Negation detection

[hammer]

One issue that @eric-czech has noticed during error analysis for relation extraction: it's hard to labeling functions that exclude negated relations.

It may be worth looking at how others have approached negation detection within the context of relation extraction and other tasks.

[hammer]

• A simple algorithm for identifying negated findings and diseases in discharge summaries (2001): chapmanbe/negex seems to be the most common baseline model for negation detection.
• DEEPEN: A negation detection system for clinical text incorporating dependency relation into NegEx (2015): jessica-glover/deepen extends NegEx by making use of dependency parse information.
• NegBio: a high-performance tool for negation and uncertainty detection in radiology reports (2018): ncbi-nlp/NegBio from Zhiyong Lu's consistently excellent NLP group at the NIH is probably where I'd start, even though it claims to be specific to radiology reports. Also cited by the Snorkel team in Cross-Modal Data Programming Enables Rapid Medical Machine Learning (2019) so Alex may have some thoughts on their approach.