I thought this would be a 'behind the scenes' look at Twitter, the company. However, it's about using Twitter as a support channel at 37signals.

Interesting that they encourage Twitter as a support channel, as a user I like Twitter for support, but as a company I absolutely loathe it. It's hard to get details and follow up, and if anything requires an explanation or sensitive details you need to get an email anyway...

There are cases when people need to follow up with an email, but there are many more people who just have quick questions that can be answered publicly with a few words or a link to what they need.

Looking closely at the main screenshot, I only see one tweet related to product support. Look forward to future posts discussing their filtering methods. Raw word list? Bayesian learning? 440+ tweets is not insignificant, but experience tells me this takes about 3 or 4 hours per day just to manually identify/categorize using a sane level of effort.

The next two parts will have more detail, but the short answer is "yes" to both a raw word list and a Bayesian filter in terms of techniques we've tried here. One of the simplifying things that makes the problem a little easier is that we don't try to classify beyond "does this need an immediate reply or not."

The description of the requirements sounds like a very close match with what ExactTarget's "SocialEngage" service offers. If you checked it out and dismissed it, I'd be interested to hear why.