UnTiCk: Unsupervised Type-Aware Complex Logical Queries Reasoning over Knowledge Graphs

Abstract

For the task of answering complex logical queries on large-scale incomplete knowledge graphs, the promising approach is to embed the knowledge graph and complex logical queries into a low-dimensional space and perform iterative reasoning to find the final answers. The general problem is that these models do not include entity types as an important feature, which reduces the reasoning potential. However, explicit type information is not always available on large-scale knowledge graphs. We innovatively propose an embedding-based framework for Unsupervised Type-Aware Complex Logical Queries (UnTiCk). Our approach implements unsupervised type constraints on multi-hop complex logical query processing. Moreover, it can capture the different representations of type features when entities are in different locations in the logical path. We designed type compatibility measurement meta-operators combined with popular Existential Positive First-Order (EPFO) neural logical operators to achieve type-aware EPFO complex query embedding. We validated the effectiveness of our framework on popular large-scale knowledge graphs by using the same embedding dimensionality as complex logical embedding methods. The results showed an average relative improvement of 1.9–12.8% on Hit@3 and up to 42.1% on the certain logical pattern.