Neurosymbolic AI for Reasoning over Knowledge Graphs: A Survey

TL;DR

This survey addresses the problem of reasoning over knowledge graphs with neurosymbolic AI, merging symbolic logic with neural embeddings to improve interpretability and data efficiency. It introduces a three-category taxonomy—logically-informed embedding approaches, embedding approaches with logical constraints, and rule learning for KG completion—and systematically situates a broad set of methods within this framework. The paper provides a comprehensive comparison, discusses trade-offs in interpretability, scalability, and long-range reasoning, and highlights practical directions such as multimodal data integration and few-shot learning. By mapping approaches to five key characteristics and summarizing available code, the work aims to guide future research toward scalable, interpretable, and knowledge-aware KG reasoning with real-world impact.

Abstract

Neurosymbolic AI is an increasingly active area of research that combines symbolic reasoning methods with deep learning to leverage their complementary benefits. As knowledge graphs are becoming a popular way to represent heterogeneous and multi-relational data, methods for reasoning on graph structures have attempted to follow this neurosymbolic paradigm. Traditionally, such approaches have utilized either rule-based inference or generated representative numerical embeddings from which patterns could be extracted. However, several recent studies have attempted to bridge this dichotomy to generate models that facilitate interpretability, maintain competitive performance, and integrate expert knowledge. Therefore, we survey methods that perform neurosymbolic reasoning tasks on knowledge graphs and propose a novel taxonomy by which we can classify them. Specifically, we propose three major categories: (1) logically-informed embedding approaches, (2) embedding approaches with logical constraints, and (3) rule learning approaches. Alongside the taxonomy, we provide a tabular overview of the approaches and links to their source code, if available, for more direct comparison. Finally, we discuss the unique characteristics and limitations of these methods, then propose several prospective directions toward which this field of research could evolve.

Figures (6)

• Figure 1: Taxonomy of neurosymbolic approaches for graph reasoning.
• Figure 2:
• Figure 3:
• Figure 4: Learning Representative Logic Embeddings. Some methods encode logical constraints as embeddings, then combine them with KGEs to make more informed predictions.
• Figure 5: EM Algorithm Methods. These approaches adopt the idea of iterative optimization from the EM algorithm dempster1977maximum to describe systems which iterate between symbolic and neural modules. Typically, the E-step involves KGC through a KGE method (the neural module), and the M-step involves updating the parameters of the symbolic module. To make the symbolic module (typically in the form of a MLN or rule mining approach) dynamic, some methods update rule confidences (§\ref{['learning_rule_weights']}). Other methods update and alter a pool of candidate rules (§\ref{['iterative_rule_mining']}). Note that approaches in this category may deviate from this portrayal.