I have some questions about your approach about wikidata-core-for-qa HOT 1 CLOSED

Hi,

Thanks a lot for your interest in our work!

1. Resolving qualifiers does not change the data, but just the way it is stored
   and probably retrieved. This can have an effect on the runtime when the dump
   is indexed appropriately, e.g. as in this project, that makes use of the pruned dump:
   https://clocq.mpi-inf.mpg.de/
   "Beyond NED: Fast and Effective Search Space Reduction for Complex Question Answering over Knowledge Bases”,
   Philipp Christmann, Rishiraj Saha Roy, and Gerhard Weikum,
   WSDM 2022.

2. We pruned several facts that are mostly not required for QA tasks.
   So what we prune are non-Wikidata facts, e.g. linkings to other KBs,
   DBs or similar. Further, entity and predicate labels are pruned from
   the dump, and the English labels maintained in a dedicated dictionary
   instead. Finally, the externalIDs are removed. They may consume quite
   some space, but can also confuse downstream QA systems. The names
   of such external predicates often have a semantically relevant surface
   form, except for the “ID”.

Further, there may well be other triples in the dump that are irrelevant
for most QA cases. For example, there is a lot of detailed information
on chemical elements, or research papers.
Therefore, I am pretty sure that further pruning is possible.

At some point we computed statistics of how many triples are pruned.
If I remember correctly, with a 2021 Wikidata dump, the size was reduced
from 12 billion to 2 billion triples, via the pruning.
To obtain the same statistics for newer dumps, one could count the lines
of the input and output files.

3. I do not have any experience with these kinds of information.
   One could potentially encode the image-data and keep it in the .ttl or .nt
   file. However, I feel that maintaining a link (either internal file path or web
   link) would be a cleaner solution.

Good luck with your project!

Regards,
Philipp

from wikidata-core-for-qa.