1. Korea Maritime and Ocean University NLP Jung Tae LEE inverse90@nate.com

3. ` 1. Window Size Two reason of choose the window with seven letter  First, significant amount of the information needed to correctly pronounce a letters is contributed by the nearby letters.  Secondly, limited by computational resources to exploring small networks => the limited size of the window also meant that some important nonlocal information about pronunciation and stress could not be properly taken into account by our model.

4. ` Mutual information provided by neighboring letters and the correct pronunciation of the center letter as a function of distance from the center letter.

5. ` 2. Changes in the network Changed network performence  Dictionary : common English word layer repeat 11input groups & 80hidden units 7input groups & 80hidden units 25 passes7% higher > 55 passes97.5%95% The number of input groups was varied from seven to eleven.

6. Changed network performence  Dictionary : common English word Adding an extra layer of hidden units also improved the performace. layer repeat Two layers of 80hidden units 7input groups & 120hidden units 55 passes97% 1passes87%85% Network with two layers of hidden units was better at generalization but about the same in absolute performance.

7. ` 3. Analysis of the Hidden Units Graphical representation of activation of the hidden units  Levels of activation in the layer of hidden units for a variety of words Phoneme, /E/ was produced by output. The input string is shown at the left with center letter emphasized. The area of the white square is proportional to the activity level. Chief_ speak_ negro nity_ least believe equa arty_ see_ appy_ each nily_ only_

8. ` Hierarchical clustering of hidden units for letter to sound correspondences.

9. `  A hierarchical clustering technique was used to arrange the letter-to- sound vectors in groups based on a Euclidean metric in the 80- demensional space of hidden units. Hierarchical clustering of hidden units for letter to sound correspondences.  Shown figure, was striking : - the most important distinction was the complete separation of consonants and vowels. For the vowels : - the next most important variable was the letter. For the consonants : - clustered according to a mixed strategy that was based more on the similarity of their sounds.

10. `  The same clustering procedure was repeated for three networks starting from different random starting states. Hierarchical clustering of hidden units for letter to sound correspondences. - The patterns of weights were completely different. - But, the clustering analysis revealed the same hierarchies. With some differences in the details, for all three networks.

11. ` 4. Conclusions NETtalk is and illustration in miniature of many aspects of learning. 1. Network start out without ”innate” knowledge in the form of input and output => network could have been traind on any language with the same set of letters and phonemes. 2. Network acquired its competence through practice, went through several distinct stages, and reached a significant level of performance 3. Network is distribute the information without single unit or link 4. The network was fault tolerant and degraded gracefully with increasing damage. => but, network recovered from damage much more quickly than it took to learn initially

12. Conclusions  NETtalk is too simple to serve as a good model for the acquisition of reading skills in humans - ex) when children learn to talk, after reprsentation for word and their meaning, they learn to read.  This approach would have to be generalized to account for prosodic features in continuous text.  Human level of performance would require the integration of information form several words at once

13. Korea Maritime and Ocean University NLP Jung Tae LEE inverse90@nate.com