About pruning spans

[zhiqihuang]

Hi,

After reading the paper, I have a question about pruning spans.

In section 5, the framework prunes the candidate spans during both training and evaluation. I think during evaluation is reasonable since the weights for word vectors are tuned.

But during the early stage of training, the weights for word vectors are not learned (e.g., random initialized in the first epoch), could keeping top lambda*T spans ordered by mention scores discard potential valid mentions?

Thanks.