Table 2.12: Results for the homographs in the training corpus for both networks.
Chapter 2. The Simulations
Control Network Fixation Network
The results in table 2.12 show that the networks have been able to learn one
of the pronunciations for most of the homographs. The control network gets 95.7% correct, whereas the fixation net gets only 85.7%, making 7 errors more. The details of the errors made can be found in table 2.13. As expected almost all of the errors made on the homographs were made for words where the frequencies of the two different
possible pronunciations are similar, with has a frequency of 0.408 and /lEd/ has
only a few exceptions (e.g. ’lead’ where /liid/ one of 0.235). It can also be seen from the
other. Thus for fixation position
example the control net pronounces the 1. The two possible correct phonemes for
word ’dove’ as /d@Vv/ at ’dove’ are either /dVv/ with
and the /V/ from the first output. Similarly the fixation net pronounces for example word ’used’ as /jUUzd/ at fixation positions 4 and 5. Again this is a combination of
two possible correct phonemes for /jUUzd/ with a frequency of 0.457.
An point to make is that the networks are able to learn homographs which have similar frequencies. Thus for example the two phonemes for the word ’read’ are /riid/ and