Here's basically how this thing works:
It takes your input (block or URL) and splits it into word-pairs... such as word pairs (it, takes), (takes, your), (your input), etc etc that could be formed from this very text.
It stores these pairs and then spits out pretty random content based on picking at random from words preceded by a particular word.
It likes a lot of text, but not too much - say between 3000 and 8000 words is optimal - some of the canned texts had to be dumped in in parts, and i'm still working on that kind of interface for you.
Apparently, it does better with text based on a poor, redundant vocabulary... so post some CNN articles in this shit. Esp the ones about war or terror.
The "seed" words are used to set up the titles and do, in their own way, influence the rest of the text. I find that influence to be too weak, but at present have no clear way of fixing it...
The more coherent the source text, the better the output, given a source text of sufficient length. In other words, illiterate babble like AIM and IRC logs will make crappy texts, and I will periodically prune them.
1) A better way to determine sentence length. It seems that a lot of times it ends a sentence just a word early, and I'd like to avoid that if possible. Anybody know about this?
2) Multi-part add, subject update. All this would require some sorta user login funk.
3) I have no reason to believe this would not work in other languages with a similar effect. Anyone having some blog text in german, spenish, french, etc, is strongly encouraged to add it and see what happens...
If you have suggestions, go on the main site and poke around til you find a contact form.
Enjoy!