# How AI Understands Words (Embeddings) When you type something into an AI system, the computer does not see words the way you do. You see meaning. The computer sees symbols. So before AI can do anything useful with language, it has to convert words into a form it can work with. That form is numbers. This is where **embeddings** come in. Embeddings are one of the most important ideas behind modern AI. They help AI systems compare meanings, find related words, understand search queries, recommend content, translate languages, and generate responses. So in this lesson, I want you to understand one big idea: **AI does not understand language by reading words directly. It represents meaning using numbers.** That may sound strange at first, but we will break it down slowly. ## What Is an Embedding? An **embedding** is a numerical representation of something. That “something” could be: * a word * a sentence * a paragraph * an image * a product * a user profile But for now, let’s focus on words. A word embedding is a list of numbers that represents the meaning of a word. For example, the word **cat** might be represented like this: ```text cat → [0.23, -0.45, 0.67, 0.12, -0.33] ``` Please don’t worry about the exact numbers. The point is not to memorize them. The point is to understand this: **The AI converts the word into numbers so it can compare it with other words mathematically.** So another word like **kitten** may have a similar list of numbers: ```text kitten → [0.25, -0.43, 0.65, 0.14, -0.30] ``` These numbers are close to the numbers for **cat**. That tells the AI that the two words are related. But a word like **airplane** would have a very different set of numbers. That tells the AI that **airplane** is not very close in meaning to **cat**. This is the heart of embeddings: similar meanings are represented by similar numbers. ## Why Does AI Need Embeddings? Computers do not naturally understand words. To a computer, the word: ```text dog ``` is just text. It does not automatically know that a dog is an animal, that it is related to puppy, or that it is different from a car. So AI needs a way to turn words into something measurable. Numbers are measurable. Once words become numbers, the AI can ask questions like: * How close is “dog” to “puppy”? * Is “king” more related to “queen” or “banana”? * Is this search query similar to this document? * Is this sentence positive or negative? * Which word is most likely to come next? Without embeddings, AI would struggle to compare meanings. It would treat words mostly as separate symbols. With embeddings, AI can begin to work with meaning as a mathematical pattern. ## A Simple Analogy: A Map of Meaning Think about Google Maps. On a map, places have coordinates. Toronto has one location. London has another. Accra has another. Places that are close together on the map usually have similar coordinates. Embeddings work in a similar way, but instead of mapping physical places, they map meaning. So imagine a large “meaning map.” On that map: ```text cat, kitten, puppy, dog, pet ``` would be close to one another. But: ```text cat, democracy, volcano, spreadsheet ``` would be far apart. The AI is not looking at the spelling of the words only. It is looking at where those words sit in this mathematical space. That space is called **vector space**. A **vector** is simply a list of numbers. So when we say a word is represented as a vector, we mean the word has been turned into a numerical position in meaning space. ## How Does AI Learn These Meanings? AI learns embeddings by looking at many examples of language. It does not learn meaning from a dictionary first. It learns from context. Here is the simple rule: **Words that appear in similar contexts often have similar meanings.** For example: ```text The cat slept on the sofa. The kitten slept in the basket. The dog slept near the door. ``` The AI sees that **cat**, **kitten**, and **dog** appear in similar kinds of sentences. They appear near words like: ```text slept, sofa, basket, door, pet, animal ``` So the AI learns that these words are related. Now compare that with: ```text The telescope captured images of a galaxy. ``` The word **galaxy** appears in a very different context. So its embedding will be far away from words like **cat** and **kitten**. This is how AI begins to build meaning. Not by being told directly. But by seeing patterns again and again. ## Distance Matters Once words become vectors, the AI can measure distance between them. If two vectors are close, the meanings are similar. If two vectors are far apart, the meanings are different. For example: ```text king → close to queen doctor → close to hospital teacher → close to school cat → close to kitten cat → far from astronomy ``` This is why embeddings are powerful. They allow AI to compare meaning using mathematics. So when you search for something online, the system may not only look for the exact words you typed. It may also look for words and documents that are close in meaning. That is why you can search: ```text cheap flights to London ``` and still get results about: ```text budget airfare to the UK ``` The words are not exactly the same. But the meaning is close. ## Direction Also Matters Embeddings do not only capture distance. They can also capture relationships. A famous example is: ```text king - man + woman ≈ queen ``` This means the AI has learned a relationship between words. The relationship between **king** and **queen** is similar to the relationship between **man** and **woman**. Another example: ```text Paris - France + Germany ≈ Berlin ``` The model learns that Paris relates to France in a similar way that Berlin relates to Germany. This does not mean the AI is thinking like a human. It means relationships have been captured mathematically. That is still impressive. A little scary too, but mostly impressive. ## What Embeddings Are Not Now let’s clear up a few things. An embedding is not the definition of a word. For example, the embedding for **cat** does not literally say: ```text A small domesticated animal that often says meow. ``` Instead, it stores the word’s position based on patterns. An embedding is also not human understanding. The AI does not “know” what a cat is the way you do. It does not remember touching a cat, hearing it meow, or watching it knock something off a table for no good reason. It only knows the patterns connected to the word. That difference matters. AI can be very good at pattern recognition without truly understanding the world like a person. ## Common Beginner Mistakes ### Mistake 1: Thinking embeddings are just random numbers They are not random after training. At the beginning, the numbers may start randomly. But during training, the AI adjusts them again and again until similar words have useful positions. ### Mistake 2: Thinking one word always has one meaning This is not always true. Take the word: ```text bank ``` It can mean a financial institution: ```text I deposited money at the bank. ``` Or it can mean the side of a river: ```text We sat by the river bank. ``` Modern AI systems use context to handle this. So the embedding for **bank** may change depending on the sentence. ### Mistake 3: Thinking AI understands because it gives a good answer A good answer does not always mean true understanding. The AI may simply be very good at predicting patterns. That is why we still need human judgment. Especially in education, healthcare, law, finance, and other serious areas. ## Mental Model Here is the simplest way to hold the idea in your mind: **An embedding is a location for meaning.** A word becomes a point in a large mathematical space. Words with similar meanings are close together. Words with different meanings are far apart. Relationships between words can appear as directions in that space. So embeddings allow AI to turn language into something it can measure, compare, and process. ## Practice Thinking Try answering these for yourself: 1. Why would **nurse** be close to **hospital**? 2. Why would **apple** be close to **banana**, but far from **airplane**? 3. Why might **teacher** and **lecturer** be close, but not exactly the same? 4. What might happen if an AI learned embeddings from biased or poor-quality text? 5. Why is context important for a word like **bank**? You do not need perfect answers. The goal is to start thinking like the system. ## Key Takeaways * AI cannot directly understand words, so it converts them into numbers. * An embedding is a list of numbers that represents meaning. * These numbers place words inside a mathematical meaning space. * Similar words have similar embeddings. * AI learns embeddings by studying how words appear in context. * Distance shows similarity. * Direction can show relationships. * Embeddings are powerful, but they are not the same as human understanding. ## What’s Next Now you understand how AI turns words into numbers. But that is only the first step. The next question is: **Once the AI has these numbers, how does it process them?** That takes us into neural network layers. In the next lesson, we will look at how data moves through layers, how each layer transforms information, and why deeper layers can learn more complex patterns.