the correlation coefficient is negative

[23LuZ]

Hi, thanks for your helpful repo.

When i try to reproduce the experiment results as you mentioned, something seems wrong. The acc of ruber-unrefer on the test set is about 0.6, but the correlation coefficient of ruber-unrefer scores and human scores is negative. And the acc of bert-ruber-unrefer on the tset set is about 0.84, but the correlation coefficient of bert-ruber-unrefer scores and human scores is also negative. (i checked human scores, they are all normal)

I am really confused, do you have any idea about this situation? Look forward to your reply~