How to select candidate answers when predict

[songjunjun]

In practice, we do not have candidate answers, so how to select from corpus?
We can compute relevance between query and response, or similarity between querys, it seems the later is better, especially when tf-idf is baseline model, so why not the later?
Thanks very much!

[XingLe0917]

I have same question about it.