flitternie / graphq_ir Goto Github PK

OR any suggestion for chinese kbqa's intermediate representation ? Thanks!

@Flitternie
hi,
From the link you offered for the Overnight dataset, there are only questions and logical forms included, but no corresponding answers are provided, right? How to get the answers? I find that we can execute the logical forms in the dataset to get the answers. Are there other methods to get the answers? Or if some other papers released the answers?
Looking forward to your reply. Very appreciate.

It a great work! Do you prepare to release the graphq_trans tool recently?

Great job! And I have a question:
When you use KQAPro, how do you get these five different divisions like "Multi-hop, Qualifier, Comparison, Logical, Count, Verify, Zero-shot"?

Hey greetings,
I was trying to use the pretrained checkpoint given in the repo to output the IR with natural questions example: "What is the longest river in China." . I would then expect the IR representation of this question according to the paper?
Unfortunately the model checkpoint in the repo gives the question back as the output. Is there anything I am missing?
Please let me know if I should be doing something different.

I'm interested in testing the implementation of the text-to-Cypher model, and it seems like that corresponds to the MetaQA dataset. Do you have an available model checkpoint for the text-to-Cypher model?

Hello,

Thanks for your wonderful work! I was wondering if you have compared the performance of your IR with other IRs like FUNQL, and s-expression since there only exists some intermediate representations.

Thanks

Hi, I found that some programs cannot be converted into graphq. For example, when I run:
from graphq_trans.kopl.translator import Translator translator = Translator() a = [{ "function": "Find", "dependencies": [], "inputs": [ "My Neighbor Totoro" ] }, { "function": "Find", "dependencies": [], "inputs": [ "German" ] }, { "function": "Relate", "dependencies": [ 1 ], "inputs": [ "original language of film or TV show", "backward" ] }, { "function": "Find", "dependencies": [], "inputs": [ "Hannah Arendt" ] }, { "function": "And", "dependencies": [ 2, 3 ], "inputs": [] }, { "function": "SelectBetween", "dependencies": [ 0, 4 ], "inputs": [ "duration", "greater" ] } ] a = translator.to_ir(a)
I got a warning " 1:130 no viable alternative at input 'Find(My Neighbor Totoro)Find(German)Relate(original language of film or TV showbackward)Find(Hannah Arendt)And()SelectBetween'", and the conversion failed.
Is it normal or how can I solve it?

Hi, sorry to disturb you again. I have a question about the test accuracies on OVERNIGHT in the paper. After reading the code, I suppose the results of GraphQ IR correspond to a single trained model. But for GraphQ IR*, you select one best-performed model for each of the eight domains. So the results of GraphQ IR* actually correspond to eight trained models rather than a single one. Is it right? I would appreciate it if you can explain it to me.