AI and Enterprise Knowledge Integration: Part 2

Posted on: November 5, 2018 by Thierry Caminel

The great Renaissance man Leonardo da Vinci once said: “Learn how to see. Realize that everything connects to everything else.” This mighty challenge is more or less where I left off in Part 1 of this blog.

In most big companies, the “knowledge landscape” is fragmented. Data lives in autonomous silos, in various formats (databases, documents…) and suffers from “semantic incoherency” (contradictory meanings). Humans can integrate it into usable knowledge, but with lots of time and effort.

This is a big problem for extending the use of AI in digital transformation to assist or augment complex knowledge intensive work. Of course, (what I called) AI 2.0 technologies like statistical machine learning and Deep Neural Networks (DNN) do have amazing numbers of use cases, but they are designed for specific “narrow” tasks with homogeneous data and, above all, can’t reason.

To move forward, we need to “connect up” different forms of enterprise knowledge, with the help of technologies from the “Symbolic AI” tradition, where meaning and reasoning take center stage.

Symbolic AI in brief

Symbolic AI focuses on high-level "symbolic" (i.e. human readable) representations of problems, logic and search, what John Haugeland famously dubbed "Good Old-Fashioned Artificial Intelligence".

One outcome was Expert Systems, which connect symbols in “If-Then” relationships to make deductions. While seen as a failure, their descendants – Business Rules Engines – achieved widespread adoption. Natural Language Processing (NLP) also began in Symbolic AI, although it has evolved with statistical machine learning.

All communication relies on symbols with meaning (i.e. they are “semantic”) and, by linking them together, we create knowledge and enable reasoning. Let’s take an example of how we human beings acquire knowledge over the years.

Human learning and Symbolic AI

Consider a baby. The first time a baby encounters a cat, the mother might say “cat”. Later the baby meets another fuzzy, purring animal, and mother again says “cat”. Perhaps the third time, the baby points and says “cat” … and language is born! It doesn’t take long for the baby to learn related terms like cat food, cat bed, cat litter… in short, a complete “cat ecosystem”.

As a baby grows into a child and then a teenager, knowledge is acquired in new ways, through abstraction and applying rules and learning to reason. From personal experience, I wouldn’t claim teenagers are always reasonable, but they certainly can reason!

The inspiration for Symbolic AI was human cognition: learning symbols, connecting them in semantically meaningful ways and applying the rules of reasoning to generate new knowledge … everything from “cat ecosystems” to scientific theories to (what interests us here) the workings of a business.

Now let’s introduce a concept – “ontology” – which is crucial for making those connections.

What’s an “ontology”?

The term comes from Philosophy, but was brought to AI by Tom Grubber, who went on to invent Siri, Apple’s pioneering virtual assistant. Just for fun, here’s Grubber’s original scholarly definition: “An ontology is a formal explicit specification of a shared conceptualization.” A bit too scholarly, perhaps? Let me explain in simpler language.

Ontologies are symbolic models in AI to express knowledge that is best represented as a graph of interconnected concepts and that is usable by computers. In an enterprise context, the concepts can be things like clients, products, parts, suppliers and invoices, as well as their numerous interrelationships in the course of doing business.

Take a look at the example below – a graphical representation of a simplified ontology.

The rectangles (“entities”) are business concepts we work with while the arrows (“relationships”) represent how they are connected (unidirectional, bidirectional), and both of course have “attributes”.

The way we assemble and give meanings to these building blocks depends on the company, its markets and how it does business. We build ontologies for entire companies or, as often the case, for knowledge intensive domains with digital transformation in mind.

Of course, our example resembles the “conceptual data models” used for designing databases before SQL ate the world, but there’s a big difference. To turn our diagram into “computational knowledge”, ontologies have a “secret sauce” – an Ontology Encoding Language. At Atos we use OWL – the standard of the Semantic Web – to provide semantics (meanings) for data, together with logical formalisms to enable machine reasoning.

Still, however powerful the encoding language, only business focus can make an ontology meaningful. We want business knowledge for business outcomes, in business language for business people.

To achieve those outcomes, we have to expand our ontologies from structured ideas to real world facts, from connected business concepts to integrated enterprise knowledge. We do this mainly with Knowledge Graphs.

Enterprise Knowledge Integration with Knowledge Graphs

Google coined the term in 2012 when, to improve search results, it announced a Knowledge Graph that “understands real-world entities and their relationships to one another….. It’s the intelligence between these different entities that’s the key.”

As Kirk Borne, Principal Data Scientist at Booz Allen Hamilton, says “The natural data structure of the world is not rows and columns like spreadsheets or even relational databases. The natural data structure of the world is a graph.”

Knowledge graphs are “smart” because they inherit their semantics from underlying ontologies (the meaning is encoded alongside the data itself), together with the ontology’s logical formalisms for machine reasoning.

Implementing Knowledge Graphs

Usually we implement Knowledge Graphs with powerful graph databases where both entities and relationships are physically supported by the database. This makes them ideal for complex queries into connected up enterprise knowledge.

Populating knowledge graphs – and then enriching them – can be challenging, because there are so many possible sources, as in the figure below.

Lots of detailed technical work may be needed to build operational knowledge graphs, but the business benefits are worth it.

Business outcomes

Leading edge companies around the world are already leveraging knowledge graphs (and related technologies from semantic AI) to integrate enterprise knowledge and achieve impressive business outcomes.

The benefits of knowledge integration are to be found across at least the following four areas: customer experience, operational efficiency, business reinvention, and trust and compliance.

In the concluding post of this series, we’ll look at real world examples in each of these areas that show how “connecting up” enterprise knowledge opens new vistas for the transformative force of AI.

Keep an eye on this space!

Share this blog article

  • Share on Linked In

About Thierry Caminel
AI Technology and Innovation Leader and member of the Scientific Community
Thierry Caminel is Distinguished Expert, leader of the Atos Expert Community Data Intelligence domain, and a member of the Scientific Community. His main focus areas are in Artificial Intelligence, Data Analytics, Semantic Technologies, IoT, micro-services, distributed systems. Before joining Atos he worked in several startups innovating in the field of AI, IoT and Embedded Systems. Thierry holds a software engineering degree, and lives in Toulouse.

Follow or contact Thierry