基于BERT的三维地质建模约束信息抽取方法及意义  

BERT-based Method and Significance of Constraint Information Extraction for 3D Geological Modelling

作　　者：邱芹军 马凯 朱恒华[5] 刘春华[5] 谢忠[1,2] 谭永健 陶留锋 QIU Qinjun;MA Kai;ZHU Henghua;LIU Chunhua;XIE Zhong;TAN Yongjian;TAO Liufeng(School of Computer Sciences,China University of Geosciences,Wuhan 430074,Hubei,China;Hubei Key Laboratory of Intelligent Geo-Information Processing,Wuhan 430074,Hubei,China;Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang 443002,Hubei,China;College of Computer and Information Technology,China Three Gorges University,Yichang 443002,Hubei,China;Shandong Institute of Geological Survey,Jinan 250000,Shandong,China)

机构地区：[1]中国地质大学(武汉)计算机学院,湖北武汉430074 [2]智能地学信息处理湖北省重点实验室,湖北武汉430074 [3]湖北省水电工程智能视觉监测重点实验室,湖北宜昌443002 [4]三峡大学计算机与信息学院,湖北宜昌443002 [5]山东省地质调查院,山东济南250000

出　　处：《西北地质》2022年第4期124-132,共9页Northwestern Geology

基　　金：国家自然科学基金项目“地球科学知识图谱表示模式与群智协同构建”(42050101)、“基于多模态数据理解及融合的三维地质模型构建方法研究”(41871311);济南城区四维地质环境可视化信息系统平台建设项目(2018GDCG01Z0301);山东省重点研发计划(重大科技创新工程)项目“数字孪生城市四维可视化信息系统及其在济南城区的应用”(2019JZZY020105);中国博士后科学基金(2021M702991)联合资助。

摘　　要：地质报告中地质体的几何、拓扑及属性信息是三维地质建模过程中重要约束性信息。但传统的属性信息抽取方法存在覆盖率有限、局限于人工设计特征及模型泛化能力差等问题。面向三维建模任务,总结了地质报告中地质体的几何、拓扑及属性文本的特点,提出了一种基于BERT-BiLSTM-CRF的三维地质建模信息抽取方法;基于BERT预训练模型,构建融合BiLSTM和CRF的深度学习模型,通过BERT模型获取动态字符深层次语义信息,弥补静态词向量无法解决一词多义的问题,提高地质体复杂建模信息的抽取能力。以43篇地质报告为数据源进行模型性能评估,实验结果表明所提出的方法对于地质体三类属性信息抽取准确率达到90%以上,对于三维地质建模具有重要支撑作用。The geometry,topology and attribute information of geological bodies in geological reports are important constraint information in the 3 D geological modeling process.However,the traditional attribute information extraction methods have problems such as limited coverage,limited to artificial design features and poor model generalization ability.Facing the 3 D modeling task,the geometry,topology and attribute text characteristics of geological bodies in geological reports are summarized,and a 3 D geological modeling information extraction method based on BERT-BiLSTM-CRF is proposed;based on the BERT pre-training model,a deep learning model integrating BiLSTM and CRF is constructed to obtain deep semantic information of dynamic characters through the BERT model to make up for the static word vector cannot solve the problem of multiple meanings of a word,and improve the extraction ability of complex modeling information of geological bodies.The model performance is evaluated with 43 geological reports as the data source,and the experimental results show that the proposed method has an accuracy rate of over 90%for extracting three types of attribute information of geological bodies,which is an important support for 3 D geological modeling.

关 键 词：三维地质建模 属性抽取 BERT 约束信息 

分 类 号：P628.3[天文地球—地质矿产勘探]