Programme CLARIN Annual Conference 2023

The task of translation involves language understanding and generation and, in this way, naturally combines the two essential challenges in computational linguistics and language technology. In the FoTran project, we are interested in the ability of neural translation models to pick up linguistic properties and to generalize to meaningful representations when trained on large amounts of multilingual data. Our focus is on the effect of linguistic diversity on abstraction and generalization. In order to study this, we need to create the necessary resources and infrastructure, and in this talk I will first introduce the OPUS ecosystem that fuels our research. In the second part I will concentrate on the experiments, studies and developments that this ecosystem enables within and outside of FoTran. I also welcome discussions on further directions that can be taken with the multilingual infrastructure we build, looking forward to your input.