The iADAATPA~1 project coded as N° 2016-EU-IA-0132 that ended in February 2019 is made for building of customized, domain-specific engines for public administrations from EU Member States. The consortium of the project decided to use neural machine translation at the beginning of the project. This represented a challenge for all involved, and the positive aspect is that all public administrations engaged in the iADAATPA project were able to try, test and use state-of-the-art neural technology with a high level of satisfaction. One of the main challenges faced by all partners was data availability. Although all public administrations had some data available, it was clearly insufficient for high-level customization. In some cases, we had merely a few hundred words or several tens of thousand words. Each domain (field) has its own unique word distribution and neural machine translation systems are known to suffer a decrease in performance when data is out-of-domain.
展开▼
