基于循环和卷积神经网络融合的中文命名实体识别与应用

Study and Application of Chinese Named Entity Recognition Based on the Fusion of Recurrent and Convolutional Neural Network

针对中文命名实体识别中循环神经网络不能很好地处理长序列问题,以及用单一向量去表征汉字时,由于汉字存在多义性而导致识别结果不佳的问题,提出一种识别效果更好的方法——BLDC-NER模型.首先利用BERT(Bidirectional Encoder Representations from Transformers)预训练模型根据字的上下文语境生成字的动态语义向量,然后将字向量序列分别通过双向长短时记忆网络层和膨胀卷积层进行语义编码,融合2个网络层输出的语义向量,经过条件随机场得到最终结果.试验结果表明:BLDC-NER模型在训练过程中比单一循环神经网络收敛速度更快,识别效果更好,在MSRA、RESUME公开数据集上的F 1值分别达到了94.78%、95.68%;另外,将BLDC-NER模型应用在建筑施工安全事故领域,在自制的数据集上F 1值为95.24%.

Aiming at the problem that the cyclic neural network in Chinese named entity recognition cannot han-dle the long sequence problem well,and when a single vector is used to represent Chinese characters,the recog-nition results are poor due to the ambiguity of the Chinese characters,a better recognition effect is proposed,that is Method-BLDC-NER model.The BERT(Bidirectional Encoder Representations from Transformers)pre-training model is first used in the BLDC-NER to generate the dynamic semantic vector of the word according to the con-text of the word,and then passes the word vector sequence through the bidirectional long-and short-term memory network layer and the expanded convolution layer for further.The semantic encoding of fusion combines the se-mantic vectors output by the two network layers,and the final result through a conditional random field is ob-tained.The experimental results show that the BLDC-NER model converges faster than a single loop neural net-work during the training process and has better recognition results.The F1 values on the MSRA and RESUME public data sets have reached 94.78%and 95.68%respectively;in addition,the BLDC-NER model is applied in the field of construction safety accidents,and the F1 value is 95.24%on the self-made data set.