If I see correctly, the evaluation script used is not the one of the CoNLL 2018 Shared Task (https://universaldependencies.org/conll18/evaluation.html): since this one checks for cycles and multiple roots, I am wondering if your reported UAS and LAS scores take this into consideration (or, as I suspect, your predicted sentences could have cycles/more roots).
If I see correctly, the evaluation script used is not the one of the CoNLL 2018 Shared Task (https://universaldependencies.org/conll18/evaluation.html): since this one checks for cycles and multiple roots, I am wondering if your reported UAS and LAS scores take this into consideration (or, as I suspect, your predicted sentences could have cycles/more roots).