基于卷积神经网络的时序知识图谱补全研究

A Study of Temporal Knowledge Graph Complementation Based on Convolutional Neural Network

知识图谱通常由三元组头尾实体及关系组成,通过将大量三元组组织成图的形式,知识图谱可以展示实体间的复杂关系网络,这些关系可以是直接或间接的关联。传统的知识图谱中实体和关系之间的连接通常是静态的,缺乏对动态关系的建模能力,而时序知识图谱则是在三元组的基础之上增加了时间戳这一维度,能够更好地捕捉实体和关系的时序依赖性,从而支持时序推理,使得知识图谱的表达更加贴近于真实世界。但是通常包含时间信息的数据集中存在大量缺失数据,导致包含时间戳在内的四元组完整性下降,知识图谱下游应用如问答系统、推荐系统等将会受到直接影响。因此,本文提出了基于静态知识图谱嵌入模型ConvE的面向时序知识图谱补全的张量分解模型Conv-ATG (Convolutional Attention Temporal Graph),具体来说,ConvE主要针对静态知识图嵌入,基于ConvE引入了对于时间信息的表示并增加自注意力机制提高模型性能。在2个主流大规模数据集ICEWS、WIKIDATA上进行实验,实验结果表明Conv-ATG模型的性能优于现有的大多数推理方法,说明了本模型的有效性。

Knowledge graphs are usually composed of ternary head and tail entities and relationships, and by organizing a large number of ternaries into the form of graphs, knowledge graphs can display a complex network of relationships between entities, which can be directly or indirectly related. The connection between entities and relationships in traditional knowledge graphs is usually static and lacks the ability to model dynamic relationships, while temporal knowledge graphs add the dimension of timestamps on top of ternary groups, which can better capture the temporal dependencies of entities and relationships, thus supporting temporal reasoning and making the expression of knowledge graphs closer to the real world. However, there is usually a large amount of missing data in the dataset containing temporal information, which leads to a decrease in the integrity of the quaternion including timestamps, and the downstream applications of knowledge graphs such as Q&A systems and recommender systems will be directly affected. Therefore, this paper proposes Conv-ATG (Convolutional Attention Temporal Graph), a tensor decomposition model for temporal knowledge graph complementation based on the static knowledge graph embedding model ConvE. Specifically, ConvE is mainly aimed at static knowledge graph embedding, and based on ConvE, it introduces a representation for temporal information and adds a self-attention mechanism to im-prove the model performance. Experiments are conducted on 2 mainstream large-scale datasets, ICEWS and WIKIDATA, and the experimental results show that the Conv-ATG model outperforms most of the existing inference methods, which illustrates the effectiveness of this model.