Chapter 3. Discussion and Conclusion
only homograph in the corpus for which this comparison can be made. For all the other homographs, either one of the possibilities had a much higher frequency (and was thus learned) or both had a very low frequency (e.g. both 0.05) and so wouldn’t have appeared often enough during training. A similar argument can be applied as to why the fixation network made an error on the word ’pint’ in the second simulation with less words.
To conclude it can be said that there is indeed a difference in how both networks evolve during training. The difference was small for the first simulation. However, the second simulation confirmed the differences in the networks by having bigger differences in the fixation positions. The effect is also expected to be bigger for larger words, since the fixation position frequencies for larger words are naturally larger.
The present work can be considered a feasibility study to see if it is worth pursuing this research of combining real fixation data with the split-fovea model of reading. The conclusion drawn before suggests that there are indeed possibilities for future research. An interesting test, which has not been done in the current simulation due to time constraints, would be to lesion the network. More precisely, by cutting the connections between the two hidden layers (i.e. removing the Corpus Callosum), it would be possible to further examine in how far the two networks have encoded the mappings between orthography and phonology differently.
Another direction in which future research could go is to build networks that admit longer words as inputs. Longer words have larger differences in the frequencies of the different fixation positions. Also, the preferred viewing postions for longer words during reading is left of centre as opposed to right of centre for 4-letter words. It would be interesting to see what would come out of a full-scale model including words