基于图的关系推理算法研究与实现

发布时间：2018-06-04 08:14

本文选题：知识图谱 + 关系推理　；参考：《电子科技大学》2017年硕士论文

【摘要】：伴随着信息技术及互联网的快速发展,传统基于文本的的信息检索方法已经难以满足用户的需求,迫切需要一种更加高效的信息处理方法。知识图谱研究就是在此背景下产生和发展起来的,知识图谱对非结构的文本信息进行处理,抽取获得包含其中的结构化知识,最终构建成为一张巨大的知识网络,相关的应用基于这一网络为用户提供更加快速和准确的服务。关系推理技术是组成整个知识图谱系统的核心模块之一,其主要任务是:在不添加额外知识的情况下,使用知识图谱中已经存在的知识,通过加工推理获得新的知识。该技术对于知识图谱的扩充和完善,以及基于知识图谱的各类实际应用有着深刻的促进意义。本文通过对国内外关系推理领域的相关工作进行全面调研,并在此基础上提出了一种新的基于图的关系推理算法(Graph-based Relation Inferring Algorithm,GBRI),该算法主要包括两个部分:全局关系推理模块和局部关系推理模块。两个模块分别利用不同关系类型间的连通结构以及单一关系内的结构特征进行推理预测,并将两者进行融合获得最终的关系推理结果。本文的主要贡献如下:(1)提出了一种新的关系预处理方法,该方法将知识图谱中的同义关系进行合并,降低了关系特征空间的维度,使得模型学习过程中的训练数据更为充足,缓解了数据的稀疏性问题。(2)提出了全局关系推理算法,该算法首次将知识图谱以一张无向图进行建模,构造“实体-关系”图,并针对知识图谱中包含的每一种关系,在“实体-关系”图中寻找其特征路径,用于构造特征向量,最后,利用Logistic回归训练得到每类关系的预测模型。实验表明,无向图的建模方式能够有效增加所能获取到的特征路径数量,提高算法的预测准确率。(3)提出了局部关系推理算法,该算法首次考虑了不同关系类型的内部特征,在“实体-关系”图中提取每种关系的关系子图,并计算子图中不同实体间的转移概率,进而得到局部关系推理模型的预测结果。实验表明,该算法能够有效提高多对多关系数据上的预测性能。(4)本文对全局关系推理算法和局部关系推理算法的预测结果进行融合,得到了最终的GBRI。并在三个开源数据集WN18,FB15k以及FB40k上,对比了GBRI模型与其它代表性工作的预测性能,证明了GBRI模型的有效性。
[Abstract]:With the rapid development of information technology and the Internet, the traditional text-based information retrieval method has been difficult to meet the needs of users, and a more efficient information processing method is urgently needed. The research of knowledge map is produced and developed under this background. Knowledge map processes the unstructured text information, extracts the structured knowledge contained therein, and finally constructs a huge knowledge network. Related applications provide users with faster and more accurate services based on this network. Relational reasoning technology is one of the core modules of the whole knowledge map system. Its main task is to obtain new knowledge by processing reasoning without adding additional knowledge and using the existing knowledge in the knowledge map. This technique is of great significance for the expansion and improvement of knowledge atlas and the practical applications based on knowledge atlas. This paper makes a comprehensive investigation on the related work in the field of relational reasoning at home and abroad. On this basis, a new graph-based Relation Inferring algorithm is proposed. The algorithm consists of two parts: the global relational reasoning module and the local relational reasoning module. The two modules respectively use the connected structure between different relationship types and the structural features in a single relationship to predict the reasoning and obtain the final relational reasoning result by merging the two models. The main contributions of this paper are as follows: (1) A new relationship preprocessing method is proposed, which combines the synonyms in the knowledge map, reduces the dimension of the relationship feature space, and makes the training data in the process of model learning more adequate. A global relational reasoning algorithm is proposed, in which the knowledge map is modeled with an undirected graph for the first time, and a "entity-relation" graph is constructed. The feature path is found in the "entity-relation" graph, which is used to construct the feature vector. Finally, the prediction model of each kind of relation is obtained by using Logistic regression training. Experiments show that the modeling method of undirected graph can effectively increase the number of feature paths that can be obtained, and improve the prediction accuracy of the algorithm. A local relational reasoning algorithm is proposed, which for the first time takes into account the internal features of different relationship types. The relation subgraph of each relation is extracted from the "entity-relation" graph, and the transition probability between different entities in the subgraph is calculated, and the prediction results of the local relational reasoning model are obtained. Experimental results show that the algorithm can effectively improve the prediction performance of many-to-many relational data.) in this paper, the prediction results of global relational reasoning algorithm and local relational reasoning algorithm are fused, and the final GBRI is obtained. On three open source datasets WN18FB15k and FB40k, the prediction performance of GBRI model and other representative work is compared, and the validity of GBRI model is proved.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP181

【相似文献】