Regarding some articles (around 10-15% from my tests) that you find to be not of the best quality.
This is actually a problem of the initial scraping sources like Goarticles, I find the largest amount of rubbish article to be on that source. It's kinda hard to make a masterpiece out of article with shaky language.
With that said, here are 2 solutions that are going to be implemented in order to minimize this issue:
1. When more article sources, like for instance Squidoo, will be added to script - Goarticles will probably be discontinued.
2. I will add filter modules to LSI Engine that will either filter out or transform the most obvious shaky language into a proper content.