Nested Named Entity Recognition with Partially-Observed TreeCRFs

Named entity recognition (NER) is a well-studied task in natural language processing. However, the widely-used sequence labeling framework difficult to detect entities with nested structures. In this work, we view NER as constituency parsing partially-observed trees and model it TreeCRFs. Specifically, all labeled spans observed nodes tree, other latent nodes. With TreeCRF achieve uniform way jointly To compute probability of partial marginalization, propose variant Inside algorithm, Masked that supports different inference operations for (evaluation observed, marginalization latent, rejection incompatible observed) efficient parallelized implementation, thus significantly speeding up training inference. Experiments show our approach achieves state-of-the-art (SOTA) F1 scores on ACE2004, ACE2005 dataset, shows comparable performance SOTA models GENIA dataset. We release code at https://github.com/FranxYao/Partially-Observed-TreeCRFs.