Multi-source Cross-lingual Delexicalized Parser Transfer: Prague or Stanford?

We compare two annotation styles, Prague dependencies and Universal Stanford Dependencies, in their adequacy for parsing. We specifically focus on comparing the adposition attachment style, used in these two formalisms, applied in multisource cross-lingual delexicalized dependency parser transfer performed by parse tree combination. We show that in our setting, converting the adposition annotation to Stanford style in the Prague style training treebanks leads to promising results. We find that best results can be obtained by parsing the target sentences with parsers trained on treebanks using both of the adposition annotation styles in parallel, and combining all the resulting parse trees together after having converted them to the Stanford adposition style (+0.39% UAS over Prague style baseline). The score improvements are considerably more significant when using a smaller set of diverse source treebanks (up to +2.24% UAS over the baseline).