question about code

[onehaitao]

hi，your code is great, but I still have a question about code (biaffine_ner_model.py from line 249 to line 256).

 top_spans = [[] for _ in xrange(len(sentences))]
    for i, type in enumerate(np.argmax(span_scores,axis=1)):
      if type > 0:
        sid, s,e = candidates[i]
        top_spans[sid].append((s,e,type,span_scores[i,type]))

 top_spans = [sorted(top_span,reverse=True,key=lambda x:x[3]) for top_span in top_spans]

Here, you sort the predicted spans in the descending order of span sores. I read your code and find span scores is logits rather than probability , and I wonder why it is.

[juntaoy]

Hi, the reason we use the logits instead of probability is that we don't want to select spans that the system is not sure. Say we have a span that has a very low score for every category, and that means the system is unsure and to use logits directly we do understand how certain the system is and not biased by scores from other categories.

[onehaitao]

Thank you very much.