Errors with frequency ≤ 0.1
Errors with higher frequency
Chapter 2. The Simulations
Control Network Fixation Network
Table 2.4: Errors by frequency.
Table 2.4 shows the network’s performance in relation to the frequency of the incorrect words. Not surprisingly, for both networks, most errors were made on low frequency words. The control network made 129 of its 139 errors on words with a frequency lower than 0.1 and he fixation network 116 of 136. The average frequency of the errors was 0.056 for the control and 0.066 for the fixation model. The error with the highest frequencies were ’coup’ at all positions for the control with a frequency of
and ’shed’ at position 1 for the fixation net with a frequency of 0.223 (see table
and 2.6. This result also means that the networks have correctly learned all the
higher frequency irregular words that were in the training corpus.
Tables of the errors with a frequency ≥ 0.1 have been included (2.5 for the control network and 2.6 for the fixation network). Note that a complete list of the errors made by each network can be found in appendix B and C respectively. As can be seen from the tables, most pronunciations computed by the networks, even though they are wrong, are nevertheless pretty close to the actual output. Thus, for example both networks pronounce the word ’dose’ as /d@Uz/ instead of /d@Us/ in most positions. Clearly this error is only a very slight error in the sharpness with which the coda of ’dose is pronounced. In fact, looking through the errors in the appendix, this error of the network substituting an /z/ instead of an /s/ is very common. Although this seems not to be a very big error, the pronunciation /d@Uz/ would correspond to the word