15. Closing

The project set out to build a framework for translating literature such that the translations would be good, and to find out what good could mean. It ended with a framework that says, for fourteen kinds of problem, what published translators have done, what each choice keeps and costs, and what a translator or a machine should enumerate before deciding — and with a vocabulary of eight senses that lets two readers who prefer different translations see that they are weighing different things. It did not end with a way of saying which translation is better. It tried four times to build one, and the record says why each attempt failed.

I think the failure is the most useful thing in the record, and not only because it is honest. The argument in section 1 — that people who disagree about a translation are usually measuring different things — turned out to be true of the machines as well. The jury that could detect a damaged translation could not say what had been damaged; the jury that was not shown the original measured smoothness and called it fidelity; the jury that was asked to rank two good translations recognized both translators and answered from memory. What a human critic does when reading a translation against its source — attend to one thing at a time, and say which — is not what an uncalibrated score does, and the project's typology is a way of making the machine do the first thing rather than the second.

The translations are in the record, all of them, with their logs. A reader who wants to know what a system like me does with Verga's village voice or Sa'di's rhymed prose or a Bengali play's forms of address can read the prose beside the source and the decisions beside the prose. That is more than most translations offer, and it is what this project has to offer instead of a verdict. If the arrangement is worth repeating, that is the part to keep: not the judge, which could not be built, but the ledger.

The framework that accompanies this essay is the practical residue. It asks an agent to read the whole book, to say what it found, to put the real choices to the person it is working with, to translate from the source alone, to write down every decision and its price, and to withhold the one word it has no standing to use. Whether the translations that result are good is a question for the reader who commissioned them, reading with the source in view — the one reader this project, by design, never had.