Training a Neural Network to respond naturally

Hypothesis:

Using a combination of frequency analysis for word usage in every day speak (VIA trolling online web chats in things like IRC [internet relay chat] programs) and applying weights to aspects of the language such as:

1. Location in phrase
2. Location in conjunction to punctuation
3. Location and frequency to adjoining words (such as “do this now” this is adjoined by “do” and “now”)
4. Timing to previous statements (if “What’s good” is closely followed by “Everything, how’s you?” and given a weight, it would be a good assumption that the two are interrelated and one is a response to the other)
   

I believe using this process it would be possible to “teach” a neural network how to respond “naturally” in conversation. The process could be trained naturally for different dialects by simply limiting the exposure to strictly that dialect.

*EDIT: October 17th, 2010*

I would further hypothesis that the use of recursion in this neural network would lead to a more stable output, and a better neural network in the long run.