基于自适应知识选择的机器阅读理解  

Adaptive Knowledge Selection for Machine Reading Comprehension

作　　者：李泽政 田志兴 张元哲[1] 刘康[1,2] 赵军 LI Zezheng;TIAN Zhixing;ZHANG Yuanzhe;LIU Kang;ZHAO Jun(State Key Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Science,Beijing 100190,China;School of Artificial Intelligence,University of Chinese Academy of Sciences,Beijing 100049,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技术学院,北京100049

出　　处：《中文信息学报》2022年第6期117-124,共8页Journal of Chinese Information Processing

基　　金：国家重点研发计划项目(2018YFB1005100)。

摘　　要：目前针对知识增强机器阅读理解的研究主要集中在如何把外部知识融入现有的机器阅读理解模型,却忽略了对外部知识的来源进行选择。该文首先基于注意力机制对外部知识进行编码,然后对不同来源的外部知识编码进行打分,最后自适应地选择出对回答问题最有帮助的知识。与基线模型相比,该文提出的基于自适应知识选择的机器阅读理解模型在准确率上提高了1.2个百分点。The current knowledge-enhanced machine reading comprehension is focused on how to integrate external knowledge into the existing MRC model,while ignores the selection for the source of external knowledge.This article first uses the attention mechanism to encode external knowledge,then scores external knowledge from different sources,and finally selects the most helpful knowledge with respect to different questions.Compared with the baseline models,our method improves the accuracy by 1.2 percent.

关 键 词：机器阅读理解 知识增强 自适应选择 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]