Walter von Hahn * Machine Translation

9. Cooperation

Multilingual activities should always be based on cooperation. Especially, there should be cooperation with native speakers from other research institutes and foreign countries.

Moreover, the history of past or ongoing projects in MT shows that data sharing is a very efficient method of cooperation and fruitful competition. A good example is the data collection activity in speech recognition. There exist corpora on the following domains:

- ATIS,
- Wall Street Journal,
- Conference Booking,
- Appointment Scheduling.

In machine interpreting meanwhile the scenario "appointment scheduling" is used in research groups in Germany (Verbmobil), Korea, Spain, USA, Italy, and Japan.

Two of the most important activities in the long run therefore are:

• to join ELRA (European Linguistic Resources Association, mail to info@relator.research.ec.org) or/and LCD (Linguistic Data Consortium, USA, mail to ldc@pine.ling.upenn.edu),
• to join ELSNET (European Network in Language and Speech, mail to elsnet@let.ruu.nl).

Another method to cooperate is to rely on common external evaluators from other groups.

10. Organisational requirements

What we learned from successful projects can be summarized as follows:

• Define clear technical aims to the funding institution and internally, aims which you can prove afterwards,
• Define clear deliverables and deadlines
  • when applying, and
  • when employing researchers,
• Agree on intermediate presentations/publications with the project crew,
• Define a test set (and training set, if stochastic processes are involved) to assess your progress, and install regular benchmarks every six months,
• Try to achieve an early implementation and keep it operational all over the run time of the project,
• Start or keep open and friendly competition with other international groups,
• Keep in touch with other groups in your field and send them reports and concrete results regularly.

126