基于多示例学习的化学物致病关系抽取  

Chemical-Induced Disease Relation Extraction Based on Multiple Instance Learning

作　　者：冯靖焜 杨志豪[1] 罗凌 林鸿飞[1] 王健[1] FENG Jingkun;YANG Zhihao;LUO Ling;LIN Hongfei;WANG Jian(School of Computer Science and Technology,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学计算机科学与技术学院,辽宁大连116024

出　　处：《山西大学学报（自然科学版）》2019年第1期59-68,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：十三五国家重点研发计划(2016YFC0901902);国家自然科学基金(61272373;61340020;61572102)

摘　　要：化学物(药物)致病关系在生物医学领域广受关注。文章提出一种从生物医学文献中抽取化学物致病关系的方法 MICNN,首先基于多示例学习的思想,将化学物疾病实体对与多个示例对齐,然后训练一个文档级别的卷积神经网络,使用多示例同时抽取位于同一文档中句间和句内级别的化学物致病关系,有效减弱了现有方法中因单示例存在噪音而对抽取性能造成的损害。实验结果表明,MICNN仅使用了词向量特征、位置特征和实体特征三种基本特征,就在BioCreative V CDR任务提供的测试集上取得很好的性能表现。The chemical-induced disease relation has attracted much attention in the biomedical field.This paper proposes a method(MICNN)for extracting chemical-induced disease relation from biomedical literature.Based on the multiple instance learning,MICNN first aligned chemical-disease entity pairs with multiple instance,then trained a document-level convolutional neural network,and used multiple instance to simultaneously extract chemical-induced disease relation at both sentence and document level from the same document.The method is able to reduce the damage to the extraction performance caused by the noise of the single instance in existing system.The experimental results show that using only three basic features(i.e.,word vector,position and entity label),MICNN achieves excellent performance on the test set provided by the BioCreative V CDR task.

关 键 词：神经网络 多示例学习 关系抽取 化学物致病关系 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]