An introductory study into conversation theory will tell you how theoretically bad up/down is. It seems to me that the two-step process is still broken though, because "important" is not a well-defined word. The best system I have seen online, which actually does a decent job of utterance categorization, is Slashdot. Based on the Gricean Maxims and you can come up with some core cirteria for any given utterance, and /. actually covers them surprisingly well. Obviously it's still not a perfect implementation, but they have done a good balance between categorization, ease of categorization, and evaluation of categorization. They really need to grow out of the +5 limit though.
While /.'s criterion are good, they don't cover enough ground to become a universal system; it doesn't scale to other cultures. There's a lot more to this I think, but I'm just be rambling again.
I think the quality not scaling is the main issue. Adding another step to the process is just going to make people less interested in contributing their opinion at all.
In allowing people to vote up/down, I realized that there needed to be a separation between 'Yes this quote is accurate' and 'Yes I think this quote is important', since the two can be exclusive.
In the end I decided to implement a two-step process. The first step is to Agree/Disagree with the quote at hand. Once a user has voted in this way, they have an optional second step to vote for how 'Important' the quote is (important or unimportant).
I have yet to see how it will work out, but I'd love any comments you guys could provide.