Semantic Reasoning with Differentiable Graph Transformations

TL;DR

This paper presents a differentiable semantic reasoner that uses graph transformations and embeddings to perform reasoning, allowing rules to be manually written or learned from data, bridging logic and neural methods.

Contribution

It introduces a novel differentiable framework for semantic reasoning using graph transformations aligned with Description Logic, enabling rule inference from facts and goals.

Findings

Effective rule inference from data

Combines logic-based and neural approaches

Supports manual and learned rule specification

Abstract

This paper introduces a differentiable semantic reasoner, where rules are presented as a relevant set of graph transformations. These rules can be written manually or inferred by a set of facts and goals presented as a training set. While the internal representation uses embeddings in a latent space, each rule can be expressed as a set of predicates conforming to a subset of Description Logic.