That is a little technical.
Even if what we call translation outcomes, machine-translated texts, are corrected perfectly as an outcome—that is a day I dream of—we are still going to have to put them back through the system. As I said, this is an automated statistical system. The machine learns. It will not make the same mistake if it is properly corrected. So the goal is to invest in corpus development, and, at the same time, in post-editing. Investing in corpus development means feeding the software with high-quality texts that it can compare.
The system will search statistically for what is most often present in its memory. If everything there is bad, it is going to come out bad at the other end. For example, if there are 25 acceptable solutions and two that are not acceptable, the system statistically will search for the 25.