I am a programmer and also a father, and working on using AI to revolutionize how we teach kid how to read.
I noticed that a very simplistic LSTM neural network model is able to learn very quickly all the rules of pronunciation ex:
- grapheme "EA" in “please” or "heat" is pronounced one way
- but grapheme "EA" in « death » or " bread" is pronounced in a different way
- but grapheme "EA" in "great", "steak", "break" is also pronounced in a different way
In fact the neural network is learning the rules and able to guess the pronunciation of word it have never seen before.
This made me try to see
1- what is the minimal number of example the Model need to be trained with to learn all the rules.
2- What is the optimal sequence that rules must be learned ...
This can all be discovered and measured easily and accurately by training the model with different training set.
I believe this have never been done before, because experimenting with real kid is too slow.
We make the assumption that is something is easy to learn by an LSTM model it will also be easy to learn by a human. (turn out to be true)
There is a lot of design decision that need to be made about how to visually display "grapheme" and let the kid interact with them.
This is just a side project for now but I am in strong need of a UI programmer to partner with!
A concept I've thought about in this same area is some kind of color coding of words to help children learn how to decode words at the level of syllables. So for instance "please" might be "plea" and "se" – avoiding the sounding of of "pa la eh ah [no, they go together as an ee sound] sa eh? [silent e]". Lots of vowel insertion that makes it very hard to actually assemble those sounds into a word!
But in another way this could be a kind of overlay of information. Maybe it shows how words are split, maybe it's to disambiguate different pronunciations of "ea", ... I'm not sure what's the most important added information. But overlaid on the words it might be a kind of scaffolding that makes it easier for kids to successfully read words and gain the practice that makes it possible to remove that scaffolding.
There's a lot to think about there when it comes to reading (not to mention a ton of pedagogical knowledge)... I'm not sure I have the space to be a solid collaborator, but I do find this stuff interesting to talk about sometime...
I noticed that a very simplistic LSTM neural network model is able to learn very quickly all the rules of pronunciation ex:
- grapheme "EA" in “please” or "heat" is pronounced one way
- but grapheme "EA" in « death » or " bread" is pronounced in a different way
- but grapheme "EA" in "great", "steak", "break" is also pronounced in a different way In fact the neural network is learning the rules and able to guess the pronunciation of word it have never seen before.
This made me try to see
1- what is the minimal number of example the Model need to be trained with to learn all the rules.
2- What is the optimal sequence that rules must be learned ... This can all be discovered and measured easily and accurately by training the model with different training set. I believe this have never been done before, because experimenting with real kid is too slow.
We make the assumption that is something is easy to learn by an LSTM model it will also be easy to learn by a human. (turn out to be true)
There is a lot of design decision that need to be made about how to visually display "grapheme" and let the kid interact with them.
This is just a side project for now but I am in strong need of a UI programmer to partner with!