Neural Axiom Network for Knowledge Graph Reasoning

Tracking #: 2852-4066

This paper is currently under review
Juan Li
Wen Zhang
Xiangnan Chen
Jiaoyan Chen
Huajun Chen

Responsible editor: 
Freddy Lecue

Submission type: 
Full Paper
Knowledge graph reasoning is essential for improving the quality of knowledge graphs due to automatic mechanisms involved in KG construction which probably introduces incompleteness and incorrectness. In recent years, various KG reasoning techniques such as symbolic- and embedding-based methods, have been proposed for inferring missing triples and detecting noises. Symbolic-based reasoning methods concentrate on inferring new knowledge according to predefined rules or ontologies, where rules and axioms have been proved to be effective but are difficult to obtain. Meanwhile, embedding-based reasoning methods learn low-dimensional representations of entities and relations primarily by utilizing structural information, and the learned embeddings achieve promissing results in downstream tasks such as knowledge graph completion. These methods, however, ignore implicit axiom information which are not predefined in KGs but can be reflected through data. To be specific, each correct triple is considered to satisfy all axioms, as it is also a consistent triple. In this paper, we explore how to combine explicit structural and implicit axiom information to improve reasoning ability. Specifically, we present a novel NeuRal Axiom Network framework (NeuRAN) that only uses existing triples in KGs to address issues in the above methods. The framework consists of a knowledge graph embedding module that preserves the semantics of a triple, and five axiom modules that are encoded based on the characteristics of five kinds of axioms corresponding to five typical object property expression axioms defined in OWL2, including ObjectPropertyDomain, ObjectPropertyRange, DisjointObjectProperties, IrreflexiveObjectProperty and AsymmetricObjectProperty. The knowledge graph embedding module and axiom modules respectively calculate the probabilities that the triple conforms to the semantics and the corresponding axioms. Evaluations on KG reasoning tasks including noise detection, triple classification and link prediction show the efficiency of our method.
Full PDF Version: 
Under Review