While several state-of-the-art approaches to dialogue state tracking (DST)
have shown promising performances on several benchmarks, there is still a
significant performance gap between seen slot values (i.e., values that occur
in both training set and test set) and unseen ones (values