X hits on this document

PDF document

The split-fovea model of visual word - page 51 / 78

237 views

0 shares

0 downloads

0 comments

51 / 78

Chapter 2. The Simulations

46

Fixation Position

Example

Real Frequency

New Frequency

1

-WORD

0.177

2

W-ORD

0.196

3

WO-RD

0.206

4

WOR-D

0.221

5

WORD-

0.200

Table 2.16: Changed fixation frequencies. The 3rd column

quencies as used in the previous nets and the 4th column

0.177

0.367

0.467

0.617

0.407

shows the real fixation fre-

shows the new enhanced

frequencies.

Table 2.15 shows the errors made by the networks on nonwords at different fixation positions, with the last column giving the difference between the two networks. As the upper table shows, there is no real difference between the two networks as far as the consistent nonwords are concerned. More interesting for the current paper is the bottom table. There is clearly a tendency for the fixation network to get about 3 or 4 more words wrong at the first three fixation positions. However on position 4 the network makes 6 errors less than the control. Compared to the positive difference at the previous three positions, this can be considered a significant fact. Position 4 is exactly the position where the fixation network was the most exposed to due to the highest fixation position frequency, confirming the results from the words in the training corpus. Even nonwords are easier to pronounce slightly right of centre since the networks was exposed to more words at that position.

2.3

Two additional networks.

From the previous analysis of the two networks, it is clear that the difference between the two networks is very small. To try and get a more significant difference, two more networks were trained for which the differences in the fixation position frequencies

Document info
Document views237
Page views246
Page last viewedSat Dec 10 05:37:47 UTC 2016
Pages78
Paragraphs3087
Words17414

Comments