Language - Computational Explorations in Cognitive Neuroscience - page 362

Gestalt Gestalt Context

(subject) and the teacher (co-agent). This syntactic in-
formation is conveyed by the cues of word order and
function words like with .
Importantly, the surface form sentences are always
partial and sometimes vague representations of the ac-
tual underlying events. The partial nature of the sen-
tences is produced by only specifying at most one
(and sometimes none) of the modifying features of the
event. The vagueness comes from using vague alterna-
tive words such as: someone, adult, child, something,
food, utensil, and place . To continue the busdriver eat-
ing example, some sentences might be, “The busdriver
ate something,” or “Someone ate steak in the kitchen,”
and so on. As a result of the active/passive variation and
the different forms of partiality and vagueness in the
sentences, there are a large number of different surface
forms corresponding to each event, such that there are
8,668 total different sentences that could be presented
to the network.
Finally, we note that the determiners ( the, a, etc.)
were not presented to the network, which simplifies the
syntactic processing somewhat.

Encode Decode

Input Role Filler

Figure 10.27: The SG model. Localist word inputs are pre-
sented in the Input layer, encoded in a learned distributed rep-
resentation in the Encode layer, and then integrated together
over time in the Gestalt and Gestalt Context layers. Questions
regarding the sentence are posed by activating a Role unit, and
the network answers by Decoding its Gestalt representation to
produce the appropriate Filler for that role.

questions are posed to the network by activating one
of the 9 different Role units ( agent, action, patient,
instrument, co-agent, co-patient, location, adverb, and
recipient ), and requiring the network to produce the ap-
propriate answer in the Filler layer. Importantly,
the filler units represent the underlying specific con-
cepts that can sometimes be represented by ambiguous
or vague words. Thus, there is a separate filler unit for
bat (animal) and bat (baseball) , and so on for the other
ambiguous words. The Decode hidden layer facilitates
the decoding of this information from the gestalt repre-
sentation.
The role/filler questions are asked of the network af-
ter each input, and only questions that the network could
answer based on what it has heard so far are asked. Note
that this differs from the original SG model, where all
questions were asked after each input, presumably to
encourage the network to anticipate future information.
However, even though we are not specifically forcing
it to do so, the network nevertheless anticipates future
information automatically. There are three main rea-
sons for only asking answerable questions during train-
ing. First, it just seems more plausible — one could
imagine asking oneself questions about the roles of the

Network Structure and Training

Having specified the semantics and syntax of the lin-
guistic environment, we now turn to the network (fig-
ure 10.27) and how the sentences were presented. Each
word in a sentence is presented sequentially across trials
in the Input layer, which has a localist representation
of the words (one unit per word). The Encode layer
allows the network to develop a distributed encoding of
these words, much as in the family trees network dis-
cussed in chapter 6. The subsequent Gestalt hidden
layer serves as the primary gestalt representation, which
is maintained and updated over time through the use of
the Gestalt Context layer. This context layer is
a standard SRN context layer that copies the previous
activation state from the gestalt hidden layer after each
trial.
The network is trained by asking it questions about
the information it “reads” in the input sentences.
Specifically, the network is required to answer ques-
tions of the form “who is the agent?,”“who/what is the
patient?,”“where is the location?,” and so on. These