基于标签增强的机器阅读理解模型  

Label-Enhanced Reading Comprehension Model

作　　者：苏立新[1,2] 郭嘉丰 范意兴[1] 兰艳艳 程学旗 SU Lixin;GUO Jiafeng;FAN Yixing;LAN Yanyan;CHENG Xueqi(Key Laboratory of Network Data Science and Technology,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;School of Computer and Control Engineering,University of Chinese Academy of Sciences,Beijing 100190;Institute of Network Technology,Institute of Computing Technology(YANTAI),Chinese Academy of Sciences,Yantai 264005)

机构地区：[1]中国科学院计算技术研究所网络数据科学与技术重点实验室,北京100190 [2]中国科学院大学计算机与控制学院,北京100190 [3]中国科学院计算技术研究所烟台分所烟台中科网络技术研究所,烟台264005

出　　处：《模式识别与人工智能》2020年第2期106-112,共7页Pattern Recognition and Artificial Intelligence

基　　金：国家重点研发计划(No.2016QY02D0405);国家自然科学基金项目(No.61425016,61472401,61722211,61872338,61902381);中国科学院青年创新促进会项目(No.20144310,2016102);重庆市基础科学与前沿技术研究专项项目(重点)(No.cstc2017jcjyBX0059);泰山学者工程专项经费(No.ts201511082)资助。

摘　　要：抽取式问答中已有模型仅建模答案的边界,忽视人的潜在标注过程,导致模型仅学习到表面特征,影响泛化能力.因此,文中提出基于标签增强的机器阅读理解模型(LE-Reader),模拟人的标注过程.LE-Reader模型同时建模答案所在句子、答案内容和答案边界.根据用户标注的答案边界推断正确答案的句子和答案内容作为标签,监督模型的学习过程.通过多任务学习的方式融合3个损失函数.预测时融合3种建模结果,确定最终答案,提高模型的泛化性能.在SQuAD数据集上的实验验证LE-Reader的有效性.In the existing extractive reading comprehension models,only the boundary of answers is utilized as the supervision signal and the labeling processed by human is ignored.Consequently,learned models are prone to learn the superficial features and the generalization performance is degraded.In this paper,a label-enhanced reading comprehension model is proposed to imitate human activity.The answer-bearing sentence,the content and the boundary of the answer are learned simultaneously.The answer-bearing sentence and the content of the answer can be derived from the boundary of the answer and these three types of labels are regarded as supervision signals.The model is trained by multitask learning.During prediction,the probabilities from three predictions are merged to determine the answer,and thus the generalization performance is improved.Experiments on SQuAD dataset demonstrate the effectiveness of LE-Reader model.

关 键 词：阅读理解 多任务学习 答案抽取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]