Distributed Language Translation

~ the obit ~

DLT closed in 1990. I first dialed a modem in 1998, so this one sits eight years before anything I can claim to have witnessed, in a Dutch software house I have never had reason to visit, on a problem I only met much later as a user of the free translation boxes that came after it. Everything below is reconstruction from the project's own published record, which is unusually complete because the people running it wrote everything down.

Distributed Language Translation was an attempt at interlingual machine translation for twelve European languages. The idea behind an interlingua is arithmetic: if you translate directly between every pair of twelve languages you need one hundred and thirty two translation engines, and if you translate everything through one language in the middle you need twenty four. The hard part is choosing the middle language, because any natural language you pick carries its own ambiguities into the pivot and makes them permanent.

DLT's answer was Esperanto, or rather a modified version of it. Toon Witkam, an aeronautical engineer by training who had joined BSO in Utrecht in 1978, got the idea in his car in the spring of 1979 while the BBC World Service talked about Balzac and while he was vaguely aware of how much translation the European Community was drowning in. Esperanto is perfectly regular. Every noun ends in -o, every adjective in -a, and the whole grammar has no exceptions to encode. As a machine-facing interface format that regularity is worth a great deal.

~ the rap sheet ~

~ how it worked ~

The name is the important part and it is usually misread. Distributed did not mean the system ran on a cluster. It meant the translation was split in half and the halves ran in different places: your workstation turned your sentence into the intermediate language, and that intermediate text was sent over a network to any number of other workstations, each of which turned it into whatever language its user read. Because ambiguity was resolved at the source, before transmission, the pivot text did not need a human to disambiguate it again downstream.

There was a second reason for choosing Esperanto that has nothing to do with linguistics. In the late seventies everyone in networking believed link bandwidth would stay the bottleneck forever while processors got faster, so compression was worth almost any amount of computation. Witkam put an informatics intern, Gert-Jan Vlasveld, on a six-month study of how compressible Esperanto was. A Huffman coding scheme over Esperanto morphemes came out at an average of nineteen bits per word, against the five or six bytes the same word would take in ASCII. The interlingua was also the wire format.

~ what it actually achieved ~

Preliminary research ran from 1979. The funded project started in the mid-eighties, with the Dutch Ministry putting in eight million guilders across six years, roughly three million ECU, and BSO matching it from its own money. Fokker, the Dutch aircraft manufacturer, came in as the industrial partner and the Universal Esperanto Association came in on the language side.

The best result was narrow and real. Working on AECMA Simplified English, the controlled vocabulary that aircraft maintenance manuals are written in, DLT hit around ninety five percent accuracy, checking both narrow and broad context rather than just swapping terms. On general text, reports from UNESCO meetings and the like, it landed between fifty and sixty percent. Nobody was going to fund a second phase on fifty to sixty percent, and BSO could not raise the money. The project stopped in 1990, unfinished.

~ the next chapter ~

Fokker went bankrupt in 1996 and BSO was eventually absorbed into what is now Atos. Witkam wrote the project's history up in 2006, and the research output, several books and a long run of papers, was left deliberately public and detailed enough for someone else to build on. That was the stated consolation at the end: the code died, the description survived.

Interlingua approaches did not win. Statistical translation arrived in the nineties and neural translation after it, and both of them skipped the pivot language entirely by learning mappings from enormous parallel corpora that did not exist in 1985. What is quietly funny is that the pivot came back anyway. Large multilingual systems still route rare language pairs through a high-resource language, usually English, and they inherit exactly the ambiguity problem that made Witkam reach for a constructed language in the first place. He wanted the middle to be a language with no exceptions in it. What the industry settled for was English.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »