Low-Resource Named Entity Recognition without Human Annotation


Zhenshan Bao, Yuezhang Wang and Wenbo Zhang, Beijing University of Technology, China


Most existing approaches to named entity recognition (NER) rely on a large amount of highquality annotations or a more complete specific entity lists. However, in practice, it is very expensive to obtain manually annotated data, and the list of entities that can be used is often not comprehensive. Using the entity list to automatically annotate data is a common annotation method, but the automatically annotated data is usually not perfect under low-resource conditions, including incomplete annotation data or non-annotated data. In this paper, we propose a NER system for complex data processing, which could use an entity list containing only a few entities to obtain incomplete annotation data, and train the NER model without human annotation. Our system extracts semantic features from a small number of samples by introducing a pre-trained language model. Based on the incomplete annotations model, we relabel the data using a cross-iteration approach. We use the data filtering method to filter the training data used in the iteration process, and re-annotate the incomplete data through multiple iterations to obtain high-quality data. Each iteration will do corresponding grouping and processing according to different types of annotations, which can improve the model performance faster and reduce the number of iterations. The experimental results demonstrate that our proposed system can effectively perform low-resource NER tasks without human annotation.


Named entity recognition, Low resource natural language processing, Complex annotated data, Cross-iteration.

Full Text  Volume 11, Number 14