Language - Computational Explorations in Cognitive Neuroscience - page 364

epoch trial EventGp

busdriver
spread
spread
jelly
jelly
jelly
with
with
with
with
utensil.
utensil.

tick
0
0
1
0
1
2
0
1
2
3
0
1

Event
busdriver_busdriver
spread_spread
spread_busdriver
jelly_jelly
jelly_busdriver
jelly_spread
with_jelly
with_busdriver
with_spread
with_jelly
utensil_knife
utensil_busdriver

cnt_se
0
0
0
0
0
0
0
0
0
0
0
0

resp
busdriver
spread
busdriver
jelly
busdriver
spread
jelly
busdriver
spread
jelly
knife
busdriver

Now press Step again on the process control panel
to present the next event.
Notice that the question asked about each word when
it is first presented is about that word itself, so the net-
work should be asked about the action of the sentence
at this point. However, the next question will go back
and ask about the agent again.

0 0

0 1

0 1

0 2

0 2

0 2

0 3

0 3

0 3

0 3

0 4

0 4

Press Step again.
You should see that the Role input asks about the
agent while still presenting the second (action) word in
the Input . Thus, the network has to learn to use the
information retained in the gestalt context layer to be
able to answer questions about more than just the cur-
rent input word. You can now step through the remain-
ing words in the sentence, observing that all previous
questions are reasked as each new word is presented.

, !

Figure 10.28: Output of the Trial 0 TextLog for the
trained network, showing sequence of questions that are asked
after each input.

that correspond to the answering of the role/filler ques-
tions — these question-level events are counted by the
tick column, and are labeled in the Event column.
The first part of the Event label is just the input word
(same as EventGp ), and the second part is the correct
answer to the current question. The actual output re-
sponse of the network is shown in the resp column,
and cnt_se is a count of the number of output units
with errors (this will be 0 when the network is produc-
ing the correct answer).
There is no reason to expect the network to produce
the correct answer at the start because the weights are
random. Thus, you will probably see that there is a
non-zero cnt_se value and the resp column does not
match the second part of the Event column. We can
now look at the network activations to see what hap-
pened in the minus and plus phases.

Press Step several more times until the EventGp
word has a period after it, and all the questions have
been asked for that last word, as indicated by the period
after the word in the Event column.
You might have also noticed that the network actually
seemed to start remembering things pretty well — this
is illusory, and is due to the effects of the small weight
updates after each question (it is basically just remem-
bering to say the same thing it just said). As we will see
in a moment, the testing that we will perform rules out
such weight changes as a source of memory, so that we
know the gestalt context representation is responsible.
Nevertheless, it is quite possible that people actually
take advantage of this kind of priming-like memory (see
chapter 9) to encode recently processed information, in
addition to the activation-based memory represented by
the gestalt context layer. Thus, this model could provide
an interesting way of exploring the interactions between
weight-based and activation-based memory in sentence
processing, but we won't explore this idea here.
After the sentence is over, the gestalt context is
cleared, and the next sentence is processed, and so on.

Press the act_m button in the network display to
view the minus phase activation states.
You should be able to use the labels on the units to
verify that the input word was presented on the Input
layer, the agent ( ag ) Role unit was activated, and the
network produced the output shown in the resp col-
umn in the Filler layer.

, !

Now press the act_p button to see the plus phase
activations.
Here, you should see that the correct answer was
provided in the Filler layer. Also notice that the
Gestalt Context layer is zeroed out for this first
event — it is automatically reset at the start of each new
sentence.

Te s t i n g

Now, let's evaluate the trained network's performance,
by exploring its performance on a set of specially se-
lected test sentences shown in table 10.15. Because the
network takes a considerable amount of time to train,