Back to the previous topic
Miikkulainen |
|---|
|
Miikkulainen (1996) trained a network to receive a
sentence one word at a time, and then assign roles to
nouns and verbs in an agent-action-patient triple. For
example, the girl saw the boy got mapped into
agent=girl, act=saw, patient=boy, represented as
|girl saw boy|. His network was divided into
three subnetworks: a parsing network, a segmenter
network, and a stack network. The combination of these
three network was able to divide input sentences in
such a way as to allow it to correctly process
sentences with higher levels of phrase embedding than
ever seen during training. Once divided, phrases were
assigned case-role representations. Each subnetwork was
trained independent of the other and with no knowledge
of what the relationship between the system's input and
the system's output was (since none of the networks
see both the outside inputs and outputs). The average
unit error for the parser subnetwork was of 0.019.
Center embedding greatly diminished the percentage of
correctly remembered agents.
Because each subunit was trained independently of the others, the author mentioned that his system should be seen as a model for human performance rather than learning. He also mentions that "Other, more cognitively valid learning mechanisms may be possible", and suggested that such systems should be studied.
|
Back to the Table of Content