X hits on this document

PDF document

The split-fovea model of visual word - page 15 / 78





15 / 78

Chapter 1. Introduction and Background


of hidden units with connections from and to the phonological units. Networks using this architecture are called attractor networks. The advantage is that the clean-up units help the phonological output layer to settle into the correct pattern. Attractor networks have an important property. They can form so-called attractor basins. This means that if the network is in state that it has not encountered before but that is close to a known legal state, then the current state will be attracted to the legal state and the network will eventually settle in that legal state. Figure 1.5 is an example of this. Note that the figure is simplified, using only 2 phonetic features. An attractor basin in this net has in reality 11 dimensions since it has 11 phonetic features. This enables the network for example to repair noisy representations since a particular noisy output will always be attracted to a valid state.

Another key difference between the two models it that Harm and Seidenberg (1999) used a phonological representation that is much closer to actual phonetic features. A single phoneme was represented by 11 units, each of which corresponded to a specific phonetic feature (e.g. voice, nasal, round...) and whose activation could vary between

  • -

    1 and 1 (see figure 2.2 in chapter 2).

This model was trained on 7839 monosyllabic words. Performance on words included in the training set was high, scoring 99% of words correctly. For nonwords, the net- work’s performance was significantly better than the older network’s, scoring 84% of words correctly. In this respect, the model achieved to eliminate one of the major crit- icisms of the older model, namely the low scores on nonword reading.

It has to be noted that the non-implemented parts of the complete triangle model as shown in figure 1.3 nontheless constitute an important part of the reading process. For example Plaut et al. (1996) suggest that there are two different ways in figure 1.3 to pronounce words. The first one is the direct route from orthography to phonology used by the Seidenberg and McClelland (1989) model and the second one is the indirect

Document info
Document views102
Page views102
Page last viewedTue Oct 25 01:59:51 UTC 2016