基于四元数门控图神经网络的脚本事件预测  

Script event prediction based on a quaternion-gated graph neural network

作　　者：车飞虎 张大伟[1] 邵朋朋 杨国花 刘通 陶建华[1,2,3] CHE Feihu;ZHANG Dawei;SHAO Pengpeng;YANG Guohua;LIU Tong;TAO Jianhua(National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;School of Artificial Intelligence,University of Chinese Academy of Sciences,Beijing 100049,China;CAS Center for Excellence in Brain Science and Intelligence Technology,Beijing 100190,China)

机构地区：[1]中国科学院自动化研究所模式识别国家重点实验室,北京100190 [2]中国科学院大学人工智能学院,北京100049 [3]中国科学院脑科学与智能技术卓越中心,北京100190

出　　处：《智能系统学报》2023年第1期138-143,共6页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61831022,61901473)。

摘　　要：脚本事件预测需要考虑两类信息来源:事件间的关联与事件内的交互。针对于事件间的关联,采用门控图神经网络对其进行建模。而对于事件内的交互,采用四元数对事件进行表征,接着通过四元数的哈密顿乘积来捕捉事件4个组成部分之间的交互。提出结合四元数和门控图神经网络来学习事件表示,它既考虑了外部事件图的交互作用,又考虑了事件内部的依赖关系。得到事件表示后,利用注意机制学习上下文事件表示和每个候选上下文表示的相对权值。然后通过权重计算上下文事件表示的和,再计算其与候选事件表示的欧氏距离。最后选择距离最小的候选事件作为正确的候选事件。在纽约时报语库上进行了实验,结果表明,通过多项选择叙事完形填空评价,本文的模型优于现有的基线模型。Two types of information sources are essential for script event prediction:the correlation between events and the inner interactions within one event.For the first information source,we use a gated graph neural network to model the correlation between events.For the inner interactions within one event,we use quaternion to model the event,and then we use the Hamilton product of quaternion to capture the inner interactions of four components.We propose to learn event representation by combining quaternion and a gated graph neural network.This approach considers the interaction of external event diagrams and the dependence within an event.After obtaining the event representation,we use an attention mechanism to learn the context event representation and the relative weight of each candidate context representation.Next,we calculate the sum of the context event embeddings through the weights,and then we calculate the Euclidean distance between the context event embedding sum and the candidate event embedding.Finally,we choose the candidate event with the smallest distance as the right candidate event.The results of experiments conducted on the New York Times corpus show that our proposed model is superior to the existing state-of-the-art baseline models through evaluation using a multiple-choice narrative cloze test.

关 键 词：四元数 门控图神经网络 事件表示 脚本事件预测 注意力机制 事理图谱 图神经网络 事件交互 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]