Complex Logical Reasoning over Knowledge Graphs

2 years ago

Complex Logical Reasoning over Knowledge Graphs

Bo Xiong - ESR 5, University of Stuttgart

Reasoning is a long sought-after goal of Artificial Intelligence. In this presentation, we focus on logical reasoning over Knowledge Graphs (KGs), which aims at answering complex logical queries in an end-to-end manner. This is a very challenging because the real world KGs tend to be large, incomplete and contain noisy information. Traditional methods like subgraph matching cannot deal with incomplete KGs and the matching process is very slow, especially for large-scale KGs. Embedding based methods can solve these issues and have better robustness and efficiency. We review three papers that solve these issues, by a continual embedding of queries in the vector space, including Query2Box, BetaE and ConeE. Query2Box models queries as boxes but can only deal with conjunctive queries. To accept full First-Order Logical (FOL) queries, which includes negative queries, BetaE and ConeE model queries as beta distributions and convex cones, respectively.

The slides of the presentation can be found here.