Closed aced125 closed 3 years ago
Neuralcoref has been nothing but a headache since it was added. I think a better strategy that I want to move to is to bring your own coreference model that works with spacy. I really want to get rid of the spacy/coreference dependency because it has been causing people issues with installation.
Hey authors,
Great repo so far. An issue: when I try to do run the
body
in the example (on the Chrysler building sale) through the neuralcoref code in the repo, it doesn't actually work...For example, here is running the
body
through neuralcoref, and examining the clusters.This the code used at the moment in
modelprocessors.py
.However, if we try this, instead, it works:
Clearly, there are a lot of issues here (e.g "Blackstone Group (BX) bought Blackstone Group (BX) for $1.3 billion 2015").
So it is almost better that this repo is working without neuralcoref.
However, neuralcoref gets 65 F1 on OntoNotes, whereas in 3 years the state of the art has progressed to Bert or Span-Bert (~80 F1). So maybe, we should use those instead?
https://github.com/mandarjoshi90/coref