Ontology and semantics

TL;DR:

Ontology in IT is a way to structure knowledge about the world in the form of related categories and their properties.

For example, in the "Game of Thrones" ontology, there are categories "houses", "characters" and connections between them. When we fill the ontology with real data, we get a knowledge graph.

Semantic layer is a more abstract concept that includes all ways of giving meaning to data.

Together they help AI systems better understand the context and provide more accurate answers.

Dynamic ontologies are used to identify hidden connections in large amounts of data, which helps make more informed business decisions.

Ontology

Ontology is a formal representation of categories of objects, their properties, and interactions.

If we take Game of Thrones as an example, the ontology will include categories such as "great houses", "castles", "characters", etc.

In this ontology, the connections between categories can be as follows:

• a → allies with → b

• x → rules → y

• z → swore allegiance → w

In the application to LLM, ontology is a description of information in a structured form. And if we add the information itself to this structure (for example, Baratheons, Starks, Winterfell, etc.), we get a graph.

That is, if ontology is a data model, then a graph is an implementation of this model.

The task of ontology is to create a bridge between raw data and their semantic concepts, thereby providing a convenient format that is both understandable to humans and can be processed algorithmically.

Many modern systems that use large language models apply RAG approaches - that is, first, information relevant to the user's query is selected from a structured data repository, and then this selection is fed to the LLM to form a meaningful response. So, part of such RAG systems are built on graph databases, such as Neo4j.

Using a graph schema that includes entities and their relationships, LLM systems can better understand the context, more accurately select relevant information, and minimize hallucinations when forming a response.

In a simple case, creating a graph schema, or ontology, involves defining the concepts of the subject area and their relationships, and is usually done by experts.

However, when working with large volumes of data, hidden relationships begin to emerge sooner or later, for example, the dependence of sales on the weather forecast, or the connection between the quality of order delivery and World Cup football matches. And such dependencies are usually very difficult to understand and predict. It is almost impossible to identify such connections expertly.

Dynamic Ontology

Dynamic ontology comes to the rescue. There are systems that allow building ontologies automatically based on raw data. That is, the system itself identifies significant entities and their relationships.

Platforms like Palantir allow accumulating raw data from various sources (e.g., ERP, CRM, IoT sensors, clickstream, video stream, etc.), identifying patterns in them, predicting the course of business processes, and forming logical and reasoned decisions.

By integrating and processing data, predicting events and actions, such platforms build a digital model of our world, with which both users and AI agents can interact.

Superbrain

You can go even further. If we have a system to which all our business applications and data stores are connected (say, SAP, Salesforce, Jira, Ms Office, etc.), and this system can make business decisions (e.g., whether to order car maintenance or send a reminder to a client), then this system will not only be able to suggest what needs to be done to improve business metrics but also do it independently.

Everything would be fine, but ....

But no matter how attractive the prospects of creating a superintelligence for business are, such technological platforms bring a number of new tasks. A dynamically created ontology still needs human validation. Hence, processes of management, control, versioning, testing, and monitoring of ontologies arise. Can you already imagine a department that will deal with Ontology Governance? I won't even mention the technical complexity.

In everyday life

Is ontology really that difficult?

In fact, ontologies in graph structures are quite common. There are even standards for describing ontologies. The most famous are RDF and OWL.

Resource Description Framework (RDF) is a framework for modeling and exchanging data on the internet. It includes a set of specifications for representing data in the form of triplets: subject-predicate-object.

For example: Sansa Stark → Ruler of → Winterfell

Web Ontology Language (OWL) is a language for creating ontologies on the internet. It is based on RDF and provides a way to describe conceptual entities (classes) and their relationships.

Semantic Layer

In conclusion, a few words about semantics.

Unlike ontology, which is expressed in a structured form (in the form of code or a graph), has standards, frameworks, and its own language, semantics (or semantic layer) is a generalized concept, the essence of which is to give meaning to data.

Semantic layer is an umbrella term that can represent a description of a data model, metadata, interaction rules, policies, API specifications, caching and data transformation mechanisms, and any other information that allows our data to have some meaning.

The task of the semantic layer is to ensure data consistency and transparency.

Imagine that we have sales data from a CRM and delivery data from a logistics system. The semantic layer may include the definition of a "problem order", a description of order mapping between systems, delivery time calculation formulas, etc.

Unlike the strict structure of ontology, the semantic layer can be represented in text format, configuration files, code, diagrams, etc.

What else to read on the topic

1. Graph Model Ontology, An Introductory Guide (graph.build)

2. Foundational Ontologies in Palantir Foundry (Medium)

3. Semantic Model vs Ontology vs Knowledge Graph (Medium)

4. Connecting AI to Decisions with the Palantir Ontology (Medium)