Words
- If all you want is some simple language parsing (or production)
that's all you really need.
- However, real human language is actually processed (by humans)
via neurons.
- How can you represent words in neurons.
- The simplist way is to have orthogonal CAs. (Each neuron
is in at most one CA).
- As text engineering parsing is usually done by lexemes, and
if you make each work of exactly one lexical cat, you
can radically increase your efficiency.
- However, you can also take advantage of overlapping encoding.
- A lot of people think this really improves space performance,
but it remains the same order.
- The real key is that you can use overlapping encoding to
share semantics.
- We take word net syn set hierarchies, and encode a word by its
own neurons, plus those of the syn sets.
- You can then get semantic generalisation.
- Kailash Nadh and I have used this to generate a system that resolves
PP attachment ambiguity better than any other system.
