融合语义路径与语言模型的元学习知识推理框架  

A Meta-learning Knowledge Reasoning Framework Combining Semantic Path and Language Model

作　　者：段立[1] 封皓君 张碧莹 刘江舟 刘海潮 DUAN Li;FENG Haojun;ZHANG Biying;LIU Jiangzhou;LIU Haichao(College of Electronic Engineering,Naval University of Engineering,Wuhan 430033,China;Team 91202,Chinese People's Liberation Army,Huludao 125004,China)

机构地区：[1]海军工程大学电子工程学院,武汉430033 [2]中国人民解放军91202部队,葫芦岛125004

出　　处：《电子与信息学报》2022年第12期4376-4383,共8页Journal of Electronics & Information Technology

摘　　要：针对传统推理方法无法兼顾计算能力与可解释性,同时在小样本场景下难以实现知识的快速学习等问题,该文设计一款融合语义路径与双向Transformer编码(BERT)的模型无关元学习(MAML)推理框架,该框架由基训练和元训练两个阶段构成。基训练阶段,将图谱推理实例用语义路径表示,并代入BERT模型微调计算链接概率,离线保存推理经验;元训练阶段,该框架基于多种关系的基训练过程获得梯度元信息,实现初始权值优化,完成小样本下知识的快速学习。实验表明,基训练推理框架在链接预测与事实预测任务中多项指标高于平均水平,同时元学习框架可以实现部分小样本推理问题的快速收敛。In order to solve the problems that traditional knowledge reasoning methods can not combine computing power and interpretability,and it is difficult to learn quickly in few-shot scenarios,a Model-Agnostic Meta-Learning(MAML)reasoning framework is proposed in this paper,which combines semantic path and Bidirectional Encoder Representations for Transformers(BERT),and consists of two stages:base-training and meta-training.In base-training stage,the graph reasoning instances is represented by semantic path and BERT model,which is used to calculate the link probability and save reasoning experience offline by fine-tuning.In meta-training stage,the gradient meta-information based on the base-training process of multiple relations is obtained by this framework,which realizes the initial weight optimization,and completes the rapid learning of knowledge under few-shot.Experiments show that better performance in link prediction and fact prediction can be achieved by the base-training reasoning framework,and fast convergence of some few-shot reasoning problems can be achieved by the meta-learning framework.

关 键 词：知识推理 语义路径 双向Transformer编码表示 模型无关元学习 

分 类 号：TN912.3[电子电信—通信与信息系统] TP391.1[电子电信—信息与通信工程]