Open alexanderkoller opened 5 years ago
Well, the Anchoring F-score holds us down anyway. I thought we could add two columns to amconll that indicate where for every word, the span which it induces begins and where it ends. I was thinking of a simple tagging approach, like the relative-POS-based encoding in this paper: https://www.aclweb.org/anthology/N19-1077 (which I find interesting that this works so well!).
I was thinking about training an NN to predict spans too. In what way would the relative-POS-based encoding solve our problem?
Whatever we do, it should be a simple solution, otherwise we won't be finished in time.
In the paper, this "pointer" encodes where to find the head, in our case, we would use two pointers: one pointing at the token where the span begins, and one where the span ends. Not sure, whether I understood your question.
Yes, I think this is the simplest way of using a NN in this case. What did you think about?
If EDM is used in the mtool evaluation, figure out a way to predict the spans better.