Closed cdhx closed 1 year ago
Hi,
We are happy that you find our work useful!
The QA system used in this work is GraftNet, for which the code is publicly available at https://github.com/haitian-sun/GraftNet. The work was published at EMNLP 2018 (https://arxiv.org/abs/1809.00782).
We also have a recent QA system from our group, that you might want to check out at https://explaignn.mpi-inf.mpg.de. The work is accepted at SIGIR 2023, and the code and pre-trained models are publicly available here: https://github.com/PhilippChr/EXPLAIGNN.
Regards, Philipp
Thank you for your kind response.
I have another question regarding the migration of graftnet to LC2. Does the migration solely involve changing entity linking from Freebase to Wikidata?
Additionally, if I intend to implement a baseline on the Wikidata dataset, e.g. LC2, would you kindly provide any recommended baselines?
Once again, I truly appreciate your patience and assistance.
GraftNet requires a set of KB-facts as input. In our experiments, we gave it either KB-facts as provided by the CLOCQ search space reduction functionality, or KB-facts retrieved for disambiguated entities (via ELQ, TagME,...). So having an entity linker + a method to retrieve KB-facts for linked entities (e.g. CLOCQ) would be sufficient.
For complex QA, you could e.g. compare with UNIQORN (https://uniqorn.mpi-inf.mpg.de) in KB-only setting, that also has results readily available on LC-QuAD 2.0. Again, EXPLAIGNN in KB-only setting would also be an option.
Finally, I would also like to point to our latest benchmark release, CompMix (https://qa.mpi-inf.mpg.de/compmix/), that has complex questions with answers grounded in Wikidata. A write-up with additional details will follow soon. We would also provide results of EXPLAIGNN. Questions may often require heterogeneous sources to be answered, but comparison in KB-only setting would also be possible. In case you are interested, let me know, and I could share results of EXPLAIGNN in KB-only setting on this dataset.
Hello,
Connecting you with Soumajit and Jesujoba, who can help you with intermediate outputs and results of UNIQORN on LC-QuAD 2.0, so that you can more easily compare your method with UNIQORN.
Regards, Rishi.
On Thu, May 11, 2023 at 2:25 PM Philipp Christmann @.***> wrote:
GraftNet requires a set of KB-facts as input. In our experiments, we gave it either KB-facts as provided by the CLOCQ search space reduction functionality, or KB-facts retrieved for disambiguated entities (via ELQ, TagME,...). So having an entity linker + a method to retrieve KB-facts for linked entities (e.g. CLOCQ) would be sufficient.
For complex QA, you could e.g. compare with UNIQORN ( https://uniqorn.mpi-inf.mpg.de) in KB-only setting, that also has results readily available on LC-QuAD 2.0. Again, EXPLAIGNN in KB-only setting would also be an option.
Finally, I would also like to point to our latest benchmark release, CompMix (https://qa.mpi-inf.mpg.de/compmix/), that has complex questions with answers grounded in Wikidata. A write-up with additional details will follow soon. We would also provide results of EXPLAIGNN. Questions may often require heterogeneous sources to be answered, but comparison in KB-only setting would also be possible.
— Reply to this email directly, view it on GitHub https://github.com/PhilippChr/CLOCQ/issues/8#issuecomment-1543903052, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB4DFHWT5BQXOH766IYGD2LXFTLFHANCNFSM6AAAAAAX4EG5TU . You are receiving this because you are subscribed to this thread.Message ID: @.***>
Thank you for your patience!
I will consider using UNIQORN as a baseline and then contact Soumajit and Jesujoba if needed.
Thanks again for your help!
Thank you for publishing this tool. It has been really helpful to me!
I am also interested in the QA system mentioned in your paper, "Beyond NED: Fast and Effective Search Space Reduction for Complex Question Answering over Knowledge Bases," especially for LC-QuAD 2.0.
I was wondering if this QA system is open-sourced? If it is, could you please provide me with a link or any relevant information on how to access it?
Thank you again for your contributions to the community. I am looking forward to your reply.