Knowledge Graphs: Artificial Knowledge for Artificial Intelligence

Intelligence is simply being good at thinking: at using what you know to make sense of what you don't.? That might be understanding new kinds of things, getting around incomplete information, or drawing different kinds of conclusions. There's always something new or different to think about. In complex or challenging environments, new things come at you every second.

A bunch of us techies would like to get computers to do some of this thinking for us. That way, they will help humans to think even faster, broader, and deeper. With help, we can find and synthesize information more easily. We can understand other people better. We can make diagnoses more reliably. Or find more effective solutions for social and climatic challenges.

If intelligence means doing a good job of using knowledge, then a key challenge for AI is to stuff as much knowledge into computers as we can.? In short, for artificial intelligence, we need artificial knowledge, otherwise machines can only go through the motions.

Humans store the vast majority of available knowledge in one of three places.? In the brains of individual people, in spreadsheets or databases, or in vast archives of written documents like what you would find in a library. People are very good at gathering, organizing, and storing knowledge; think of all the experts you've heard of or met.? But it's very, very difficult to “translate” brain activity into a format that computers can make sense of and then build on. We simply don't know how to do that yet.

Spreadsheets are way easier for computers to manipulate but they contain only a few kinds of information:? mostly numbers. For example, they can't store recipes, videos, or stories in an easy-to-use way. They're just not as informative as we would like.

We're left with written documents like books, articles, emails, text messages, reports, web pages, etc. as our best chance to get knowledge for computers. There are gazillions of these documents, about every conceivable topic, so this sounds very promising. On top of that, many humans – especially researchers – look at written documents and feel like they understand them effortlessly. Great! We can shovel tons of documents into computers (they think) to fill our computers with all the knowledge we can find and that will make them smarter computers. That was part of the reasoning behind the Big Data movement, which succeeded in making trillions of words of documents available for analysis.

Unfortunately, life just ain't so simple.?

An ocean of letters isn't the same thing as a mountain of know-how. Reading is only easy when you're reading about familiar topics written in familiar words. You can make sense of the words because you already know about the topic and the words.? But there are lots of situations where getting information from letters isn't easy. Maybe you encounter a new topic in school, but it's explained with familiar words – like in a really good textbook.? That's harder but still possible. Or maybe you come across a familiar topic that's been described with rare words and unusual, convoluted sentences –? like in research articles or literary works.? This is often very hard, but again still possible because we have enough knowledge to guess what the new words mean. In an extreme case, you might try to read something about a familiar topic that's been written in a totally new language, say Mongolian. That takes even very intelligent people a long time. It's very slow and difficult in part because to succeed at this we have to do it with a little help from our friends – like teachers and dictionaries.?

Now think of computers. They try to “read” or get information from letters in the worst possible scenario:? they start have no knowledge of familiar topics and no familiar wording.? They don't have even a smidgen of knowledge about the world to start with.? They're drowning in letters with no foothold. Computers simply don't have what they need to make sense of the letters or words or sentences, no matter how many we provide. What AI has given us recently is a very effective way to learn the patterns of the letters and words. So much so that it can imitate extremely well how humans would create sentences. But conceptual knowledge is still out of reach.

On to Artificial Knowledge

The next big challenge for AI – the missing link – is to create ways to describe the parts and combinations of the concepts that make up knowledge and meaning. That way, we can show computers how to link spellings with concepts. We can teach them strategies to combine concepts in new ways. And we can show them how to fill in gaps when they get incomplete information.??

Fortunately, researchers have put a huge amount of effort into understanding just how to describe knowledge. They actually started centuries ago (!) but most recently they made a lot of progress during the phase many call “first-generation” AI, some 50 years ago. So we're by no means starting from scratch; we know it's possible.??It just wasn't possible to do at scale with these older approaches.

To understand how to build knowledge, think about how we describe new concepts to people.? For children, we start building concepts for concrete things in the world. We use short definitions (like from a dictionary) or longer explanations (like from an encyclopedia). In person, we point at and describe different parts of these things. These descriptions “unpack” each concept to list the characteristics that are typical and the ways in which it's different from other concepts. They also include some of the ways this concept is related to others. There are typical kinds of information that we include in definitions:? the kind of thing our concept is about, its parts, what it's used for, where it comes from, etc. So we can actually judge how complete and detailed specific definitions are.?

In sum, it's practical to think of a concept as a collection of different kinds of facts about the same thing. These facts can be organized and expressed in different ways, most commonly as a sentence or paragraph (for humans) but also as lists, branching structures (like taxonomy trees), or networks (like knowledge graphs) to make things more convenient for computers.? Note that any given concept or collection of facts will usually have many different names both in one language and across many. -- hundreds! These are what we usually call synonyms. And the same fact can be described with many different (but equivalent) sentences – what we usually call paraphrases.? In practice, this diversity in names and phrases is a huge blocker for search and other uses of computers:? we express our needs with one word, but writers of the information we need often use other words.???

When we create and store concepts (represented by these definitions or facts) in consistent, systematic ways we start to build artificial knowledge.? Knowledge that computers can and increasingly do manipulate in a range of useful ways.? In techno-babble, we call these (sometimes enormous) collections of interconnected concepts and facts knowledge graphs or ontologies.??

One way to build these collections is to “translate” regular sentences – with all their synonyms and paraphrases – into facts for computers where the terms only ever mean one thing (which we document carefully and store the synonyms elsewhere). That way, we (and the computer) always know what a label means.? Since the same concept is connected in many different ways to many other unique concepts, the collection of facts ends up looking like an interconnected network or graph. Avoiding synonyms and paraphrases ends up making knowledge graphs thousands of times smaller than the same knowledge would be if we stored it in sentences. Which in turn makes the knowledge that much easier to accumulate, process, and check.

Scientists are already building ontologies like these to share and compare their data and their knowledge. Lawyers are busy building knowledge graphs like these to search much more precisely for precedent in law and for evidence in client data, even when the wording varies dramatically. Financial service companies build and use ontologies to understand patterns of behavior that look very different but are all cases of fraud. AI researchers leverage knowledge graphs to get chatbots to provide more sensible and relevant answers. Artificial knowledge guides how businesses aggregate data, how they break down siloed data, how they infuse knowledge into their processes where they once had only data.

Artificial knowledge is the next frontier for artificial intelligence.?

For more information about the world of artificial knowledge, you can read about:?

• How knowledge graphs are like fancy databases of artificial knowledge
• How data is the new mud so we need more artificial knowledge
• The limits of spelling-based reasoning without artificial knowledge

Steps for building artificial knowledge:

• Knowledge Graphs add layers of value, part 1
• Knowledge Graphs add layers of value, part 2
• Knowledge graphs add layers of value, part 3?

And you can follow my newsletter on Knowledge Architecture for more.