Determining the difficulty of Word Sense Disambiguation.

Highlights: [•] We explore estimating WSD performance on a range of ambiguous biomedical terms. [•] We evaluate the difficulty predictions against the output of two WSD systems. [•] Supervised methods are the best predictors but limited by labeled training data. [•] Unsupervised methods all perform well and can be applied more widely. [•] Best performance was obtained using the relatedness measure proposed by Lesk. [ABSTRACT FROM AUTHOR]