Title: Addressing Barriers to Reproducible Named Entity Recognition Evaluation

Abstract: To address what we believe is a looming crisis of unreproducible evaluation for named entity recognition tasks, we present guidelines for reproducible evaluation. The guidelines we propose are extremely simple, focusing on transparency regarding how chunks are encoded and scored, but very few papers currently being published fully comply with them. We demonstrate that despite the apparent simplicity of NER evaluation, unreported differences in the scoring procedure can result in changes to scores that are both of noticeable magnitude and are statistically significant. We provide SeqScore, an open source toolkit that addresses many of the issues that cause replication failures and makes following our guidelines easy.