Chapter 2. The Simulations
were enhanced by 10 (see table 2.16). To calculate the new frequencies, the following formula was used:
f i = x i + 1 0 ∗ x i
w h e r e f i i s t h e n e w f r e q u e n c y , x i i s t h e o l d f r e q u e n c y a n d m i s t h e l o w e s t o l d f r e q u e n c value (0.177 in position 1). Note that these new frequencies have no relationship to any real data of fixation positions while people are reading. It is purely introduced to enhance and examine possible difference between the two networks. To reduce training times, the training corpus consisted of 200 words (1000 events) taken from the original corpus chosen at random, corresponding to about 10% of the original corpus. Again training was stopped after 2750 epochs, although tests on the models were done throughout the training to monitor the progress of the networks. y
only made errors on 10 events, pronouncing ’boor’ as /bUrr/ instead of /bU@r/ at all
positions and ’chic’ as /SiIk/ instead of /Siik/ at all positions.
Both are low frequency
words having a frequency of 0.050 and 0.080 respectively. got 99% of the events correct.
The control network thus
The fixation net made errors on the same two words, pronouncing ’boor’ as /bUOr/ and ’chic’ as /SIIk/ at all positions. Furthermore the fixation net makes one more error on one other word, the word ’pint’ (freq 0.145) pronouncing it as /pEInt/ instead of /p&Int/ in all positions. Thus, in total, the network had 15 events wrong, which corresponds to 98.5% correct. The reason why the fixation net gets ’pint’ wrong might be the same as the reason why it got the homograph ’read’ wrong (see section 3.2.6). Similar words in the smaller corpus all pronounce /pin/ as ’pIn’ (e.g. pins). The end letter ’t’