DeepBank_OneOne

Background

In early 2014, we are preparing a minor update of DeepBank to fix a handful of imperfections in the original release:

• missing segments: wsj09e, wsj12e, wsj20e
• synchronize final REPP treatment of ellipses and profiles
• benefit from bug fixes in EDS export
• corrections in head assignment for DT conversion
• pick up DM conversion improvements (from SDP release)
• include GML-based WDC profiles in release

Candidate fixes to apply (which would require an increment of the grammar version):

• ESD-related renaming (e.g. PT and its values)
• unknown_card and its CARG
• known sources of systematic annotation consistency (if any; e.g. splitting U.S. as sandwiched period).
• rule naming (recover the one case using an underscore in the first field)
• predicate name on not ... but

n-nh_j-cpd_c => n-j_j-cpd_c n-nh_j-t-cpd_c => n-j_j-t-cpd_c n-nh_v-cpd_c => n-v_j-cpd_c

Finally, a few aspects of ‘packaging’ that did not make it into the first version

• export into MRG-style trees with PTB-style tokenization
• investigate relatively high failure percentage in robust meaning construction (and DM conversion)