Mama please can I have an ice cream? Susan: Just a minute! Someone scarfed the last ice cream sandwich, right? Doug: Oh. Susan: How about cookies? Doug: Tammy, can I please have one of your twisters? Tammy: That's the only one. Gary had two. Mother: What's the matter? Is this the great ice cream debate? Words like vacation and dreaming can only be learned through linguistic context. But the precise nature of this learning is a mystery. I think it will remain a mystery for a long time. Roger Brown 57: Linguistic Determinism and the Part of Speech: Three groups of 3 and 4 year-old children were shown the same picture (novel object in novel action on novel stuff) told: In this picture you can see sibbing. In this picture, you can see a sib. In this picture you can see sib. i.e. - the sentences used verb syntax, count noun syntax, and mass noun syntax. The children were shown three pictures - and they were then asked to "show me another picture of" sibbing / a sib / sib, and they tended to construe the verb as referring to the action in the picture, "a sib" as an object, and "sib" as the substance. e.g. From the sentence: John learned to drive a zoop: Can tell you a lot about what zoop refers to. The same skills are needed in metaphoric transfer - e.g. In his nightmare, John saw himself driving a spoon. where spoon now acquires properties of a car. But is it the syntax or the semantics that children are learning from? E.g. Fisher, Hall, Rakowitz and Gleigman (94): give a syntactic account for 3 and 4 year old's understanding of the contrast between: The bunny is zorking the ball to the elephant The bunny is zorking the ball from the elephant But [Pinker 94a, Lingua how could a child use syntax to learn semantics] claims that the children are making this inference based on the semantic properties of the prepositions. 196 When it comes to learning at least some nominals, one can do fine without syntax. "Syntax is not necessary for at least some nominal learning." 198 But children are sensitive to syntax for example, in distinguishing count noun vs ma ss noun usage. [Leslie McPherson 91] two and three year old's sensitivities to syntax - ambiguities in "little" - which can be a mass quantifier - "a little water" - or a size modifier - "a little cup". Then they were taught a new word "vax" in mass / count noun contexts. When asked "Give me a little vax" - they then found them responding appropriately. Similar studies [Soja 92] with a pile of stuff - mass-noun usage - substance - like clay; count-noun usage - bounded collective - pile, or puddle [Bloom 94a] fep = bell sounds in rapid sequence: these are feps (count) vs This is fep; there really a a lot of fep here. (mass) when asked to make a fep / make fep even 3-year olds made single sounds vs a lot of sounds. VERBS AND ARGUMENTS Letitia Naigles 90: events with two participants vs single - a duck would repeatedly push a rabbit into a bending position at the same time that both the rabbit and duck were waving their free arms. 25 month old children shown: The duck is gorping the bunny (transitive) The duck and the bunny are gorping (intransitive) when asked then to "find gorping" between two videos, one with duck and rabbit arm waving, no pushing, and the other with duck pushing rabbit, no arm waving, two year olds looked longer at the causal scene when they had heard the transitive verb and vice versa for the intransitive. (p.202) [Naigles and Kako 93] actions need not be causal - frog touching a duck but no causal effect - learn it in a transitive context better than in a neutral frame (Look! Gorping!) --> transitively framed sentences imply object affectedness. [Pinker 89] [Cynthia Fisher et al 94] - 2 yr olds: in contexts where actions could be described with commonplace words - "the bunny is nading the elephant" - bunny was giving food - "feeding", vs The elephant is daking = eating. [Pinker 94a] claims that children can access existing lexical items via syntactic structure. But further work Fisher etal 94 show phrasal descriptions, e.g. "licking it off the spoon" instead of "eating". [Fisher 96] - She's pilking She's pilking her She's pilkingthe ball from her both 3 and 5 year olds are sensitive to transitive/intrans (syntactic) usage, as well as in vs from (lexicalized) usage in idtintifying the agent in questions like "who's pilking the ball to her?" 203 [taylor and gelman 88] new object described either as "a zav" or "a zav one" (adj) --> result in other objects of same kind being described as zav vs those with same colour / pattern / texture properties. [Smith etal 92] nouns have "same shape bias", not adjs Adjs pick out properties or subkinds? i.e. do these constitute an ontological distinction? [Wasman and Kosowski 90] - 2 to 5-yr olds are focused by nouns onto taxonomic categories. (following [Markman and Hutchinson 84]) but adjs did not [Taylor and Gelman 88] - adjs pick out properties [Waxman 90]: adjs denote subkinds; nouns pick up taxonomies - e.g. good does not pick out a subordinate category. [Bolinger 67] PREDICATION "This dog is fep" vs MODIFICATION "the fep dog" predication gives NP --> entity; modification is restrictive, it picks out a subclass of teh category referred to by the noun. [Prasada 97] - 2 to 3-yr olds more likely to give an adj a restrictive interpretation when it appears prenominally, as a modifier, than when it appears as a predicate. 204 There are reasons to doubt that children relate meaning and form through a sensitivity to statistical correlations, observing that certain forms and certain meanings just happen to go together. Correlational learning is a fine way to learn arbitrary relationships, such as shapes of English letters and the sounds they correspond to, but the relationship between syntax and semantics is not an arbitrary one. V with 1 NP arg John sleeps V with 2 NP arg John kissed the dog V with 3 NP arg John gave Fred the book Is it really possible that this one-to-one mapping between the number of NP arguments and the number of entities encoded in the meaning of the verb is learned through a sensitivity to correlations? Is it really reasonable that this one-to-one mapping between the number of NP arguments and the number of entities encoded in the meaning of the verb is learned through a sensitivity to correlations? That is, children just happen to notice that verbs with one NP argument have meanings like _sleep and verbs with three NP arguments have meanings like _give? This would suggest that children could just as well have learned the mapping the other way around and that the pattern is an accident of English; some languages should use one NP argument for give and three with sleep. But this never happens. Generalization appears to reflect a linguistic universal: an isomorphism between the conceptual structure of a predicate and its syntactic structure (Chomsky 81). It is striking support for the existence of this universal that the same relationship is also found in spontaneous communication between deaf children. SEMANTIC BOOTSTRAPPING (Pinker 84): Children use the mappings between syntax and semantics not to infer the meanings of words from their syntax but the other way around - they infer their syntactic category from their meanings. (Bloom 99, Grimshaw 81). For instance, given that dog refers to an object and big refers to a property, children can infer that dog is a noun and big is an adjective and hence, on hearing "big dog" can infer that adjectives precede nouns within an NP. Once children know this, they can use this grammatical knowledge to infer the syntactic category of words whose meaning they do not already know; "big idea" tells them that idea is a noun, for example. 208 [JABBERWOCKY example And as in uffish thought he stood ] Grammar draws relatively crude distinctions, picking out ontological kinds (such as individuals vs stuff) and subtypes of events (such as 1-participant vs 2). Word meanings are much more fine-grained. Children can learn object names, substance names, and proper names without the aid of linguistic cues. But what about verbs? [Gleitman 90 and Fisher_etal-Gleitman 94] argue that syntax plays a significant role for the acquisition of many verbs. For one, parents almost never use verbs to name actions (Gleitman 90, Tomasello 92]. In a no-audio TV adults were showin interactions between mothers and children. But everytime a word was used there would be a beep. When told to expect Nouns, they got nouns right 45% of the time. For verbs though, they got it right only 15% of the time. Gleitman and Gleitman 97 suggest that this is in part because of the poor temporal correspondence between verbs and what they refer to. 209 DIRECTLY OBSERVABLE VERBS: verbs that correspond to events or activities not directly observable are harder (e.g. "thinks"). VERBS ARE NOT UNIVERSAL: In French and Spanish, verbs encode trajectory: il descend l'escalier en courant "LIT: He goes down the stairs in running" whereas in English traj of motion is in the preposition : He is running down the stairs" PERSPECTIVE: A Giving X to Y or Y receiving a from a? Verbs also encode perspective. Gleitman 90, p. 17: "Verbs seem to describe specific perspectives taken on ... events by the speakers, persepectives that are not 'in the events' in any direct way ... since verbs represent not only events but the intent, belief and perspectives of the spearkers on these events, the meanings of the verbs can't be extracted solely by observing the events." Syntax may help resolve this [But so could the prep from / to]. Similarlty non-observable verbs may be learned from syntax: John ___ that Bill is upset is consistent with an action that takes a proposition as its subject. But verbs like dream may be as hard to teach a child as a noun such as nightmare. SYNTAX: helps the learner by providing event structure.