.png){kind=icon}
.png){kind=icon}
Return
Semi-supervised named entity recognition with data augmentation by structured consistency training
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127464.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Named entity recognition (NER) is a fundamental task in the natural language understanding pipeline, yet state-of-the-art NER models often require a large corpus of human-annotated sequences. In this work, we present a novel semi-supervised NER method by leveraging data augmentation on bountiful unlabeled data. Consistency training is utilized to regularize model predictions to be invariant to small noises induced by the data augmentation. Moreover, to capture the structured prediction nature of named entity recognition, we propose an unsupervised structured consistency loss to promote consistency in inter-token dependencies between the original and the augmented sequences. In principle, the proposed structured consistency training can be used to convert any supervised NER model to a semi-supervised one and thus can readily leverage the advancement in supervised NER. We further provide a theoretical analysis of the required number of labeled sequences to achieve a certain error rate. The experiments are conducted on two NER datasets, and the results demonstrate the effectiveness of the proposed approach.
Keywords:
Named entity recognition
Semi-supervised learning
Data augmentation
Consistency training