Digression analysis in text ?

People seem to go away from the topic from time to time. Quite often they are adding information to or expanding on the topic, to help the listeners, to think. (or prevent them from clear thinking. ) That is, what they say doesn't always fit into the logical flow of the utterance. This may be a kind of important information for text miner. So what are the cues for digression? As a lecturer, I may say "well, now, let's stop here for a second", "That reminds me ...", "By the way", etc. And when to back to the main topic? "Now, let's return to ..", "Anyway, ..", "Getting back to the subject ...". In order to get a full comprehension, it'd be better not to ignore these, but, how can we let machine learn this?