The night train trick that made translation less guessy
In the mail car of a night train, a clerk squints at a long strip of station notes. The stops are quick, so the clerk reads it once, then scribbles a tiny pocket card. A partner uses that pocket card to write fresh notes in a different language as the carriage rattles.
The timing is the killer. The first line the partner needs might depend on something near the end of the strip, so fixes arrive too late. Older setups tried to cope with big hand-made lists of phrases and rules, but they get heavy and snap when wording changes.
Then they try a cleaner split. One part reads the whole strip and squeezes it into one pocket card. The other part writes the new sentence one word at a time, guided by what it has already written and what the pocket card says. Takeaway: one learned summary can steer a full translation.
They also find the pocket card gets better when it is rewritten a few times before the partner sees it. Like passing the card down a line of clerks, each one making it clearer. With a deeper chain, the summary holds together, and the word choices wobble less.
Then comes the odd bit. The clerk flips the strip and reads from the end first, while the partner still writes in the normal order. It feels wrong, but details needed early are now read more recently, so it is easier to link the right cause to the right word.
Scaled up to lots of paired sentences, this two-part setup, plus the flipped reading, gives stronger translations than an older phrase-and-rule machine. It can even act as a second opinion, scoring a shortlist of options. Unknown words get stamped as UNKNOWN, so the rest still makes sense.
By the end of the run, the mail car feels different. Instead of lugging around more and more rule lists, the team trusts the pocket card the system learns for itself. And that little flip of the strip, reading backwards, turns out to be the quiet helper.