Neuro-Symbolic Techniques for and with Ontologies and Knowledge Graphs

The summit will survey current techniques that combine neural network machine learning with symbolic methods, especially methods based on ontologies and knowledge graphs.

Ontologies are representations of a knowledge domain. They define the concepts, relationships, properties, axioms and rules within that domain, providing a framework that enables a deep understanding of that subject area. Knowledge graphs are structured representations of semantic knowledge that are stored in a graph. Ontologies and knowledge graphs are used to enable machine reasoning and semantic understanding, allowing a system to draw inferences and to derive new information and relationships between entities.

Neural network and other machine learning models, such as LLMs, are trained on large corpora, learning the patterns and connections between words and images. Hence, although their “knowledge base” is broad, it is also sometimes incorrect and/or biased, and don't explicitly understand the semantics or relationships in that content.

Consequently, neural network and traditional AI techniques are complementary. The Fall Series of the summit explored the similarities and distinctions between ontologies and LLMs, as well as how they can be used together. The Main Summit Series will examine the more general topic of neuro-symbolic techniques, especially how one can leverage the complementary benefits of neural networks and of ontologies and knowledge graphs.

Fall Series Theme

Ontologies and Large Language Models (LLMs) such as OpenAI's GPT-4 represent two different but related concepts within the fields of artificial intelligence and knowledge representation.

Ontologies are representations of a knowledge domain. They define the concepts, relationships, properties, axioms and rules within that domain, providing a framework that enables a deep understanding of that subject area. Ontologies are used to enable machine reasoning and semantic understanding, allowing a system to draw inferences and to derive new information and relationships between entities.

On the other hand, LLMs are machine learning models that aim to generate human-like responses (including text and images) based on an input (“prompt”). They are trained on a large corpora of (mostly online) text, learning the patterns and connections between words and images. Hence, although their “knowledge base” is broad, it is also sometimes incorrect and/or biased. LLMs generate new content based on their training data, but don't explicitly understand the semantics or relationships in that content. This mini-summit explores the similarities and distinctions between ontologies and LLMs, as well as how they can be used together. In addition, the success of LLMs has generated much interest in AI and machine learning. This can be leveraged to promote the benefits of, and increase awareness of, the value of ontologies.

Fall Series Topics

• Semantic Understanding and Knowledge Representation: Both ontologies and LLMs represent “knowledge”. Ontologies explicitly capture data’s semantics regarding entities and their relationships, and are based on a formal, structured, axiom-based/logical representation of a domain of knowledge. LLMs, in contrast, have implicit, probabilistic representations of “knowledge”, based on the patterns in their training data.
  • "What do we mean by knowledge representation?” Internal vs external vs formal knowledge, a sliding bar
• Technical Assistance and Hybrid Systems: Both ontologies and LLMs can be used to build question-answering systems, information extraction systems, chatbots and a variety of technical assistants. They can be used together to improve functionality and correctness. For example, LLMs can be used to extract information from text and aid in its mapping to an ontology. LLMs can make ontologies more accessible and usable for non-expert users. In turn, ontologies can be used to formulate prompts to the LLM, or to validate the responses of the LLM.

Main Series Tracks

• Track A. Foundations and Architectures
• Track B. Large Language Models, Ontologies and Knowedge Graphs
• Track C. Applications
• Track D. Risks and Ethics