Partially Supervised Named Entity Recognition via the Expected Entity Ratio Loss

Abstract We study learning named entity recognizers in the presence of missing annotations. approach this setting as tagging with latent variables and propose a novel loss, Expected Entity Ratio, to learn models systematically tags. show that our is both theoretically sound empirically useful. Experimentally, we find it meets or exceeds performance strong state-of-the-art baselines across variety languages, annotation scenarios, amounts labeled data. In particular, significantly outperforms previous methods from Mayhew et al. (2019) Li (2021) by +12.7 +2.3 F1 score challenging only 1,000 biased annotations, averaged 7 datasets. also that, when combined approach, sparse scheme exhaustive for modest budgets.1