Negation’s Not Solved: Reconsidering Negation Annotation and Evaluation

Abstract:

Objective To characterize and ameliorate the weaknesses of clinical negation detection techniques across corpora that have different annotation schema. Materials and Methods Named entities and negation attributes have been manually annotated in several corpora, including the new SHARPn NLP Seed Corpus, the 2010 i2b2/VA NLP Challenge assertion annotations, the MiPACQ Corpus, and the NegEx Test Set. We analyze each of these corpora through the lens of two contrasting systems: a rule-based NegEx baseline, and the SHARP machine-learning-based attribute discovery tool. Results