Incorporating chunking information for Chinese named entity recognition using neural networks

Most named entity recognition (NER) systems use supervised machine learning methods, including linear models and neural networks. However, both methods require large amounts of annotated data. In this paper, we utilize chunking resources to improve the performance of NER without annotating more data and investigate two approaches to incorporate chunking information in the character-level neural network framework for Chinese NER. The first approach is multi-task learning, which falls into the framework of sharing common feature representation of Chinese chunking and NER task. The second approach is based on stacking, which uses the features provided by a trained chunking model to guide the NER model. Experimental results on OntoNotes 4.0 corpus show that compared with the models only using the NER corpus, the NER models which utilize the chunking resources can improve the performance of NER.