结合原型网络的远程监督命名实体识别方法

Distantly Supervised Named Entity Recognition Combined with Prototypical Networks

针对利用远程监督标注文本实体过程中存在实体类别标注错误问题导致模型难以有效区分各实体的类别特征,影响模型精准度的问题,本文提出一种利用原型网络过滤训练语料中标注错误样本的远程监督命名实体识别方法,利用预训练的原型网络编码正确标注实体生成类别原型表示,过滤语料中距类别原型较远的样本.实验表明,使用原型网络有效地提高了语料的标注质量,提升了模型性能.

Aiming at the problem of entity category labeling errors in the process of using distant supervision to label text entities,it is difficult for the model to effectively distinguish the category characteristics of each entity and affect the accuracy of the model.A named entity recognition(NER)method was proposed in this paper.It was designed to use pre-trained prototypical network coding to correctly label entities to generate category prototype representations,and to filter those far away samples from category prototypes in the corpus.Experiments show that the use of the prototype network can effectively improve the annotation quality of the corpus and improve the performance of the model.