Document-Level Named Entity Recognition With Q-Network

Named entity recognition (NER) is typically viewed as a sequence labeling problem, where the solution is optimized in sentence level. In this paper, we explore utilizing document-level label consistency to improve NER performance by reinforcement learning. The process entails searching similar mentions and reconciling the labels, which are repeated until sufficient evidence is collected. We employ a Q-network, trained to maximize the total reward that reflects labeling accuracy while penalizing extra effort. On three publicly available datasets, our approach achieves an F-1 of 90.75%, outperforming the ensemble method (89.80%) by 0.95% and the best base tagger (88.16%) by 2.59% (The source code is available at https://github.com/KrisWentaoWong/dqnner/).