Greg Detre
25/10/00
rules vs connectionist model
backpropagation � non biological plausible
distributional evidence
if you can get a connectionist system to learn language without innate language acquisition device
argument against Chomsky??? negative evidence is almost non-existent
how can we test hypotheses?
synapses being pruned away, leaves just positive evidence
argument from conservatism � child is only producing successful utterances
doesn�t work with whole sentences, though it might work just with a smaller domain, e.g. irregular past tenses
Learnability Problem: how do children recover from their errors?
part of a wider theory of learning in general
requires also working out which hypotheses to test (which structures are open to manipulation + transformation, e.g. inflection, word order)
Chomsky explains hypothesis-creation as being innately determined
but may require negative evidence to prune back incorrect hypotheses
Baker�s paradox � subset of the learnability problem
Pinker: child�s problem with the dative case
that some verbs take the dative like �give�, e.g. �the boy gave ___ the girl [a ball]�
different for �deliver�
feels like an exposure thing � we don't change closed class word usage
how does the child know which/how to use? can't possibly have tried all the permutations
could this be explained by Chomskian toggle parameters???
assuming a productive, non-conservative, no negative evidence child
don't underestimate the opwer of positive input � we learn to speak what we hear
�
Saxton � nonsense words, negative vs positive evidence for regularising
positive evidence children produced zero correct responses!!!
5 week period �/span> 40% with positive evidence(???)
negative evidence �/span> 50% correct ina� very short time
expect them to use positive evidence at a younger age (learning words, high metabolism, + synaptic connections)
children who don't speak don't get negative evidence at all
poverty of the stimulus
no negative evidence
positive evidence is under-specified and insufficiently detailed to correctly induce from surface structure � deep structure
yet by the age of 5, children speak syntactically as well as adults
learnability problem; Baker�s(???) paradox; poverty of stimulus; u-learning � all hinge around lack of negative evidence
�learning� � in Chomsky, we don't learn, we have already a rule system
that we have a system of rules is indisputable, e.g. over-analysation
what do you mean by a �rule system� � is connectionism one? it�s an implicit one, though
look at nonsense words, e.g. Jabberwocky poem
rules as preset likely/probabilistic patterns
in cultures where children don't get spoken to, siblings help
speech perception � categorical � even babies are best with phonemic variation
is it language specific? no, because chinchillas do it too
across boundaries employed by language
(but that�s because our language has evolved to take advantage of our particular auditory filtering)
parallel with phonemic contrast atrophy and legitimate grammar
what about negative evidence in the form of facial expression, nature of response
cognitive evidence is vocalised, grammatical errors might not be verbalised but still apparent � 18 month babies are desperate for feedback