改进路由机制的元学习少样本文本分类模型  

Few-shot Text Classification Model Based on Meta-learning with Improved Routing Mechanisms

作　　者：荆沁璐 冯林[1] 王旭 龚勋[2] 胡议月 JING Qin-lu;FENG Lin;WANG Xu;GONG Xun;HU Yi-yue(School of Computer Science,Sichuan Normal University,Chengdu 610101,China;School of Computing and Artificial Intelligence,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]四川师范大学计算机科学学院,成都610101 [2]西南交通大学计算机与人工智能学院,成都611756

出　　处：《小型微型计算机系统》2023年第11期2392-2400,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61876158)资助;四川省重点研发项目(2023YFS0202)资助。

摘　　要：深度学习模型已在文本分类领域得到了广泛应用.然而,深度神经网络在处理少样本文本分类任务时其有效性易受噪声、同类样本点分布不均衡等问题的影响.为此,提出改进路由机制的元学习少样本文本分类模型.模型对胶囊网络的动态路由机制做出两种改进,针对噪声干扰问题,提出基于交互信息的路由机制,捕获同类文本间的交互信息来引导模型加强重要特征,减弱噪声影响;针对文本样本点分布不均衡的问题,提出基于距离系数的路由机制,引入距离系数指导权重分配过程以更好地划分原型空间.然后,将二者学习到的类原型进行融合,以充分捕获少样本文本特征信息.实验结果表明,相对其它少样本文本分类任务的基线方法,该文模型具有更优的少样本文本预测能力,并且收敛速度更快.Deep learning has been widely used in text classification.However,deep neural networks face challenges on few-shot text classification,such as text noise and unbalanced data distribution.Therefore,a few-shot text classification model based on meta-learning with improved routing mechanisms is developed.The proposed model makes two strategies to improve the dynamic routing mechanism.We propose the Cross Information augmented Routing,which captures cross information among the same category data,to reduce unnecessary noise effects and strengthen important features.Meanwhile,we propose the Distance Coefficient guided Routing,which introduces distance coefficient,to delineate the prototype space reasonably.Two higher-quality prototypes are learned through the above routings.Then,the learned prototypes are fused to utilize textual information.Compared with some classical few-shot text classification methods,the proposed model has better performance and faster convergence ability.

关 键 词：少样本文本分类 路由机制 元学习 深度学习 胶囊网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]