A Weakly Supervised Chinese Named Entity Recognition Method Combining First-Order Logic

Named entity recognition is a key prerequisite for many tasks. However, the high cost of entity annotation limits feature learning and generalization capabilities of models. To address this problem, this paper integrates the weakly supervised method with first-order logic for Chinese named entity recognition. Firstly, a knowledge base is established by using first-order logic, tailored to the characteristics of the Chinese named entity recognition dataset. Secondly, self-training approach is introduced to address the issue of suboptimal feature learning in the model, stemming from a limited number of entity types. Lastly, the first-order logic knowledge base is incorporated into self-training approach to rectify mislabeling in the training process, which improves the generalization ability. The F1-score on the public datasets ACE05 and MSRA are improved by 2.56% and 0.35% respectively.