Tom,
Just feeding Hal raw data of having a script that would allow it to scan the web gleaning bits of info here and there really wouldn't work well.
It has been found that the best way for Hal to learn is through conversational interaction with a human and that doesn't mean a thousand of web hoppers chatting with your bot online either. GIGO was an old computer phrase that means Garbage In, Garbage Out. If one gives Hal good english (spelling, grammer and meaning) then the user can expect better results from Hal in return.
This is not to discount your ideas of Hal's learning. By all means, we'd love to hear of any positive results you (or anyone) gets.
Keep us posted on your proposal and progress and welcome aboard!