X hits on this document

PDF document

The split-fovea model of visual word - page 27 / 78





27 / 78

Chapter 2. The Simulations


word correctly and produces the correct output. Yellow units stand for an activation of +1, whereas blue units stand for an activation of -1. Any other colors are somewhere in-between1.

Phonological Representation.

The phonological output layer consists of 66 units in

total. The layer is made up of six slots containing eleven units each. These eleven units correspond to the eleven different phonemic features which were used by Shillcock and Monaghan (2003) (see figure 2.2). As can be seen from the figure, each phoneme has a very specific set of features associated with it. In the output layer, these features are represented by the activation values of the different units, ranging between -1 and 1. Some more special English phonemes (for example ’x’ as in the Scottish ’loch’) are not represented in this table and words containing these were discarded. The phono- logical representation was of the onset-nucleus-coda (CCVVCC) form (e.g. Plaut et al. (1996)). The first two slots are reserved for the onset consonants, the middle two for the nucleus vowels and the last two for the coda consonants. For words that have less than 6 features (e.g. back b&k) the remaining empty slots have all their features set to -1. The network’s task therefore is to map the slot-based orthographic input into the

just described phonological representations at the output.


The training corpus for both networks consisted of English four letter words

taken from the CELEX corpus (Baayen et al., 1995). The words were taken from the

CELEX word form corpus rather than the word lemma corpus since the word forms

are what is actually read and the fixation position data used thus applies to the word

forms rather than the word lemmas. After eliminating the words that did not fit the

ONSET-NUCLEUS-CODA structure, there were 1988 words left. Since there are five

different positions in which the words can be presented to the net, the total number of

events presented to the network is 9940.

1Note there are no clean-up units in this architecture. This network does not need clean-up units to learn its task

Document info
Document views138
Page views138
Page last viewedThu Oct 27 07:12:25 UTC 2016