Natural Language Parsing
- The agent has to accept commands from the user.
- The input process is for the user to type text.
- I've developed a fLIF parser which works on one rule and
only has a few words.
- I'm currently expanding it to work on several grammatical
constructs and several words leading to several hundred
interpretable sentences.
- Hold on, this gets a bit hairy. Remember it's all implemented
in fLIF neurons.
- There is stack which can be bound to specific words using
fast bind neurons.
- There is a stacktop.
- The initial word is presented.
- As the stacktop is 0 and there is input, a push is done.
- This causes the stacktop to be incremented and the 1st stack
item to be bound to this word. (The new word is then read in).
- A test net activates the second then first item to see if a rule
applies.
- In this case, no rule applies, but after the second push, the
rule VP -> VP NP-obj is ignited. (The sentence is Follow me.)
- It binds the NP-obj to the object slot of the VP (using fast bind
neurons).
- It then pops the stack and activates the clear network.
- Clear just wastes time, but does activate the first stack element
so that its binding and the VP-obj binding is supported.
- The result is that a meaningful semantic frame is produced.
- The clearing net is not really psycholinguistically plausible, and
I have a better mechanism without a stack in mind.
