Currently the primary UD dependencies are used but what we actually need is the secondary dependencies since they better reflect the semantics. There are two problems:
- the training should be made to work with DAGs instead of trees
- not all treebanks have secondary dependencies but it looks like it is possible to recreate them automatically.