Bi-Lattice LSTM Model with Self-Attention for Chinese NER

In this paper, we investigate a bi-lattice-structured LSTM model for Chinese NER based lattice LSTM model, which encodes a sequence of input characters as well as all potential words that match a lexicon. Compared with lattice LSTM model, our model fully combines the forward and backward information of each character in the word. And we also introduce the highway model to solve the gradient problem to some extent. In order to capture the global sequence information from multiple subspaces, we introduce the self-attention mechanism. Experiments on various datasets show that our model outperforms both original model and other advanced models without relying on external resources such as dictionaries and multi-task joint training.