Words
- Now I mentioned that CAs act as categorisers of inputs.
- So, I could (theoretically) build a CA that recognised the
typed word "dog" or one that recognised the spoken word "dog"
or indeed one that recognised real animals.
- That CA could then persist after the input had passed.
- What I really did on the parser was had symbolic input
that ignited a particular CA when the parser was on the word
"dog".
- The words were orthogonal, and in some versions of the parser
the underlying semantics was orthogonal.
- However, you can get some semantic overlap by having similar
words sharing neurons.
- So, I grabbed words from WordNet and the associated synsets.
- I encoded a word as 10 neurons for it, and 10 neurons for each
element (up to around 5?) of the synset hierarchy.
- So, girl and boy only differed on the root, while girl and cat differed
on 4 of 6 features.
