An accurate framework for carrying out multi-lingual discourse segmentation with BERT is described, trained to identify segments by casting the problem as a token classification problem and jointly learning syntactic features like part-of-speech tags and dependency relations.