relational matirx is very sparse in Chinese dataset. When I tested my implemention of multi_head_selection, the loss of relation was very small. Because most relations were 'N', the samples are unbalanced across categories. I tried some other methods to ease this problem, but failed.
Therefore, I wonder whether you meet this problem too?
relational matirx is very sparse in Chinese dataset. When I tested my implemention of multi_head_selection, the loss of relation was very small. Because most relations were 'N', the samples are unbalanced across categories. I tried some other methods to ease this problem, but failed. Therefore, I wonder whether you meet this problem too?