WuXuan374
commented
1 year ago Hi, we haven't tried to adapt this framework to the GrailQA dataset, so we can only provide some suggestions based on our understanding. We'd like to follow the sequence in README.md.
- (2) Parse SPARQL queries to S-expressions. This part of code is modified based on RNG-KBQA. Since GrailQA is also based on Freebase, this part of code should work. But it will be better if you run a coverage test on SPARQLs from GrailQA, to check if there are SPARQLs our code fails to handle.
- (3) Retrieve Candidate Entities and (4) Retrieve Candidate Relations.
These two steps aims to provide candidate entities/relations for our multi-task model. Both
migrating our method for GrailQAandusing results of other linking toolsshould work. - (5) Prepare data for multi-task model. We guess there should be some file format issue for a new dataset.
- (6) Generate Logical Forms through multi-task learning.
We think the multi-task model itself should work without modification. But similar as (2), when converting S-expression back to SPARQL in
eval_topk_prediction_final.py, you may need to run a coverage test for unhandled cases from GrailQA.
These are possible changes we can think of at current stage. Hope this helps.
luoxindi
Hello, thanks for your repository, it inspires me a lot. For the QA dataset, I wonder how to adapt this framework to the GrailQA dataset? What changes do I need to make to this framework？