Knowledge Graph

A knowledge graph is a special way of storing knowledge. Instead of storing sentences or table rows, it stores individual things and the relationships between them. A thing can be a person, a city, a company, or a movie. Between these things lie named connections: “was born in,” “works for,” “is the capital of.” If you draw this out, a network of points and labeled arrows emerges. This network is exactly what the word graph refers to, and the labeling of the arrows is what turns it into knowledge rather than just a pile of data.

Why facts in a network are worth more than facts in a table

In a normal table, each piece of information stands on its own. You can look things up in it, but it’s hard to reason further with it. A knowledge graph, on the other hand, allows chains of inference. The question “Which directors have made films with actors from my hometown?” links three relationships in sequence. In a table, this would be an elaborate search; in a graph, it’s simply a path along the arrows.

A second advantage is extensibility. A table has fixed columns, and a new kind of information often forces a new column for all entries. In a graph, you simply attach an additional arrow to a single thing. That’s why knowledge graphs are well suited to knowledge that keeps growing and is never fully complete.

Uniqueness also matters. In the graph, “Paris” is not merely a word but a specific entry with its own identifier. The city in France and the small town in Texas are two different points in the network. A text document cannot make this distinction, but a graph can.

Triples, ontologies, and the question of where the data comes from

The smallest unit of a knowledge graph is called a triple and consists of three parts. One example: “Marie Curie” – “received” – “Nobel Prize in Physics.” Millions of such triples together make up the graph. Because all entries have the same simple form, a computer can easily search and combine them.

To keep the network from ending in chaos, there is a set of rules for the allowed types and relationships. This rulebook is called an ontology or schema. It specifies, for example, that a person can be born, whereas a company is founded. Such rules prevent the same relationship from ending up in the graph in five different spellings.

Knowledge graphs are filled in three ways. Some facts come from existing databases, such as library catalogs. Others are entered by hand by people, as with the open knowledge collection Wikidata. And a growing share is extracted automatically from text, with a program recognizing names and guessing the relationships between them. The automatic route is the fastest and, at the same time, the most error-prone.

From Google's info box to banks' fraud detection

The most well-known example is the box next to Google’s search results showing date of birth, occupation, and related people. It comes from Google’s own Knowledge Graph, which gave the method its name. Voice assistants also draw on this when answering questions like “How old is the chancellor?”

In companies, knowledge graphs often serve to link separate systems. Banks use them to detect suspicious money flows, because unusual chains of transfers become visible as patterns in the network. Pharmaceutical companies connect genes, active substances, and diseases to find possible new drugs. Online stores suggest products by analyzing connections between items, brands, and buyer groups.

In current news, the term frequently comes up alongside language models. Such models sometimes invent facts because they string words together based on probability. A connected knowledge graph, by contrast, provides verified information with a traceable source. Incidentally, a common misconception is to mistake a knowledge graph for a form of artificial intelligence. It does not think; it merely organizes knowledge so that programs can work with it.