X hits on this document

PDF document

The split-fovea model of visual word - page 61 / 78





61 / 78

Chapter 3. Discussion and Conclusion


work was presented comparatively more words at position 4 than at any other position. This corresponds to the reading behaviour of humans for 4-letter words when they read words in a text. Note that the differences in the fixation position frequencies were very small. The difference between the highest frequency position (position 4 with a frequency of 0.221) and the lowest frequency position (position 1 with a frequency of 0.177) was only 0.044. Another interesting fact is that the difference between the pre- ferred position of the control network (position 2) and position 4 was only 0.024. Even though these differences are very small, they nevertheless managed to induce a change in the networks behaviour and counter the natural preference of the network for words fixated slightly left of centre. Thus even these very small differences in the initial con- ditions of the networks manage to change their behaviour. For the second simulation the differences between the frequencies of the fixation positions was enhanced. This simulation confirmed the previous findings. Whereas before, the fixation position right of centre was only slightly preferred, there now was a much bigger preference for a fixation right of centre reflecting the initial conditions of this network.

Another interesting aspect of this simulation was the fact that for the homographs, the control network pronounced the word ’read’ as /rEd/ whereas the simulation net pronounced the word as /riid/ (in all fixation positions). This hints at a difference in the two networks in how information is stored in the hidden layers. Both of these possibilities appeared roughly the same number of times during training. Nevertheless both networks pronounced the word differently. This might come from the fact that the fixation net sees the word more often as ’rea-d’ and thus would use what it has learned about ’ea’, namely that ’ea’ appears most often as /ii/ in the corpus. The control net on the other hand will have a preference to encode the ’ead’ in read as /Ed/ since that is most frequently the case for words ending in ’ead’ (e.g. ’dead’). Note that there would be no problem for either of the nets to learn either of the pronunciations if ’read’ was not a homograph. Since the network here has two possibilities, what it has learned before about other similar words will influence the final result on ’read’. This was the

Document info
Document views228
Page views237
Page last viewedFri Dec 09 08:11:33 UTC 2016