Open Hevia opened 1 year ago
We need a method of performing coreference resolution:
Other methods worth exploring:
Papers:
An important aspect is we are not yet at the phase where we have a dataset ready to be labelled, so approaches that work with zero additional training data are desired
Scispacy's hearts patterns capture nothing sadly
Textacy does a good job of extracting SVO triples, combine it with Coreferee and it might extract some useful relations. It does sadly miss some crucial multi-token spans, but does well on single token spans!
Try out Hearst patterns from: https://github.com/allenai/scispacy first since it’s the lowest resource technique to get a rough prototype working
EDIT: Sadly does not capture much