X hits on this document

PDF document

The split-fovea model of visual word - page 59 / 78





59 / 78

Chapter 3. Discussion and Conclusion


fully learned both regular and exception words. The learning of the exception words is an important test for any connectionist model. Since adherents of the Dual-Route theories claim that it is impossible for a single system to learn both exception and nonwords (e.g. Coltheart et al. (1993)), the fact that the model was able to learn the exception words is a first step to validate the model. The networks were also able to learn most homographs correctly, although the control network was slightly better at this. The difficulty in pronouncing these words comes from the fact that they have two different pronunciations associated with them. This makes them difficult to learn since both pronunciations will interfere with each other during training. Nevertheless the networks were able to learn one of the pronunciation for almost all of the homographs in the training corpus.

The second important test with respect to the claims of the Dual-Route theory is the test on nonwords. The networks were tested on two sets of nonwords, consistent and inconsistent. Both networks scored between 81% and 85% of the nonwords correctly for both sets, with the control net being slightly better than the fixation net. This score is good enough to conclude that both networks have successfully managed to pronounce nonwords. It has to be noted that most connectionist networks are trained on more word presentations. Thus for example the Shillcock and Monaghan (2003) model was trained on 10 million words, whereas the current models were only trained on 4 million words to reduce training times. It is expected that a longer training time would improve the nonword performance of the network since a lot more of the low frequency words especially would be presented more often than in the current simu- lations. Additionally this would also improve the overall performance of the network on the words in the training corpus, since the majority of errors were made on low frequency words to which the network was not often exposed.

To conclude, it can be said that both networks were able to learn both exception words and nonwords. A single system is responsible for learning both of these types of words.

Document info
Document views110
Page views110
Page last viewedTue Oct 25 17:38:06 UTC 2016