Neurosymbolic AI for Reasoning Over Knowledge Graphs: A Survey

Area of Science:

• Artificial Intelligence
• Machine Learning
• Knowledge Representation

Background:

• Neurosymbolic AI merges symbolic reasoning and deep learning for enhanced capabilities.
• Knowledge Graphs (KGs) are increasingly used for complex data representation.
• Existing KG reasoning methods are often rule-based or embedding-based, creating a dichotomy.

Purpose of the Study:

• To survey existing neurosymbolic reasoning methods for Knowledge Graphs (KGs).
• To propose a novel taxonomy for classifying these neurosymbolic KG approaches.
• To identify limitations and suggest future research directions in the field.

Main Methods:

• Literature review of neurosymbolic reasoning techniques applied to KGs.
• Development of a new classification taxonomy with three categories: logically informed embeddings, embeddings with logical constraints, and rule-learning approaches.
• Comparative analysis of methods, including tabular overviews and source code links.

Main Results:

• A comprehensive survey of neurosymbolic KG reasoning methods.
• A novel taxonomy categorizing approaches into three main types.
• Identification of the strengths and weaknesses of current neurosymbolic KG methods.

Conclusions:

• Neurosymbolic reasoning offers a promising path to bridge the gap between symbolic and subsymbolic AI for KGs.
• The proposed taxonomy provides a structured framework for understanding and comparing diverse methods.
• Further research is needed to fully realize the potential of neurosymbolic AI in KG reasoning, focusing on interpretability and expert knowledge integration.