Temporal Knowledge Graph Completion with Approximated Gaussian Process Embedding

Proceedings of the 29th International Conference on Computational Linguistics, pages 4697–4706 October 12–17, 2022. 4697 Temporal Knowledge Graph Completion with Approximated Gaussian Process Embedding Linhai Zhang1 Deyu Zhou1∗ 1 School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China {lzhang472, d.zhou}@seu.edu.cn Abstract Knowledge Graphs (KGs) stores world knowl- edge that benefits various reasoning-based ap- plications. Due to their incompleteness, a fun- damental task for KGs, which is known as Knowledge Graph Completion (KGC), is to perform link prediction and infer new facts based on the known facts. Recently, link pre- diction on the temporal KGs becomes an active research topic. Numerous Temporal Knowl- edge Graph Completion (TKGC) methods have been proposed by mapping the entities and rela- tions in TKG to the high-dimensional represen- tations. However, most existing TKGC meth- ods are mainly based on deterministic vector embeddings, which are not flexible and expres- sive enough. In this paper, we propose a novel TKGC method, TKGC-AGP, by mapping the entities and relations in TKG to the approx- imations of multivariate Gaussian processes (MGPs). Equipped with the flexibility and ca- pacity of MGP, the global trends as well as the local fluctuations in the TKGs can be simul- taneously modeled. Moreover, the temporal uncertainties can be also captured with the ker- nel function and the covariance matrix of MGP. Moreover, a first-order Markov assumption- based training algorithm is proposed to effec- tive optimize the proposed method. Experimen- tal results show the effectiveness of the pro- posed approach on two real-world benchmark datasets compared with some state-of-the-art TKGC methods. 1 Introduction Knowledge Graphs (KGs) provide an efficient way to store world knowledge. Various KGs such as DBpedia (Auer et al., 2007), NELL (Carlson et al., 2010), YAGO (Suchanek et al., 2007) and Freebase (Bollacker et al., 2008) have been constructed and benefited downstream applications such as informa- tion retrieval, question answering, etc (Hao et al., 2017; Zhang et al., 2021). Generally, a fact in KG ∗Corresponding author. (2008-11-04, Obama, president, USA) Time Line 𝑎���𝑎���𝑚���𝑎���𝑚���𝑚𝑖𝑚���𝑖𝑚���𝑚���𝑖𝑎��� −𝑎���𝑎𝑚���𝑎���𝑎��� 𝑎���𝑎���𝑚���𝑚���𝑖𝑚���𝑦 −𝑎���𝑎𝑚���𝑎���𝑎��� 𝑚���𝑚���𝑚���𝑎���ℎ𝑎𝑚���𝑚���𝑖𝑎��� 𝑚���𝑚���𝑚���𝑎���𝑎���𝑚���𝑚��� −𝑎���𝑎𝑚���𝑎���𝑎��� (2008-11-05, Obama, criticize, Iraq) (2012-11-03, Obama, demand, Japan) (2008-11-02, Obama, praise, Mexico) (2008-11-01, Obama, reject, South Korea) Figure 1: Illustration of three types of embedding-based TKGC methods which share the same embedding posi- tion function. Blue points denote the ideal position for the embeddings of entity Obama, red lines denote the embedding evolution function estimated by three types of methods, and shade areas denote the uncertainty es- timated by density-based and stochastic process-based methods. can be represented as a triple (s, p, o), where s (sub- ject) denotes head entity node, o (object) denotes tail entity node and p (predicate) denotes relation edge between them. However, in the real world, some facts are time-aware. For example, the fact (Joe Biden, presidentOf, the United States) is not valid until January 21, 2021, the United States presi- dential inauguration of Joe Biden. Therefore, some KGs store time-aware facts or events as the quadru- ple …

Content truncated to 500 words