Abstract
Dependency parsers are almost ubiqui- tously evaluated on their accuracy scores, these scores say nothing of the complex- ity and usefulness of the resulting struc- tures. The structures may have more com- plexity due to their coordination structure or attachment rules.
As dependency parses are basic structures in which other systems are built upon, it would seem more reason- able to judge these parsers down the NLP pipeline. We show results from 7 individual parsers, including dependency and constituent parsers, and 3 ensemble parsing tech- niques with their overall effect on a Ma- chine Translation system, Treex, for En- glish to Czech translation.
We show that parsers’ UAS scores are more correlated to the NIST evaluation metric than to the BLEU Metric, however we see increases in both metrics.