vdobrovolskii
commented
2 years ago Hi, thanks for the interest!
I am not sure I completely understand the question. I'll try to answer the way I understand it, please clarify if needed.
First of all, coreference resolution with word representations works the same way as with span representations, i.e. each mention can be only assigned to one cluster.
Then, the assumption is that there is one-to-one correspondence between all the valid spans in the text and their head words. That is, in "A big black cat sat on a mat" there will three valid spans for Ontonotes: "a big black cat" -> "cat", "sat" -> "sat" and "a mat" -> "mat".
So when the coreference links are found between individual words, each word is assigned to only one coreferent cluster and is then converted to only one span, so there are unique non-overlapping coreferent clusters of spans.
Now there are (very few) cases when two spans in the original Ontonotes dataset share the same head word. Almost all of them are related to conjunction. In those cases "A" and "A & B" are different spans with the same head word, "A". In our implementation such cases were simply discarded from the training set, because they were few and we were able to perform well, even though we couldn't predict any of such cases during inference.
To adopt this system under requirements that such cases should work, I would try assigning an artificial "head word" to such spans, for instance, the conjunction itself.
yhcc
First of all, thank you to release the code. Really nice work, so glad to see advances without enumerating span combinations. But I have one question about the "word coreference" part. Since the one word may be assigned to different coference clusters, such as for "a b c d e", "a b c" and "c" are both coferences and belong to different cluseters (if 'c' is the head word for "a b c"), how will this "word coreference" stage deal with it?