Despite achieving parsing scores in the range of 93% (UAS) and 
92% (LAS) on German newspaper text, neural dependency parsers still 
struggle, for example, with predicting the core arguments for verbal 
predicates and often confuse direct and indirect objects, or attach 
prepositions to the wrong heads. Their performance drops even more 
dramatically when the target domain is different from the one the parser 
was trained on. In this talk, we investigate different techniques to 
make recent parsers more robust, namely (1) incorporating 
subcatergorisation frame (SCF) information into dependency parsing and 
(2) improving out-of-domain parsing by re-ranking.

SCFs provide syntactic information on the type of argument that a verb 
can take. In my colloquium talk last year, I showed that this 
information has the potential to significantly improve core argument 
prediction by extending a state-of-the-art parser for German (Dozat & 
Manning, 2017) with gold SCF information. This year, in the first part 
of the talk, I report on follow-up results for predicting SCFs and 
incorporating them into the parser.

In the second part of the talk, I present work in progress toward 
improving parsing German Twitter data using re-ranking. We propose to 
augment a neural network re-ranker (Zhu et al., 2015) with point-wise 
mutual information (PMI) of preposition-head pairs from the 
Subcategorisation Frame Database (Scheible et al., 2013) and SCF-aware 
margin for the hinge loss function, and report some preliminary parsing 
results on the new TweeDe dataset.