Chapter 2. The Simulations
Errors by Phoneme Length
Table 2.7 shows the errors by phoneme length for both networks. This table is included in the report because it is interesting to see for which target length the network has the most problems. Intuitively one would imagine that the most errors would occur with long phonemes since the there are more phonetic features to be learned and hence more possibilities to make a mistake. To be able to compare the networks’ performance at different phoneme length, the table contains a column labelled percentage, which gives the percentage of errors of a particular length. For example for the control network there are 1775 events of phoneme length 3. The number of errors is 41 which is 2.31% of the total number of events for length 3.
However, the results indicate that learning words which have a larger feature bun- dle associated with them is not necessarily harder than learning words with a shorter phoneme. Thus, for both networks the highest percentage of errors (2.31% for control and 2.99% for fixation) were made on words with three phonetic features, whereas the 4 and 5 letter words have comparably low error percentages.
Different Output Positions
Table 2.8 show the errors the network made at the different output slot positions. As described before, the phonological output of the network is divided into six slots in total with an onset-nucleus-coda structure. The onset consists of two consonants, the nucleus of two vowels and finally the coda of two consonants again. For a word that has only one phonetic feature in:
the onset position, the first slot is occupied (i.e. the onset is left aligned)
the nucleus position, the first slot is occupied (i.e. the nucleus is left aligned)
the coda position, the second slot is occupied (i.e. the coda is right aligned.