基于BiLSTM-CRF的中文地质时间信息抽取  

Geological Time Information Extraction from Chinese Text Based on BiLSTM-CRF

作　　者：刘文聪 张春菊[1,3] 汪陈 张雪英 朱月琴[4] 焦守涛[4] 鲁艳旭 LIU Wencong;ZHANG Chunju;WANG Chen;ZHANG Xueying;ZHU Yueqin;JIAO Shoutao;LU Yanxu(School of Civil Engineering,Hefei University of Technology Mefei 230009.China;MOE Key Laboratory of Virtual Geographical Environment,Nanjing Normal University,Nanjing 210023,China;Key Laboratory of Urban Land Resources Monitoring and Simulation.Shenzhen Guangdong 518034,China;Development Research Center,China Geological Survey,Beijing 100037,China)

机构地区：[1]合肥工业大学土木与水利工程学院,安徽合肥230009 [2]南京师范大学虚拟地理环境教育部重点实验室,江苏南京210023 [3]自然资源部城市国土资源监测与仿真重点实验室,广东深圳518034 [4]中国地质调查局发展研究中心,北京100037

出　　处：《地球科学进展》2021年第2期211-220,共10页Advances in Earth Science

基　　金：自然资源部城市国土资源监测与仿真重点实验室开放基金“面向国土资源信息感知的定性位置空间语义计算”(编号:KF-2020-05-084);国家自然科学基金项目“‘文本—地图’结合的地理知识图谱构建方法”(编号:41971337)资助。

摘　　要：时间信息贯穿于地质现象和事件产生、发展、消亡的整个过程中,反映了地质现象和事件的状态和演变过程。特别是,地质时间表达通常与成矿内在机制和时空演化规律有关。设计并实现了基于深度学习的通用时间和地质时间信息抽取方法。结合地质矿产文本中时间信息的描述特点,将时间信息划分为通用时间信息与地质时间信息两种类型,并对两种时间信息类型进行细分;基于自主研发的"交互式矿产信息标注软件",采用交叉验证及意见反馈模式构建了地质时间信息语料库;实现了基于双向长短期记忆神经网络-条件随机场(BiLSTM-CRF)的时间信息抽取方法;并与主流的卷积神经网络(CNN)和条件随机场(CRF)模型的抽取结果进行了比较。实验结果表明,基于双向长短期记忆神经网络-条件随机场的时间信息抽取效果最好,对总体时间抽取的F1值达到95.49%,较好地解决了地质文本中时间信息的规范化表达和结构化抽取问题。Time information runs through the entire process of the creation,development and extinction of geological entities,reflecting the state and evolution of geological entities.In particular,the expression of geological time is usually related to metallogenetic mechanism and space time evolution regularity.This paper designs and implements a universal time and geological time information extraction method based on deep learning methods.Combining the description characteristics of time information in the Chinese text of geological and mineral resources,the time information in geological reports and documentation is divided into two types:universal time information and geological time information,and the two types of time information are subdivided.The self-developed geological time information corpus is constructed using cross-validation and opinion feedback mode.The time information extraction method based on BiLSTM-CRF is realized,and this method is compared with CNN and CRF.The experimental results show that the BiLSTM-CRF model is better than the mainstream model in time information extraction,and the F1-Measure of the overall time extraction reaches 95.49%,which solves the problem of standardized expression and structured extraction of time information in geological text.

关 键 词：地质时间 时间抽取 时间信息语料库 条件随机场 双向长短期记忆神经网络 

分 类 号：P628[天文地球—地质矿产勘探]