The approach I have adopted in most of my work with microblogging texts (or microposts as I usually call them) is a combo of corpus linguistics and discourse analysis. Essentially this involves playing around quantitatively with language patterns and then focusing-in on anything that looks particularly interesting. I have largely been working with a 100 million word corpus of tweets, HERMES (Olympian God and herald), my technical slave scraped for me last year using Twitter's streaming API.
The most common three word pattern (3-gram) in the corpus is 'THANKS FOR THE'. Generally people are thanking each other for a FF (follow friday) mention or a RT (retweet). For example:
@User Hi – thanks for the RT. Much appreciated :)