【作者】 于波；

【导师】 沈继红；

【作者基本信息】 哈尔滨工程大学 ， 数学， 2021， 硕士

【摘要】 电磁频谱战条令其内容概念丰富,包含综述,关系,作战的组织、规划、实施等,反映了外军对于作战的考虑,所以本文以电磁频谱战条令为基础,采用知识图谱来解决文本发现问题。知识图谱的构建是一个系统性的问题,针对非结构化的文本数据,采用自上而下的方法,结合本体建模、自然语言处理、知识表示学习等技术构建军事条令知识图谱。本论文创新性的研究工作如下:针对电磁频谱战条令的内容结构特点,本文整合本体构建、知识抽取、图谱补全技术,提出了一个电磁频谱战条令知识图谱构建框架,并以这个框架为基础进行了知识图谱构建,展示了本框架的实用性。由于需要对文本数据进行规范化处理,本文根据有关条令中的描述和定义,结合领域背景,提炼出文本描述特点和实体关系特征,对电磁频谱条令定义4个类和13种关系,构建了电磁频谱战本体模型。利用分词、词性标注和依存句法分析模型对数据进行预处理,采用规则对实体和关系进行抽取。传统的链路预测模型通过最小化损失函数来构建实体和关系向量,尽管这些方法在简单关系上取得了进步,但存在度量函数简单,间隔参数不具有适用性等问题,在建模复杂关系方面缺乏优势。为解决以上问题提出自适应嵌入模型,通过抑制噪声维度和改进边缘参数实现模型的自适用性。本文在电磁频谱条令中对提出的模型进行了验证,并取得了良好的效果。为进一步验证模型有效性,本文在公开数据集WN18和FB15K上进行了实验。结果显示,相对于传统转换模型,本模型将数据集WN18的平均排名降低120以上,正确实体比例提升2%以上;将数据集FB15K的平均排名降低30以上,正确实体比例中提升6%以上。在公开数据集上的结果也表明,本模型因为具有自适应特性,能更好的适应复杂环境。更多还原

【Abstract】 Electromagnetic spectrum doctrine is rich in content and concept,including overview,relationship,organization,planning and implementation of operations,which reflects the foreign military’s consideration of operations.Therefore,based on electromagnetic spectrum doctrine,this paper uses knowledge graph to solve the problem of text discovery.The construction of knowledge graph is a systematic problem.Aiming at the unstructured text data,this paper uses the top-down method,combined with ontology modeling,natural language processing,knowledge representation learning and other technologies to construct the military doctrine knowledge graph.The innovative research work of this paper is as followsAccording to the content structure characteristics of electromagnetic spectrum warfare doctrine,this paper proposes a framework of electromagnetic spectrum warfare doctrine knowledge graph construction by integrating ontology construction,knowledge extraction and map completion technology,and constructs the knowledge graph based on this framework,which shows the practicability of this framework.Due to the need for standardized processing of text data,this paper,according to the description and definition in relevant regulations,combined with the domain background,extracts the characteristics of text description and entity relationship,defines 4 classes and 13 relationships for electromagnetic spectrum regulations,and constructs the ontology model of electromagnetic spectrum warfare.The model of word segmentation,part of speech tagging and dependency parsing is used to preprocess the data,and rules are used to extract entities and relationships.The traditional link prediction model constructs entity and relation vectors by minimizing the loss function.Although these methods have made progress in simple relations,they have some problems,such as simple measurement function,inapplicability of interval parameters,and lack of advantages in modeling complex relations.To solve the above problems,this paper proposes a adaptive embedding model.Different from the traditional model given global parameters,this paper weights the loss measure by operator and linearly correlates the interval,and realizes the self applicability of the model by determining the parameters in the training process.In this paper,the proposed model is verified in the electromagnetic spectrum regulations,and good results are achieved.In order to further verify the effectiveness of the model,experiments are carried out on open datasets wn18 and fb15 k.The results show that compared with the traditional transformation model,this model reduces the average ranking of data set wn18 by more than 120,and increases the proportion of correct entities by more than 2%;reduces the average ranking of data set fb15 k by more than30,and increases the proportion of correct entities by more than 6%.The results on public data sets also show that the model can better adapt to complex environment because of its adaptive characteristics.更多还原