Randomly Wired Graph Neural Network for Chinese NER.

Recurrent Neural Networks (RNN), tracking the features of characters and words sequentially, have achieved great success in Chinese Named Entity Recognition (NER). However, such chain-like models are inadequate to capture hierarchical and nested entities due to their poor structures. Besides, existing graph-based models on NER are limited to using graphs for modeling text sequences, which aims at obtaining an expected text-to-graph structure to segment the entities’ boundaries by hand design. Thus, we propose a Randomly Wired Graph Neural Network (RWGNN) in terms of neural network architecture, which forms a novel context encoder and automatically generate multi-directional wired pattern in graph networks by using random graph algorithms. We also incorporate lexical information with reorganized and original word representations to enhance the global dependencies. Our experimental results show that RWGNN has very promising results consistently on five different domain datasets. Furthermore, ablation experiment, complexity analysis, case study, study on sentence length, and inference speed of RWGNN also be conducted to evaluate the rationality of our model.