X hits on this document

PDF document

The split-fovea model of visual word - page 40 / 78





40 / 78

Chapter 2. The Simulations


does not seem to have an effect on the performance in the different output regions. The only noticeable difference is in the coda slot, where the fixation network makes 7 more errors than the control network. This can be explained by the fact that the fixation net receives proportionally more words in the input that are fixated at the third position. This means that the last letter of the word is the only letter in the right field (left hemisphere). This makes the task faced by the left hemisphere of getting the right pronunciation for the letter more difficult because it does not have information about the previous letters. For the network to compute the correct sound nevertheless, the connections between the two hidden layers are crucial. Because the net sees the last letter of a word proportionally more often alone in the right visual field, the exact pronunciation of this letter is harder to learn, which is reflected in table 2.8.

Table 2.8 also shows that both network make the most errors for phonetic features occurring in the onset. An obvious difference between the onset/coda slots and the nucleus slot is that the nucleus slot consits entirely of vowels whereas the other two consist entirely of consonants. Since there are only 8 different phonetic features for the vowels compared to 24 for the consonants (see figure 2.2). It should therefore be easier to for the network to learn the vowels since they appear at a much higher fre- quency. In this case the expected results should clearly be noticeably less errors in the nucleus position versus the other two positions. The reason for the high error rate in the nucleus slot lies in the unpredictability of the pronunciation of vowels in the English language (e.g. hint, mint but: pint). This makes it a very difficult task for the network to learn the correct pronunciation of the vowels in a word, especially since their pro- nunciation depends on the preceding and the following letters, Thus the hidden layers of the networks have to encode bigrams and trigrams to correctly compute the output for the vowels. The pronunciation of the consonants on the other hand is much more predictable. The difficulty for the consonants lies in the fact that there are so many different ones (24). The difference between errors in the onset slot (53 for control; 51 for fixation) and the coda slot (33 for control; 40 for fixation) arise the fact that the

Document info
Document views152
Page views152
Page last viewedFri Oct 28 21:50:15 UTC 2016