DensePhrases for non-answerable questions

[tteecchh]

Hello!

Thanks for the amazing work on phrase retrieval. I was recently working on a question answering project and stumbled upon this work. Our project however involves non-answerable questions as in Squad2.0.

To prevent answering, I tried to threshold the score obtained in the meta data of the results, but it seemed that it was not well calibrated (since it is not trained on non-answerable sets)

Could you please guide me on how I could train this model to perform a good distinction between answerable and non-answerable questions. Otherwise, advice on properly using score to achieve this would also be great.