Perform like an Engine: A Closed-Loop Neural-Symbolic Learning Framework for Knowledge Graph Inference

Proceedings of the 29th International Conference on Computational Linguistics, pages 1391–1400 October 12–17, 2022. 1391 Perform Like an Engine: A Closed-Loop Neural-Symbolic Learning Framework for Knowledge Graph Inference Guanglin Niu1, Bo Li1,2∗, Yongfei Zhang3,4, Shiliang Pu5 1 Institute of Artificial Intelligence, Beihang University, Beijing, China 2 Hangzhou Innovation Institute, Beihang University, Hangzhou, China 3 Beijing Key Laboratory of Digital Media, Beihang University, Beijing, China 4 State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China 5 Hikvision Research Institute, Hangzhou, China {beihangngl, boli, yfzhang}@buaa.edu.cn, pushiliang.hri@hikvision.com Abstract Knowledge graph (KG) inference aims to ad- dress the natural incompleteness of KGs, in- cluding rule learning-based and KG embedding (KGE) models. However, the rule learning- based models suffer from low efficiency and generalization while KGE models lack inter- pretability. To address these challenges, we pro- pose a novel and effective closed-loop neural- symbolic learning framework EngineKG via incorporating our developed KGE and rule learning modules. KGE module exploits sym- bolic rules and paths to enhance the semantic association between entities and relations for improving KG embeddings and interpretabil- ity. A novel rule pruning mechanism is pro- posed in the rule learning module by leverag- ing paths as initial candidate rules and employ- ing KG embeddings together with concepts for extracting more high-quality rules. Experi- mental results on four real-world datasets show that our model outperforms the relevant base- lines on link prediction tasks, demonstrating the superiority of our KG inference model in a neural-symbolic learning fashion. The source code and datasets of this paper are available at https://github.com/ngl567/EngineKG. 1 Introduction Typical knowledge graphs (KGs) store triple facts and some of them also contain concepts of enti- ties (Bollacker et al., 2008). The KGs have proven to be incredibly effective for a variety of applica- tions such as dialogue system (Zhou et al., 2018) and question answering (Huang et al., 2019). How- ever, the existing KGs are always incomplete which restricts the performance of knowledge-based ap- plications. Thus, KG inference plays a vital role in completing KGs for better applications of KGs. The existing KG inference approaches are usu- ally classified into two main categories: (1) Rule learning-based models such as AMIE+ (Galárraga ∗Corresponding author. (2) Compression Stroke: KG Embedding Module triples paths concepts (1) Intake Stroke: Inject the matched rules into Compression Stroke rules (3) Expansion Stroke: Rule Learning Module (4) Exhaust Stroke: Update rule set newly learned rules updated rule set Link Prediction KG embeddings Figure 1: The brief architecture of our closed-loop framework for KG inference EngineKG that performs like a four-stroke engine. et al., 2015) and AnyBurl (Meilicke et al., 2019) mine rules from KGs and employ these rules to predict new triples by deduction. However, rule learning-based models suffer from low efficiency of the rule mining process and the poor general- ization caused by the limited coverage of inference patterns. (2) KGE technique learns the embed- dings of entities and relations to predict the missing triples via scoring each triple candidate, including TransE (Bordes et al., …

Content truncated to 500 words