Relation Extraction

[Hevia]

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

[Hevia]

We need a method of performing coreference resolution:

• https://explosion.ai/blog/coref
• https://github.com/explosion/coreferee (might be the same system as above, unsure if neuralcoref is still being maintained)

Other methods worth exploring:

• https://github.com/mmxgn/spacy-clausie
• https://github.com/erre-quadro/spikex (No longer actively maintained)
• https://textacy.readthedocs.io/en/latest/api_reference/extract.html#triples (works well!)
• https://github.com/ibm/zshot
• https://github.com/jakelever/kindred (this might prove especially useful for gene RE)
• https://github.com/graphbrain/graphbrain (How are they building their Semantic Hypergraph?)
• https://github.com/explosion/holmes-extractor
• https://github.com/dair-iitd/OpenIE-standalone
• https://github.com/dwadden/dygiepp (Especially useful since labels are standardized + spaCy integration)

Papers:

• https://aclanthology.org/C92-2082.pdf

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

[Hevia]

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!