汉译藏传佛教典籍中的神灵命名实体识别方法研究

Research on Deities Named Entity Recognition Method in Tibetan Buddhist Classics Translated in Chinese

命名实体识别是自然语言处理中的一项基础性关键任务。针对汉译藏传佛教典籍中各种神灵名称难以识别的问题,提出一种基于BERT预训练语言模型、双向长短时记忆网络(BiLSTM)和条件随机场(CRF)的多神经网络融合方法BERT-BiLSTM-CRF-a。该方法使用BERT代替浅层网络训练字向量,充分表征字的多义性;引入注意力机制的权重思想将BiLSTM层的前向和后向隐层向量加权后再拼接,进一步提高了上下文特征的有效利用率;最后使用CRF模型输出序列上的最优标注结果。实验表明,该方法在测试集上准确率达95.2%,较传统的BiLSTM-CRF模型提升7.6%,召回率也高出8.7%,因此能够应用于汉译藏传佛教典籍中神灵名称识别任务。

Named entity recognition is a basic task in natural language processing,but also an important foundation of knowledge graph construction.In view of the fact that the naming patterns of various deities are not fixed and difficult to identify in Tibetan buddhist classics translated in Chinese,a multi-neural network fusion method BERT-BILSTM-CRF-a was proposed based on the BERT pre-training language model,Bidirectional Long Short-Term Memory(BiLSTM)and Conditional Random Field(CRF).The model used the BERT pre-training method instead of traditional shallow neural network to train word vector to fully represent the polysemy of the words.In addition,the weight thinking of attention mechanism is introduced to weight the forward and backward LSTM hidden layer vectors before concatenate to further improve the effective utilization of context features.Finally,the CRF model is used to label texts and the to output optimal label resulted on the sentence sequence.The experimental results showed that the model has an accuracy of 95.2%in the test set,7.6%higher than BiLSTM-CRF model,and the recall rate is also 8.7%higher than that of BiLSTM-CRF model.Therefore,BERTBILSTM-CRF-a can be effectively applied to the task of deities named entity recognition in the field of Tibetan Buddhism.