Report on the MT Roadmap workshop organised in conjunction with the 9th International Conference on Theoretical and Methodological Issues in machine Translation (TMI2002, March 13-17, Keihanna, Japan).

The workshop was the fourth in a series of ELSNET workshops aimed at the creation of a broadly supported roadmap for human language technologies.

For more information on earlier and future workshops visit http://www.elsnet.org/roadmap.html.
The URL of the workshop is http://www.elsnet.org/roadmap-tmi2002.html.

The workshop was organized by Steven Krauwer and Laurie Gerber, and it was attended by some 30 participants.

The aim of the workshop was to identify major challenges for MT. As for an individual researcher or developer the solution of a single detail problem in his dissertation or in a prototype he is developing may count as a major challenge, some indicative quantification of the size of a real challenge was given: Participants were asked to imagine that they were in charge of a major R&D programme, and that they had 100 million USD to spend on one single problem over a period of up to 5 years.

The challenges were grouped in three classes, based on the perspectives one could adopt:

• research challenges
• provider challenges
• user challenges

It turned out that one important header was missing: human resources, especially cross-training of linguists and computer scientists.

Some participants objected that under the present economical conditions even fantasizing about such amounts of R&D money was hard. Others said that massive funding for a single programme to solve a single problem would not make sense because there is not such a thing as a big problem, and that it would be much more effective to fund a large number of independent small size projects.

Under the heading research problems the following candidate challenges were collected:

• Developing a formal theory of translation
• Developing a semantic theory
• Eliminating the knowledge acqusition bottleneck
• Using translation memories (bitexts) and machine translation together in a product
• Creating permanent shared language repositories (sharing), including huge, word aligned multitexts
• Robust speech recognition (based on speech and other signals such as gestures, facial expressions, etc) to meaning.
• Moving towards a theory of crosslingual communication aids for situation dependent solutions

The following challenges from the user perspective were proposed:

• Language plug-ins for mobile phones (for transactions rather than full fledged interpretation)
• Help with the hard part of foreign languages.
• Large MT evaluation from user perspective.
• Standard control menu language (for cross-language communication by means of small menu driven devices)
• Crosslingual sign-reading eyeglasses (foreign language signs or messages are read by a small camera, and the translation is projected in the user's glasses)
• Learning from user feedback (via post-edition tools), and predicting user needs, constructing user models
• Web search and translation with CLIR.
• Automatic stenography (TV, conferences)

Industry challenges:

• Language plug in for cellphone, but as a service
• Ways to stick language books into MT system
• Using TM (bitexts) & MT together in a product
• Coverage of Minority languages.
• Masssively annotated multitext.
• Exploiting markup.

When trying to look back and identify achievements that in hindsight would count as overcoming challenges the following items were proposed:

• Free on-line web translatioin (however poor the quality)
• Preservation of markup
• Commercial speech-to-speech translation

There was a general agreement that it would be an interesting follow-up exercise to try to integrate the points listed above in the timeline contained in Bernsen's roadmap document based on the first roadmap workshop in November 2000.