融合字词特征的中文嵌套命名实体识别

Chinese Nested Named Entity Recognition Integrating Improved Representation Learning Method

命名实体识别(named entity recognition,NER)是信息抽取的重要组成部分,也是时下研究热点之一。由于中文没有明确的词边界,中文命名实体识别常使用基于字进行词嵌入的方法,但基于字的词嵌入方法会使文本表征学习变得困难。因此,本文提出使用CNN(convolutional neural networks)结合Bi-LSTM(Bidirectional long short-term memory)作为表征学习方法。方法首先使用ANN(artificial neural network)对词向量进行学习,确保词向量的表征能力。然后,方法使用CNN对局部词特征进行学习,通过字词特征结合并使用Bi-LSTM学习文本上下文信息,有效提升文本表征学习能力。此外,针对命名实体识别中的实体嵌套问题,本文使用现有的合并标签标注方法并结合CRF进行序列标注。为验证本文方法有效性,本文选用几个基线模型与本文模型进行对比实验。实验结果表明,本文方法具有较优的识别性能。

Named Entity Recognition(NER)is an important part of information extraction,and it is also one of the current re⁃search hotspots.Since there is no clear word boundary in Chinese,Chinese named entity recognition often uses word-based word em⁃bedding methods,but word-based word embedding methods make it difficult to extract features of text.Therefore,we proposes to use CNN(convolutional neural networks)combined with Bi-LSTM(Bi-directional long short-term memory)as a representation learning method.The method first uses ANN(artificial neural network)to further learn the word vector to ensure the characterization ability of the word vector.Then,we use CNN to learn local word features,combines word features and uses Bi-LSTM to learn text context infor⁃mation,which effectively improves the learning ability of text representation.In addition,in view of the entity nesting problem in named entity recognition,we uses the existing combined tag labeling method and combines CRF to carry out sequence labeling.In order to verify the effectiveness of the method in this paper,we selects several baseline models to conduct comparative experiments with the models in this paper.Experimental results show that this method has better performance.