Greg Detre
30/10/00
morphology � grammar that builds words out of pieces (morphemes)
acoustic signal � phoneme � syllable � morpheme �word-phrase � sentence
connectionist vs symbolic
associative memory/network
interconnected
stems + inflections are associated with each other (not necessarily in whole form � perhaps only phenomenologically) � these links are strengthened by usage
connectionist � no two routes � homogenous
regular and irregular are treated identically
when adults irregularise, it might be because the novel verb might trigger a similar-seeming irregular verb�s response
rule application � impervious to phonological form
2 types of frequency: type/token
type depends on form
token = �instance�???
they only over-regularise only about 5-10% of the time
but as �sang� becomes represented stronger through � exposure, they make fewer mistakes
connectionist account explanation of over-regularisation
the first verbs they learn are irregular
then there is a critical mass when the regular > irregular
so it learns to apply the rule everywhere
Marhcman 1997 � phonological significance of regular verbs
children more likely to get correct the verbs with many similar regular verb forms � but she didn't control for similar-sounding irregular verbs (�enemes�???)
�= an argument for connectionist
Pinker 1999 regular verbs that sound like irreguar enemes are added to memory
time taken for adults to inflect regular/irregular verbs
parallel � search for irregular through associative memory
and apply regular rule
hence occasional over-regularisation in conversation because the irregular didn't get a chance to block
Pinker �irregulars can never be faster�
why are new forms regularised?
because higher probability (86%) are regular (only 180 verbs)
minority default is not salient������� = either probabilistically highest/novel forms default
German � plural is minority �s�
Arabic plural � sound is broken (31 common types)
Plunkett + Nakisa � counted 800 Arab random verbs, showed which form they learn + generate
Raid + Farah 99 � Palestinian Arabic
Nakisa, Plunkett + Hali(???) � plotted nouns in phonological space
irregular plurals cluster, whereas the regular are more scattered
Clarksen doesn't give details � very iffy
connectionist modelling = finding implicit probabilistic rules, but not psychological rules
connectionist model vs account???
a model has specifics and implementation � weight, synapses etc.
account = whether any connectionist model could do whatever in principle
how they claim the mind works, not just man-made NNs
91 Plunkett model didn't have u-learning
93 incremental learning, u-learning
dual route � Wernicke/Broca, double dissociation
mind has an innate linguistic/specific device for extracting the rule
wherease connectionists prefer to rely on universal cognitive rule-finding mechanisms for extracting systematicities in language
why focus on inflectional morphology (rather than derivational)
inflection = change of grammatical class