Chapter 2. The Simulations
control network whereas for example positions 2 and 3 have higher mse’s and higher error rates. Lookin at the change in the errors as well as in the mse, the degree of change between the networks mostly matches the relative fixation frequency at these positions. The only position which is clearly contrary to what would be expected from the fixation data is position 1. This position has been the least frequent to be presented to the network. Nevertheless, the fixation network makes less errors and the mse change is less than would be expected for this position.
The training corpus contained 14 homographs (70 events since there are 5 different fix- ation positions for each word). Homographs are words with the same orthography but a different phonology (e.g. read in ’to read’ → /riid/ or ’to have read’ → /rEd/). These have not been included in the discussion so far. The reason for treating the homographs separately is the following. Since the network has no way of knowing in which con- text the homographs appear, it is impossible for it to learn the correct pronunciation for both the phonemes of the word. For this reason these words were excluded from the previous analysis because they would inevitably introduce some errors because the network simply could not learn both pronunciations. However, it is interesting to look at the homographs on their own. More precisely, the question is whether the network is able at all to learn at least one of the possible pronunciations correctly since the two different possibilities will compete and interfere with each other during training. For words where one of the pronunciation has a much higher frequency than the other one this shouldn’t be a problem and the network is expected to learn the pronunciation with the higher frequency since that is the one that will dominate during training with little interference from the low frequency one. On the other hand, when both frequencies are similar, both possibilities will be presented roughly the same number of times dur- ing training and it will be hard for the model to settle for one of the two at the end of