【作者】 王斌；

【导师】 吴亮；

【作者基本信息】 中国地质大学 ， 地图制图学与地理信息工程， 2022， 硕士

【摘要】 在数字化和智能化快速发展的时代,长期积累的海量基础地质成果资料促进了地质信息化和知识化的发展。然而,面对海量的地质资料,特别是地质报告,如何从地质认知的角度从地质报告中抽取地质图件及文本信息,并以结构化的形式进行存储和表达。这些问题是当前地质信息抽取任务面临的挑战。因此,本文以地质报告这一主要地质成果记录载体作为数据源,建立面向地质报告的地质图文语义表征模型,并提出了面向地质剖面图的半自动矢量化构建方法及顾及语句上下文信息的地质实体和关系联合抽取方法,为地质知识的演化与推理分析提供了基础理论与方法。本文的主要研究工作如下:(1)提出了基于地质报告的地质剖面图图件和地质文本信息结构化表征模型。针对地质剖面图图件,提出构建矢量地质剖面图岩层对象化的概念,使其每个岩层具备几何特征和属性信息;针对地质报告地质文本信息,通过地质三元组形式进行抽取,并利用图结构形式存储和表达;通过地质内容词的概率统计、词云展示和内容词语的向量表征和聚类结果,在此基础上构建标注标准,实现了以24种地质关系为特征的地质报告信息抽取数据集的构建。该结构化表征模型设计为地质剖面图的构建与地质文本信息抽取工作提供了基础框架。(2)提出了融合多源地质数据的半自动化地质剖面图构建方法。本文利用区域地质调查研究报告和数字高程模型作为数据源,建立了“剖面线-边界线-对象化”的半监督式构建流程,实现了构建矢量的对象化地质剖面图的目的,方便实时编辑和内容更新;并且,该方法为在缺乏钻孔等野外测量地质数据情况下,融合多源数据构建地质剖面图提供了一种新的思路。(3)提出了顾及语句上下文信息的地质文本三元组抽取方法。本文在考虑地质文本的语料特征的基础上,针对地质报告文本存在复杂关系、重叠关系下难以完整提取地质三元组的问题,通过基于Bi LSTM的轴向注意力模块和多尺度图卷积单元对地质实体和关系进行表征,提出顾及上下文信息的地质实体和关系联合抽取方法。实验结果表明本文提出的方法F1-Score达到了90.05%,具有良好的性能。最后,在地质报告图文信息抽取关键技术的支持下,本文实现了地质报告图文信息抽取及结构化表征的原型系统,为地质资料信息抽取工作提供了一种新的管理途径和数据分析工具。更多还原

【Abstract】 In the rapid development of digitization and intelligence,the long-term accumulation of massive basic geological achievements promotes the development of geological informatization and knowledge.However,in the face of massive geological data,especially geological reports,how to extract geological maps and text information from geological reports from the perspective of geological cognition,and store and express it in a structured form is the challenge faced by current geological information extraction.Therefore,in terms of the geological report,the main achievements of geological record carrier,as a data source,to establish a geological graphic and text representation model.Also,a semi-automatic vectorized construction method for geological profiles and a joint extraction method of geological entities and relations considering sentence context information are proposed,providing the analysis of the geological evolution of knowledge and reasoning with the basic theory and method.The main research work of this thesis is as follows:(1)A structured representation model for geological reports based on geological profiles and textual information is proposed.Aiming at the geological profiles,the concept of constructing a vector geological profile is proposed to objectify the rock layers.Hence,each rock layer has geometric features and attributes information.For the geological text information of the geological report,extract it in the form of geological triples,and store and express it in the form of graph structure;through the probability statistics of geological content words,word cloud display and vector representation,and clustering results of content words,a labeling standard is constructed,and realized the construction of a geological report information extraction data set featuring 24 geological relations.The structural representation model provides a basic framework for the construction of geological profiles and the extraction of geological text information.(2)A semi-automatic geological profile construction method integrating multi-source geological data is proposed.In this thesis,using regional geological survey report and digital elevation model as data sources,a semi-supervised construction process of "profile line-borderline-objectification" is established to realize the purpose of constructing vector objectified geological profiles,which is convenient for real-time editing and content updating.Moreover,this method provides a new idea for the construction of geological profiles by integrating multi-source data in the absence of field measurement geological data such as boreholes.(3)A triplet extraction method of geological text considering sentence context information is proposed.In this thesis,on the basis of considering the geological text corpus characteristics,aiming at the problem that it is difficult to completely extract geological triples under the complex relations and overlapping relations in the geological report text,the Bi LSTM-based axial attention module,and multi-scale graph convolution unit are used to characterize geological entities and relations,and a contextual information-based joint extraction approach is proposed.Experimental results show that the proposed method F1-Score achieves 90.05% and has good performance.Finally,with the support of the key technology of graphics and text information extraction from geological reports,this thesis implements a prototype system for graphic and textual information extraction and structured characterization of geological reports,which provides a new management approach and data analysis tool for geological data information extraction.更多还原