This is a extension of the graph-based dependency parser. Instead of training POS tagger and dependency parser separately, it trains them at the same time to avoid the error propagation.
With the parser above, you have to run a POS tagger before running the parser because POS tags are a important feature that is used in the parser. Therefore, the error of the POS tagger is propagated to the parser. To avoid this propagation, a tagger and a parser is trained using the shared features.