Using tool to validate semi-automatically created question/answers pairs

[almugabo]

This tool has the potential to encourage sharing more data on QA from various domains. One idea I think it is worth looking into, is to combine it with semi-automatically created questions/answers pairs (and documents) and use it to validate them. For example by pre-populating the questions/answers fields.

In the excellent video tutorial https://www.youtube.com/watch?v=YhVgl70Tn_k it is mentioned at the end an open source system which - if I understood it correctly - could help in this. [Close?) Can someone share the link and we can explore if it can be used for that end ?

[fmikaelian]

Hi @almugabo

In the tutorial @andrelmfarias is talking about UnsupervisedQA, a project released recently by Facebook. The idea is to create basic question-answer pairs automatically, allowing to accelerate the creation of QA datasets.

It would be very interesting indeed to add pre-propulation & validation to the cdQA-annotator to speed up the process of annotation. Aside UnsupervisedQA, we could also add a pre-trained reader model in the loop to suggest answers to the user annotating (you would still have to write a question though).

Any help on this would be very welcomed! 😃