Closed yhcc closed 2 years ago
No worries! I am glad this work might be interesting to anyone :)
"head2span" contains tuples [head, span_start, span_end], so indeed the span in question (175, 192) has been filtered out.
I checked the case and (175, 192) is "Commander - in - chief Zhu De and Vice Commander Peng Dehuai of the Eighth Route Army", which is what I mentioned in issue #2: because the span contains two conjuncts ("Commander - in - chief Zhu De" and "Vice Commander Peng Dehuai"), the heads of the left conjunct and the whole phrase are the same.
In such cases the data preparation script just picks the shortest option.
As I have already said, if this is critical, you can try changing the data preparation script to pick something else as span head in such cases, for instance, the conjuction itself ("and" in this example). I think this should work, but it will require retraining the model, becase it has not seen such heads and has not been trained to predict such spans.
So when conducting evaluation, the code will use the span_clusters
to evaluate (if the word_level_conll is False). I got it, thank you for your patient reply.
Yes, during evaluation span_clusters are used, so all the spans that cannot be predicted are treated as false negatives.
Feel free to ask any further questions :)
Sorry to interrupt you again. I find some span presented in the
span_clusters
not presented inhead2span
, such as a processed sample below[175, 192]
in thespan_clusters
is not presented in thehead2span
(The most similar one might be[181, 175, 182]
, and[175, 182]
is also a span belong to another coreference cluster). Do you have any idea why this may happen? Is this because the head word overlap, the processed script only keep one span?