Creating linguistic resources for automated translation

A major difficulty in developing automated language translation is that you need a system with a fairly extensive vocabulary from which it can learn, before any degree of reliability or accuracy is possible. The LC-STAR project developed just such a vocabulary.

“First, we created large lexica for several language databases,” explains project coordinator Ute Ziegenhain of Siemens in Germany. “Secondly, we developed a demonstrator that could automatically translate speech to speech for output to another interface.”

Having finished on 31 January 2005, the IST programme funded-LC-STAR developed vocabularies, called lexica, and bodies of writings, called corpora, for some 13 languages in all, ranging from Italian and Greek to include Arabic, Chinese, Hebrew and Russian. These linguistic databases comprise a minimum of 100,000 entries per language.

The lexica and corpora are needed to train such systems for reliable, automated speech-to-speech translation (SST). Once developed, the various SST components (flexible speech recognition, high-quality text-to-speech synthesis and speech-centred translation) can be integrated into speech-driven interfaces embedded into mobile appliances and network servers.

The team also produced a working demonstrator called ‘Gaia’, which is a telephone server capable of translating between the project partners’ languages of English, Spanish and Catalan within a single register. LC-STAR focused on the tourism register, however Ziegenhain stresses that the system can be opened up to any domain if it is provided with sufficient vocabulary.

Results already in use

LC-STAR project results are already in use by Siemens within its own speech recognition and speech synthesis systems. They have also been supplied to the European Language Resources Association (ELRA) for further dissemination. ELRA makes available a variety of language resources for language engineering and the evaluation of language-engineering technologies.

In addition, LC-STAR vocabularies and machine-translation technology have been incorporated into the ongoing TC-STAR project. TC-STAR is a long-term effort (six years) focused on advanced research into core technologies for speech-to-speech translation – its goal is to make a breakthrough in reducing the gap between human and machine performance.