Open apoorvumang opened 2 years ago
Hi Apoorv, I have tried EXAQT on CronQuestions last year but due to some reasons (one is the number of CronQuestions is large and another is I was busy with other things) I did not complete it. If you are trying EXAQT on CronQuestions, the TagMe (I use WAT, improved TagMe https://sobigdata.d4science.org/web/tagme/wat-api) and ELQ (https://github.com/facebookresearch/BLINK) for NERD are needed. CLOCQ (https://clocq.mpi-inf.mpg.de/) is used to extract one-hop facts for the NERD entities of each question and is also used to extract two-hop temporal facts for the completed GST subgraphs. To obtain question-relevant facts, we need to create the training dataset and train the BERT classifier. The pipeline of EXAQT is a little bit long so I think I can restart the work of trying EXAQT on CronQuestions and share the data with you.
Since CronQuestions dataset has gold topic entities, I think the NERD step can be removed from the EXAQT pipeline.
Hi Zhen, thanks for the response
Since CronQuestions dataset has gold topic entities, I think the NERD step can be removed from the EXAQT pipeline.
Yes we can probably do away with the NERD step in the pipeline
The pipeline of EXAQT is a little bit long so I think I can restart the work of trying EXAQT on CronQuestions and share the data with you.
That would be extremely useful! Please let me know if you make any progress on this front. I am also trying to understand the steps of EXAQT, and even data of intermediate steps would be useful
You're welcome. When the pipeline is done, I will let you know.
Hi Zhen
Did you have any luck in trying EXAQT on CronQuestions dataset (https://github.com/apoorvumang/CronKGQA)? If not, what kind of dataset processing do you think is needed to get EXAQT to work on CronQuestions?
Thanks Apoorv