结合词典方法的图注意力网络实体抽取模型

Entity Extraction Model of Graph Attention Network Combined with Dictionary Method

常见的命名实体识别模型主要关注字词特征的抽取,对上下文语义信息的捕捉与挖掘不够充分,对中文字词边界模糊、语义歧义等问题效果较差;对此,提出了一种结合词典方法的图注意力网络实体抽取模型,能够有效减少错误信息在网络中的传播,增强对词典信息的有效利用;并通过融入“BMES”图结构和注意力机制,建立字词之间的通道,深度挖掘字词在不同语境下的关系;为了验证模型的有效性,在MRSA和Weibo数据集上进行了测试,模型在MRSA上的F1值可达94.12%,与传统方法相比提高了1.42%;后者在模型下的F1值可达61.30%,较传统方法提升了2.51%;实验结果证明,结合词典方法和交互图的图注意力网络模型在MRSA和Weibo数据集上实体识别的准确率优于传统方法,且具有一定的泛化能力.

Common named entity recognition models mainly focus on word feature extraction,insufficient capture and mining of contextual semantic information,and have poor results on problems such as fuzzy boundaries and semantic ambiguity of Chinese characters and words.In this regard,an entity extraction model of graph attention network combined with dictionary method is proposed,which can effectively reduce the propagation of error information in the network and enhance the effective use of dictionary information.By incorporating the“BMES”graph structure and attention mechanism,the channel between words is established,and the relationship between words in different contexts is deeply mined.In order to verify the validity of the model,tested on MRSA and Weibo data sets,the F1 value of the model on MRSA can reach 94.12%,which is increased by 1.42%compared with the traditional method.The F1 value of the latter under the model can reach 61.30%,which is 2.51%higher than that of the traditional method.Experimental results show that the GAN model combined with dictionary method and interactive graph is superior to traditional methods in entity recognition on MRSA and Weibo data sets,and has certain generalization ability.