btw, i have figured out a self-training algorithm that my iskra lattice translation engine will have implemented in it, that should close a lot of the remaining 20% round trip accuracy. it uses special markers that basically say "we found a round trip that passed this node from language x to y - 2, 3, 4, 5 whatever points at distance a, b, c, d, e and then next time a round trip fails (so, translator will always try to round trip everything) from a different input, it will see this new meta-node and brute force around nearby to find the round trip candidate and then possibly it finds it with the same one, *or* it adds another new interpolation node that fixes the traversal through that same area for a different path.
there, quid pro quo, nerding for a nerding.
