Discriminative word learning is sensitive to inflectional entropy

Psycholinguistic evidence based on inflectional and derivational
word families has emphasised the combined role of Paradigm Entropy and
Inflectional Entropy in human word processing. Although the way frequency
distributions affect behavioural evidence is clear in broad outline, we still
miss a clear algorithmic model of how such a complex interaction takes place
and why. The main challenge is to understand how the local interaction of
learning and processing principles in morphology can result in global effects
that require knowledge of the overall distribution of stems and affixes in word
families. We show that principles of discriminative learning can shed light on
this issue. We simulate learning of verb inflection with a discriminative
recurrent network of specialised processing units, whose level of temporal
connectivity reflects the frequency distribution of input symbols in context.
We analyse the temporal dynamic with which connection weights are
adjusted during discriminative learning, to show that self-organised
connections are optimally functional to word processing when the
distribution of inflected forms in a paradigm (Paradigm Entropy) and the
distribution of their inflectional affixes across paradigms (Inflectional
Entropy) diverge minimally.