X hits on this document

PDF document

The split-fovea model of visual word - page 34 / 78

215 views

0 shares

0 downloads

0 comments

34 / 78

Fixation Position

Computed Output

Correct Output

5

/dVts/

/dOts/

2

/s@Ud/

/t@Ud/

dots

0.074

toad

0.062

Chapter 2. The Simulations

a. Regular

Word Frequency

29

Fixation Position

Computed Output

Correct Output

5

/Jiik/

/Siik/

1

/Tiik/

/Siik/

4

/h&Vnd/

/h&Ind/

3

/mjUIl/

/mjUUl/

b. Irregular

Word Frequency

Table 2.3: Errors on the regular and irregular words for the fixation net.

chic

0.080

chic

0,080

hind

0.113

mule

0.083

on a list of regular, ambiguous and exception words taken from Plaut et al. (1996). Note that for this test only words that appeared in the training corpus were taken. The lists of the words can be found in appendix A. The list of regular words contains 92 elements (corresponding to 460 events) , the list of irregular words contains 61 elements (305 events) and the list of ambiguous words contains 19 elements (95 events). Table 2.2 shows the results of these tests. The control network only made one error on the irregular word list, getting all the regular and ambiguous words correct. The error was made on the 0.08 frequency word chic, pronouncing it as /Jiik/ instead of /Siik/ at fixation position 5. The fixation net made 2 errors on the regular words, 4 on the irregular words and none on the ambiguous words. The details of the errors can be found in table 2.3. All but one of the errors are made on low frequency words with a frequency less than 0.1. These tests show that both networks have learned to pronounce both regular and exception words.

Document info
Document views215
Page views224
Page last viewedThu Dec 08 02:12:37 UTC 2016
Pages78
Paragraphs3087
Words17414

Comments