It's a bit like T9 typing but in reverse. You do an isolation of which words can be constructed out of every key your finger slides over, assign higher weight to the ones where sharp turns or longer dwell times were, and then take the most statistically likely result, and suggest the next 3 possible results in the top bar.