1968 SystranPeter Toma / Systran — Russian–English for the USAF, then the European Commission from 1976. Still trading.
1970 LogosLogos Corporation — English–Vietnamese for the US Air Force, later German and commercial European pairs.
1976 MétéoTAUM, Université de Montréal — Weather bulletins, English–French. A narrow domain in daily production for decades.
1981 ALPSAutomated Language Processing Systems — Interactive translator workstation rather than full automation. An early move to assistance.
1982 ATLASFujitsu — Japanese–English. Part of the Japanese commercial wave of the early 1980s.
1984 PIVOTNEC — Interlingua-based Japanese–English, one of the few committed interlingua products.
1985 METALSiemens / University of Texas — German–English transfer system, heavily engineered and expensive to license.
1991 PC-TranslatorLinguistic Products — Desktop MS-DOS translation for a few hundred dollars. The market moves to the personal computer.
1993 Power TranslatorGlobalink — Consumer desktop translation for Windows, sold in retail boxes.
1994 PROMTPROMT — Russian–English and later many pairs. Survived the transition to web and mobile.
1997 Babel FishAltaVista / Systran — Free web translation at scale. The first time most people used machine translation at all.
1998 TranslaticaPOLENG / PWN — Polish–English, one of the strongest of the smaller-language commercial systems.
1999 ParaLinkPROMT — Web and desktop translation aimed at Russian-speaking consumers.
2006 Google TranslateGoogle — Statistical phrase-based, later neural. Coverage over quality, which proved the right trade.
2017 DeepLDeepL / Linguee — Neural from the start, built on the Linguee parallel corpus.
Commercial machine translation systems and what happened to them
The survival pattern is not what a reader might expect. The systems that lasted were not the
linguistically sophisticated ones. Systran, architecturally the oldest and least elegant
design in the catalogue, outlived nearly every research-derived competitor because it had
institutional customers who needed gist translation of large volumes and did not care how it
worked.
The interlingua products were the opposite case. NEC's PIVOT and the European Eurotra project
took the theoretically cleanest route, and commercially it was the weakest: an interlingua
pays off only across many language pairs, and no vendor had enough of them to recover the cost.
Read across the catalogue, the commercial machine translation systems that endured were the
ones whose buyers wanted volume rather than accuracy.
The desktop era and the history it left behind
Between roughly 1991 and 1999 translation software became a boxed consumer product. This
period is the worst documented in the whole history of the field, because the artefacts were
shrink-wrapped floppies and the companies were small. Much of what is known about ParaLink,
Translatica, PC-Translator and their competitors comes from catalogue entries compiled at the
time rather than from anything the vendors preserved.
That is why a directory of translation software written in 1996 is now a primary source. The
compendium reproduced here was maintained while those products were still on sale.
Where the catalogue comes from
Entries are drawn from the vendors' own documentation of the period, from the trade press that
reviewed the desktop products, and from the institutional procurement records that survive for
the large installations. Where a claim could only be sourced to marketing material it is
described as a claim rather than repeated as fact, which is why several entries say what a
system was sold as doing rather than what it did.
The gaps are real and worth naming. Almost nothing survives for the Eastern European and
Soviet systems of the nineteen-seventies, and the Japanese commercial wave of the early
eighties is documented far better in Japanese than in English. A catalogue assembled from
English-language sources will understate both.
How to read a catalogue entry
Three things about these systems are worth holding in mind while reading the list. The first
is that a vendor's language count is not a capability count: a system advertising forty pairs
usually had two or three it handled well and the rest by pivoting through English, which
compounds every error twice.
The second is that price signalled the buyer, not the quality. The institutional systems cost
as much as a building because they were sold to organisations translating millions of words a
year, and their output was never meant to be read as finished work. The desktop packages of
the nineteen-nineties cost a few hundred dollars and were sold to people who expected exactly
that, which is where most of the era's disappointment came from.
The third is that survival in this catalogue is a poor guide to merit. The systems still
trading are the ones whose customers needed volume and tolerated error, not the ones whose
linguistics were best. Several of the most sophisticated designs here shipped once and left no
documentation behind at all.
What was the first commercial translation software?
Systran, founded in 1968 and running for the United States Air Force from 1970, is normally given the title. Earlier systems existed as research installations rather than products.
Why did Babel Fish matter so much?
It was free, it was on the web, and it handled arbitrary text. Before 1997 machine translation was something institutions bought; afterwards it was something anyone could try, which reset public expectations of what these systems were for.
Are the older systems still available?
Almost none. Systran and PROMT continue in modern form; the desktop packages of the nineteen-nineties are effectively lost software, which is why the catalogue entries describing them have value.
Why are the institutional systems so poorly known?
They were sold to a handful of buyers under licence, never reviewed in the consumer press, and documented only in manuals that stayed with the customer. A system running daily inside a government translation service for fifteen years can leave less public trace than a boxed product sold for one season.
