Minds and Computers : An Introduction to the Philosophy of Artificial Intelligence

Figure 19.3
The standard case.

---

Consequently, we need to add another hidden unit to accommodate
double ‘s’ contexts (see Figure 19.5).
Note that we could have implemented a slightly di
fferent solution to
the double ‘s’ problem. It makes no di
fference which letter ‘s’ we inhibit
the production of the phoneme /s/ for, so long as it is only produced
once. As such, we could just as well have inhibited the firing of the
output node for an ‘s’ with a following ‘s’ rather than a preceding one.
The next three words in our test set – ‘asia’, ‘asian’ and ‘asiatic’ –
can all be accommodated with the addition of a single hidden unit.
No word in English with the letter combination ‘asia’ is such that
the ‘s’ is pronounced as /s/. We can therefore add a hidden node to the
network which detects the context ‘asia’ and inhibits the firing of the
output unit.
Exercise 19.2
Augment the network by adding a hidden unit and setting
weights and thresholds to detect the context ‘asia’ and inhibit
the firing of the output unit.
Accommodating each of the following two words in the test set –
‘is’ and ‘as’ – requires more specificity. While we want our context
194
  
Figure 19.4
Context sensitivity.

---

detectors to be as general as possible, these two cases don’t admit of
contextual generalisation. We might be tempted, for instance, to
inhibit the firing of the output unit for any context in which ‘is’ appears
at the end of a word. While this would then also accommodate ‘his’,
the network would subsequently make an incorrect determination
when presented with ‘this’. Similarly with ‘as’, ‘has’ and ‘gas’.
Consequently, we need to add hidden units which detect just the
words ‘is’ and ‘as’. This is where the utility of having a node in each
input pool for detecting a space becomes apparent. Detecting just the
word ‘is’ involves detecting the context ‘_is_’, where ‘_’ represents a
space.
Exercise 19.3
Augment the network by adding hidden units and setting
weights and thresholds to detect the contexts ‘_as_’ and ‘_is_’
and inhibit the firing of the output unit.
As well as there being numerous contexts with the letter ‘s’ such that the
phoneme /s/ should not be produced, there are also numerous ortho-
graphic contexts such that the phoneme /s/ should be produced which
do not contain ‘s’. The last two words in our test set are just such cases.
  
195
Figure 19.5
Context sensitivity.

---

Almost every English word containing the letter combination ‘ice’
is such that the ‘c’ is pronounced as /s/. The simplest way to accom-
modate the remaining words in our test set is to add a hidden node
which detects this context and excites the output unit such that it will
fire, as required.