In mid January, the European Commission (represented by Language Technology maestro Roberto Cencioni) brought together 250 of Europe's brightest and best translation players - academic researchers, technology providers, and LSPs - in Luxembourg to offer them some badly needed innovation funding. There was not a lot in the pot - €40 million to collaborate on new ways of making better systems - but for the first time in several years, the machine translation agenda was specifically targeted for development. Why now?
For Hans Uszkoreit, the doyen of MT researchers in Europe, one answer lies in the Gartner Hype Cycle that tracks technology visibility along a complex curve. Professor Uszkoreit is the current coordinator of the EuroMatrix project that is attempting to weave extensive collaboration around building statistical MT engines for all "official" EU languages (506 pairs), and gave a stimulating keynote at the Luxembourg meeting.
The Gartner Hype Cycle for MT
The MT Hype cycle goes something like this:
- Early hopes back in the dawn of MT - translating Russian is really decoding the English version hidden inside the source - led to overinflated hype in the early Cold War
- then plummeted to near extinction (at least in the US) when the ALPAC report put the doom on "high quality fully automatic translation".
- Due to the "paradigm shift" introduced by more statistical and machine learning methods (and pressing intelligence needs in the U.S.), MT is now climbing up the Slope of Enlightenment, and there are whispers at camp-fires on the way up of billion dollar MT markets.
In Europe, the pathway has been slightly different: R&D programs were maintained in France and some other countries during the days of the U.S. Trough, and the attempt to produce a super MT system for Europe (Eurotra) went through its own mini-cycle of boom followed by a bust in the early 1980s when it failed to deliver.
Perhaps in fact the Hype Cycle is fractal, each slope containing smaller mini peaks and troughs within the overall pattern. It may prove hard to distinguish clearly between those old "inflated expectations" and the current widespread "hype" that there will be genuine productivity when Enlightenment it attained and new projects come to fruition within the (proverbial) next five years. Or to put it more bluntly, will MT ever grow into a visible business over and beyond its role as a technology add-on?
Today, Google provides free MT as part of its "no evil" mission to manage the world's information. MT buttons are becoming a feature tacked onto websites and blogs. And communities are providing linguistic labor in exchange for ego food. Apart from a small number of large organizations and one or two large companies, is anyone really prepared to pay for the kind of investments that a serious MT industry would require? The only real success stories of monetized MT services (e.g. the French site Reverso.net) do not tend to come from "industry leaders."
We are therefore embarking on slopes that may be going in contrasting directions. It is indeed time for a paradigm shift in translation delivery, processing more content in more languages for more end users using more technology. At the same time, where can "out of the wall" automatic translation generate money for its practitioners?
It was good to see that the EC event attracted a broad range of LSPs interested in collaborating on MT experiments. But the capacity to make MT pay will also require more Enlightenment in terms of inventing sustainable business models and value-add services that bring home the bacon.