In cross-lingual dependency annotation projection, information is often lostduring transfer because of early decoding. We present an end-to-end graph-basedneural network dependency parser that can be trained to reproduce matrices ofedge scores, which can be directly projected across word alignments. We showthat our approach to cross-lingual dependency parsing is not only simpler, butalso achieves an absolute improvement of 2.25% averaged across 10 languagescompared to the previous state of the art.
展开▼