Infant artificial language learning and language acquisition

Rebecca L. Gómez and LouAnn Gerke

Infant language researchers have begun by examining four aspects of the language learner’s task. The first involves identification of word-like units in speech. The second involves encoding and remembering the order in which words occur in sentences. The third involves generalization of grammatical relations. The last involves learning at the more abstract level of syntactic categories (e.g. determiner, adjective, noun and verb). This fourth sensitivity is at the root of our unique human ability to produce and comprehend novel utterances.
[Sam: The fundamental flaw to this approach is that we are taking an adult approach to understand how infants acquire languages.]

How does our growing understanding of infant learning abilities bear on the highly constrained language learner described in the introduction? We can identify at least three ways.

First, all of the artificial-language-learning studies discussed have examined infants’ sensitivity to linguistic form in the absence of semantic content. In so far as these studies are tapping sensitivities used in real-language acquisition, they challenge many accounts in which language development is driven by a mapping between meaning and form. This is not to say that learners do not ultimately need to map the syntactic forms they encode during infancy onto meaning. Obviously they do. However, the fact that infants are able to acquire certain aspects of form prior to acquiring the meaning of these forms changes the nature of the language acquisition problem in a fundamental way.
[Sam: Baby's world of language has no meaning. They are just playing with sounds.]

A second implication of the research on infant artificial language learning concerns the specificity of the constraints on the learner. On many accounts, these constraints have been construed as being language specific, such that for every aspect of language to be acquired, the child is born with a specific constraint or parameter that guides him/her to the correct representation. Data showing that infants can use transitional probabilities to segment grammatical tone sequences contrasts with this view, suggesting that they apply statistical learning to linguistic and non-linguistic stimuli alike.The application of statistical sensitivity to the problem of word segmentation is admittedly far from the constraints discussed by linguistic nativists (involving such language-specific notions as whether or not declarative sentences in a particular language must have an overt subject).
[Sam: Chomsky has over complicated the issue by imposing meaning when baby is just playing with sounds.]

A third implication of both the infant artificial-languagelearning studies reviewed here and the myriad studies of infant language perception preceding them concerns the relevance of children’s early utterances as evidence for theories of language acquisition. One of the key observations of linguistic nativists involves errors that children do not make. As noted earlier, children never erroneously transform a statement like ‘The man who is tall is Sam’ into a question like ‘Is the man who tall is Sam? ’ The lack of such errors, along with logical arguments concerning the poverty of the stimulus, have been taken as evidence that children never consider rules based solely on linear order in sentences. Although researchers have begun to address the question of a how a statistical learner might begin to negotiate impoverished input, it is equally important to note that if the studies of infants’ early linguistic abilities tell us anything, it is that they have become sensitive to many aspects of linguistic form a year or more before they ever begin to produce multiword speech. This is not to say that all of language is acquired by the age of 12 months. However, if infant language-perception studies have one theme, it is in demonstrating the extremely complex (and often contrasting) relationship between aspects of their native language infants and young children have tacitly discerned and those they actually produce. Thus, we must exercise caution in interpreting children’s early utterances as evidence for or against the linguistic representations they do and do not entertain.
[Sam: the research has overlooked the different stages of brain development.]

A final comment is in order. Given the vast differences in artificial grammars and natural language, how do we ensure that the learning observed is representative of language learning in the real world? First, in using this approach it is important to design experiments capturing key linguistic phenomena. If we can isolate a phenomenon of interest experimentally, we can go on to test it using a wide range of manipulations, where, presumably, such manipulations are driven by our knowledge of natural language acquisition. For instance, the finding that 18-month-olds, but not 15-montholds track grammatical dependencies separated by one to three intervening syllables, suggests that we should see the same pattern with an artificial grammar designed to investigate such learning Indeed, studies in our joint laboratories show that we do. Another approach, currently being investigated by Saffran and colleagues, is to test whether the output of statistical learning can be used as input to natural language.

Ultimately, however, as with any scientific endeavor, the proof of this approach will depend on the extent to which it generates new ways of understanding the mechanisms involved in natural language acquisition. Its real promise lies in the precision it affords with respect to investigating infant learning.

http://www.ehu.es/ehusfera/neurolengua/files/2011/03/Gomez-and-Gerken-2000-TICS.pdf