Knowledge Is Flat: A Seq2Seq Generative Framework for Various Knowledge Graph Completion

Proceedings of the 29th International Conference on Computational Linguistics, pages 4005–4017 October 12–17, 2022. 4005 Knowledge Is Flat: A Seq2Seq Generative Framework for Various Knowledge Graph Completion Chen Chen1,∗, Yufei Wang2,∗, Bing Li3 and Kwok-Yan Lam1† Nanyang Technological University, Singapore1 Macquarie University, Sydney, Australia2 Centre for Frontier AI Research (CFAR), A*STAR, Singapore3 {S190009,kwokyan.lam}@ntu.edu.sg, yufei.wang@students.mq.edu.au li_bing@ihpc.a-star.edu.sg Abstract Knowledge Graph Completion (KGC) has been recently extended to multiple knowledge graph (KG) structures, initiating new research direc- tions, e.g. static KGC, temporal KGC and few- shot KGC (Ji et al., 2022). Previous works often design KGC models closely coupled with specific graph structures, which inevitably re- sults in two drawbacks: 1) structure-specific KGC models are mutually incompatible; 2) ex- isting KGC methods are not adaptable to emerg- ing KGs. In this paper, we propose KG-S2S, a Seq2Seq generative framework that could tackle different verbalizable graph structures by unifying the representation of KG facts into “flat” text, regardless of their original form. To remedy the KG structure information loss from the “flat” text, we further improve the input rep- resentations of entities and relations, and the inference algorithm in KG-S2S. Experiments on five benchmarks show that KG-S2S outper- forms many competitive baselines, setting new state-of-the-art performance. Finally, we ana- lyze KG-S2S’s ability on the different relations and the Non-entity Generations 1. 1 Introduction Knowledge graph completion (KGC) has been a fundamental task to discover unobserved facts from various knowledge graph (KG) structures, includ- ing static KGC (SKGC), temporal KGC (TKGC) and few-shot KGC (FKGC) (Ji et al., 2022). As shown in Figure 1, TKGC (in orange) contains temporal facts with timestamps, while FKGC (in green) predicts the facts with relations that only have limited or zero training instances. Typically, the solutions for KGC are graph- based, i.e., treating entities and relations as nodes and linkages. The training and inference of SKGC ∗First two authors contribute equally. †Corresponding author 1Our source code is available at https://github. com/chenchens190009/KG-S2S Figure 1: Running examples of Static (SKGC), Tempo- ral (TKGC) and Few-shot (FKGC) Knowledge Graph Completion tasks. Our proposed KG-S2S is an unified Seq2Seq framework adaptable to all of these tasks. models rely on various transitional relations over graph paths (Trouillon et al., 2016; Dettmers et al., 2018; Vashishth et al., 2020). TKGC and FKGC methods are further integrated with non-trivial com- ponents or learning paradigms to handle the ex- tra temporal information or training requirements. Concretely, TKGC models (Dasgupta et al., 2018; Goel et al., 2020; Lacroix et al., 2020) either con- struct temporal-specific sub-KG or add additional temporal embeddings into existing SKGC meth- ods. FKGC models apply the additional training scheme (e.g., meta-learning) between the frequent relations and the infrequent ones to the SKGC mod- els (Xiong et al., 2018; Niu et al., 2021). Such a methodological discrepancy leads to a great main- tenance cost and being inadaptable to emerging knowledge queries, ingestion, and presents. Nat- urally, a research question has been raised: Can we adapt the different forms of KG facts and solve these KGC …

Content truncated to 500 words