Entity

A sentence like “Tim Cook presented the new iPhone in Cupertino” contains things that a human recognizes instantly: a person, a product, a place. Precisely such uniquely nameable things are called entities in computer science. An entity, then, is not simply a word, but something that exists in the world and has a name. This includes people, companies, cities, countries, products, amounts of money, and dates. The rest of the sentence – “presented”, “the new”, “in” – merely describes what happens between these things. Software that analyzes texts first looks for the entities and then for the relationships between them.

From a lump of text to usable information

To a computer, a text is initially just a long chain of characters. It can store and display it, but it cannot compute with it. Only once it is clear that “Cupertino” is a place and “Tim Cook” is a person can meaningful questions be asked. For example: In which cities did Apple present products last year? Recognizing entities is thus the step from unstructured text to organized data.

This is especially important in the world of finance. News agencies and trading platforms analyze tens of thousands of reports every day. A program must recognize which company is meant in a report in order to match it to the correct stock price. If a report mentions “Alphabet”, it must be clear that the Google parent company is meant, and not the alphabet from grade-school lessons. Such mix-ups can get expensive.

A second reason is traceability. Language models formulate fluent answers, but their statements are hard to verify. If a system instead outputs which person is connected to which company, that individual statement can be checked in a targeted way. Entities thus make results verifiable.

Recognizing and correctly linking

The first work step is called Named Entity Recognition, often abbreviated as NER. Here, a program marks every place in the text that is an entity and assigns each one a category: person, organization, place, date. Earlier systems worked with long lists and fixed rules, for example: a capitalized word after “Mrs.” is a surname. Today’s systems learn this from many thousands of example texts in which people have manually marked the entities.

The second step is harder and is called entity linking. Here, the text passage that was found is linked to a unique entry in a database. “Paris” can be the French capital, a city in Texas, or a first name. Only the context decides: if “Eiffel Tower” appears in the same paragraph, the capital is meant. Without this step, names remain just letters without a fixed reference.

Conversely, there is the same problem in reverse: one thing can have many names. “VW”, “Volkswagen”, and “the Wolfsburg-based group” all refer to the same company. A good system traces all three back to a single entity. A typical mistake, by the way, is confusing entity with keyword. A keyword is merely a frequent word; an entity is an identifiable object with its own identity.

Entities in search engines, chats, and market data

The principle is most visible with Google. If you search for a soccer player, a box appears on the right with date of birth, club, and height. This information comes from a huge collection of entities and their relationships, the so-called Knowledge Graph. There, Google doesn’t answer the search with text passages, but with facts about a specific thing.

Voice assistants work the same way. If you say “Remind me tomorrow at eight about the dentist”, the device breaks the sentence down into a date, a time, and an occasion. These building blocks are entities, and without them no appointment could be created. Chatbots in online shops use the same procedure to pull order numbers and product names out of your message.

In business news, you often encounter the term in connection with automated analysis. Providers promise to sort reports by company and assess the sentiment associated with them. Both require that companies be reliably recognized as entities. Whoever does this first step poorly ends up producing false statistics – no matter how good the rest of the technology is.