I'm a huge fan of MT, and do think that it can be used to get pretty accurate/high quality results. You can actually have multiple people work on each HIT and then cross-check or chose the best
An algorithm you could use: try three times, prefer the most similar two ("map"), look in those for puzzle-piece matches with the nearest neighbors ("reduce").
I actually posted about our experience with MY yesterday here: http://blog.newscred.com/?p=157