(been lurking for a while so thought I'd finally jump in with a first post! Edit: looks like I made a lame post in 2006, oops)
Trellis modulation is a sort-of misnomer. You still have to use it together with other modulation schemes (eg QAM, PSK etc)
It's a way of encoding the data so that if the transmitted data gets corrupted, the receiver can 'guess' what you actually meant. For example, imagine we're playing a strange version of chess where you're allowed to take three moves in one turn.
I decide that I'm going to move my knight from w to z like this:
I then send you the list of positions: B2, C4, E5, C6. However, during transmission, one of the positions is corrupted because of noise, and you actually receive B2, C4, G5, C6, which would look like this:
When you receive this, it's obviously wrong because you know that knights can't move like that. However, because you know the rules, you can pick the most likely move that I was trying to do, and assume that B2, C4, E5, C6 is what I actually sent you and therefore repair the data yourself.
That's conceptually how it works - the actual implementation is quite tricky! It's useful because it means you can turn the transmission rate up, and recover from some of the errors. Without error protection, you have to transmit more slowly so that each bit of information has a high enough chance of getting to the far end without being corrupted. If you're allowed to lose some of the data during transmission, you can transmit much faster and deal with the consequences later.