基于两阶段网络和提示学习的少样本中文命名实体识别

Few-shot Chinese named entity recognition based on two-stage networks and prompt learning

基于提示学习的方法在处理少样本句子级分类任务时效果较好,然而在处理字符级的命名实体识别任务时,现有采用枚举手动构建提示模板的方法效率低下且性能不稳定。针对此问题,提出基于两阶段网络和提示学习的少样本中文命名实体识别方法。第一阶段网络,创新性地引入文本生成模型T5,利用提示学习的思想,将文本生成任务改造为完形填空的形式以自动生成提示模板;第二阶段网络,使用当前主流的BERT-BiLSTM-CRF网络架构进行训练。通过消融实验,探讨了提示模板更合理的嵌入方式,在Weibo、Resume和MSRA语料库上的实验结果F1值分别达到了73.58%、96.63%和95.67%,在多组对比实验中效果均为最优,表明了该方法可有效提升少样本中文命名实体识别的效果。

Prompt learning is more effective in dealing with the few-shot sentence-level classification task,but the existing method of manually constructing prompt templates using enumeration is inefficient and unstable to solve the character level task of named entity recognition.Aiming at this problem,this paper proposes a few-shot Chinese named entity recognition method based on two-stage networks and prompt learning.The first stage innovatively introduces the text generation model T5,and uses the prompt learning to transform the text generation task into a form of fill-in-the-blank to automatically generate prompt templates.The second stage adopts current mainstream frameworks BERT-BiLSTM-CRF for training.This paper explores a more reasonable embedding method for prompt templates through ablation experiments.Experimental results show that the best F1 values on Weibo,Resume and MSRA corpus reach 73.58%,96.63%and 95.67%,respectively,the highest among other models.Thus,the method can improve the recognition of Chinese named entities with few-shot.