首页|基于标签提示和门控模块的少样本命名实体识别

基于标签提示和门控模块的少样本命名实体识别

扫码查看
少样本命名实体识别旨在利用少量样本实现命名实体的自动识别.近年来两阶段原型网络在少样本命名实体识别任务上取得了较好的效果,但仍存在跨度检测假阳性和跨度分类原型不准确的问题.针对上述两类问题,该文提出一种基于标签提示和门控模块的少样本命名实体识别模型.在跨度检测阶段,利用标签提示信息优化句子表示,减少假阳性的出现.在跨度分类阶段,通过引入门控模块,显式地利用标签信息和样本原型进行融合,分别提取标签信息和样本信息中的有效信息,以获得更准确的原型表示.在多个数据集上的实验结果表明,该文所提出的方法相较于基准模型在F,值上能够取得10.63%的提升,并且消融实验也表明该文模型各个模块的有效性.
Few-shot Named Entity Recognition Method Based on Label Prompt and Gate Mechanism
Few-shot named entity recognition aims to achieve automatic recognition of named entities via a few sam-ples.Recently,the two-stage prototype network has achieved good results in few-shot named entity recognition tasks,but there are still problems of false positives in span detection and inaccurate prototypes in span classification.In response to the above problems,this paper proposes a few-shot named entity recognition model based on label prompts and gate mechanisms.Label prompt information is used to optimize sentence representation and reduce the occurrence of false positives in span detection.The gating module is introduced to fuse the label information and sample prototypes,so as to obtain a more accurate prototype representation in span classification.Experimental re-sults on multiple datasets show that the proposed method achieves an improvement of 10.63%in the F1 value.

few-shot learningnamed entity recognitionprototype network

吕明翰、黄琪、罗文兵、王明文

展开 >

江西师范大学计算机信息工程学院,江西南昌 330022

江西省科技基础条件平台中心,江西南昌 330002

少样本学习 命名实体识别 原型网络

国家自然科学基金江西师范大学研究生创新基金江西省教育厅科学技术研究项目

62266023YJS2022030GJJ2200354

2024

中文信息学报
中国中文信息学会,中国科学院软件研究所

中文信息学报

CSTPCDCHSSCD北大核心
影响因子:0.8
ISSN:1003-0077
年,卷(期):2024.38(9)