BiGCNN: Bidirectional Gated Convolutional Neural Network for Chinese Named Entity Recognition

  • Tianyang Zhao ,
  • Haoyan Liu ,
  • ,
  • Changzhi Sun ,
  • Dongdong Zhan ,
  • Zhoujun Li

25th International Conference on Database Systems for Advanced Applications (DASFAA 2020) |

Recent advances on Chinese named entity recognition (NER) are mostly based on the recurrent neural network (RNN). Since RNNs are limited in parallel processing, some works apply the convolutional neural network (CNN) to perform NER. However, existing CNN-based models fail to explicitly distinguish the preceding and subsequent contexts, so they are difficult to handle cases that are sensitive to the location of the contexts. Moreover, they pay equal attention to the context within a convolution kernel, while not all the information is useful for semantic understanding. In this paper, we propose a novel CNN-based model, Bidirectional Gated Convolutional Neural Network (BiGCNN), to differentiate the entity-related information between preceding and subsequent contexts and filter out the convolution information adaptively. By incorporating automatic segmentation and glyph information, BiGCNN outperforms state-of-the-art models on four Chinese NER datasets. Additionally, benefiting from the parallelism processing, the proposed method enjoys higher training and testing efficiency, e.g., 12.04 times faster than RNN-based models, while with better performance.