混合神经网络的中文地名识别方法

Chinese place name recognition based on hybrid neural network

针对互联网媒体数据中地名信息表达不规范、实体边界不清晰、地名简化表达问题,该文提出了一种混合神经网络的中文地名识别方法。通过ALBERT层学习字级别特征表达与BiLSTM层提取文本上下文语义特征,由CRF层获得全局最优标记序列,更有效地识别并提取中文地名。数据测试表明:相较于BiLSTM的地名识别模型,本文模型精确率提高12.89%,F1值提高10.83%;相较于BiLSTM-CRF的地名识别模型,本文模型精确率提高3.56%,F1值提高2.1%;相较于ALBERT-CRF的地名识别模型,本文模型精确率提高1.22%,F1值提高0.72%。

Aiming at the problem of irregular expression of place name information in Internet media data,unclear entity boundaries,and simplified expression of geographical names,a hybrid neural network Chinese place name recognition algorithm is proposed.The character level features are learned in the ALBERT representation layer with the context features extracted in the BiLSTM layer to form the feature matrix.Finally,the optimal tag sequence is generated in the CRF layer.The actual data test shows that:compared with the BiLSTM-based geographical name recognition model,the accuracy of the model in this paper is increased by 12.89%,and the F1 value is increased by 10.83%;compared with the BiLSTM-CRF-based geographical name recognition model,the accuracy of the model in this paper is increased by 3.56% and the F1 value is increased 2.1%;Compared with the geographical name recognition model based on ALBERT-CRF,the accuracy of this model is increased by 1.22%,and the F1 value is increased by 0.72%.