Inter-Annotator Agreement and the Upper Limit on Machine Performance: Evidence from Biomedical Natural Language Processing.

Human-annotated data is a fundamental part of natural language processing system development and evaluation. The quality of that data is typically assessed by calculating the agreement between the annotators. It is widely assumed that this agreement between annotators is the upper limit on system performance in natural language processing: if humans can't agree with each other about the classification more than some percentage of the time, we don't expect a computer to do any better. We trace the logical positivist roots of the motivation for measuring inter-annotator agreement, demonstrate the prevalence of the widely-held assumption about the relationship between inter-annotator agreement and system performance, and present data that suggest that inter-annotator agreement is not, in fact, an upper bound on language processing system performance.