What is natural language processing? How computers read

Natural language processing, or NLP, is the branch of computer science and artificial intelligence that lets computers work with human language: reading it, sorting it, translating it, answering it and writing it. "Natural" means the languages people speak, like English or Hindi, as opposed to programming languages. Autocomplete, spam filters, voice assistants, translation apps and chatbots are all NLP. What changed over 70 years is how it's done: first with hand-written rules, then statistics, now neural networks.

In short

• NLP covers two jobs: understanding language that comes in, and generating language that goes out.
• The classic approach broke a sentence into steps, words, grammar, then meaning; modern models learn all of it at once from huge amounts of text.
• The hard part has always been ambiguity: the same words can mean different things, and people resolve it with context and knowledge of the world.

Why is human language hard for computers?

Because it is ambiguous at every level, and people barely notice. Take the sentence "The bank will not lend her money." A person knows at once that bank is a place that lends, not the side of a river, because of the word lend. But her is genuinely unclear: the bank may refuse to lend money to her, or refuse to lend out her money. A listener uses context to choose. A computer has only the characters, unless it has learned what the words usually do near each other.

Programming languages are designed so that every statement has exactly one meaning. Human languages grew without a designer. They are full of words with several senses, sentences with more than one structure, sarcasm, idioms ("kick the bucket"), and missing words the listener supplies. Most of the history of NLP is a history of different ways to handle ambiguity.

How does a computer read one sentence?

Textbooks such as Jurafsky and Martin's Speech and Language Processing describe the classic pipeline as a series of steps, each building on the one before. Here is the bank sentence going through it.

1. Tokenization. Split the text into units: The · bank · will · not · lend · her · money · . That's 8 tokens, counting the full stop.
2. Parts of speech. Label each word's grammatical role: The (determiner), bank (noun), will (modal verb), not (adverb), lend (verb), her (pronoun, either the person or the owner of the money), money (noun).
3. Parsing. Work out the structure: The bank is the subject, will not lend the verb phrase. Two parses fit what follows: her as the person and money as the thing lent, or her money as one unit.
4. Word sense. Pick the meaning of each word. Bank next to lend is the financial kind.
5. Meaning and use. Turn the result into something a program can act on: a negative statement, about a refusal to lend, with one ambiguity left unresolved.

Each step can fail, and errors pass down the line: a wrong part-of-speech tag gives a wrong parse, and a wrong parse gives a wrong meaning.

How does a language model handle the same sentence?

A modern large language model skips the separate steps. It still starts by cutting text into tokens, often whole words for common words and pieces of words for rare ones, and turns each token into a long list of numbers, called an embedding, that places it near words used in similar ways. Then layer after layer of a neural network lets every token take information from every other. By the time bank has passed through those layers, its numbers have been pulled toward money and away from rivers by lend.

No one wrote a rule saying "bank near lend means finance" or labeled her as a pronoun. The model picked up those patterns from billions of sentences during training, a form of machine learning. Ask it about the sentence and it will usually say her is ambiguous, and explain both readings, because text explaining ambiguity was among what it learned from. The trade-off is that its reasons are spread across billions of numbers, and are hard to inspect. The ChatGPT note follows the same process through to a reply.

A short history: rules, statistics, neural networks

Year	What happened
7 January 1954	The Georgetown–IBM demonstration in New York: an IBM 701 translates more than 60 Russian sentences into English, using 250 words and six grammar rules
1966	Joseph Weizenbaum's ELIZA imitates a therapist with pattern matching; the US ALPAC report finds machine translation far behind its promises, and funding is cut
Late 1980s–1990s	Statistical methods take over: systems learn from large collections of text, such as translations of the same documents into two languages, instead of hand-written rules
August 2002	Paul Graham's essay A Plan for Spam popularizes filters that score each word's odds of appearing in spam
2013	Word embeddings such as word2vec represent each word as a list of numbers learned from text
September 2016	Google moves Translate to neural machine translation, starting with Chinese to English
2017	Google researchers publish the transformer, in Attention Is All You Need, the design behind today's language models
30 November 2022	OpenAI releases ChatGPT

The Georgetown–IBM demonstration made front pages, and its organizers expected that translating almost everything would be possible within three to five years. As the translation historian John Hutchins records, the 250-word dictionary held only the vocabulary needed for carefully selected sentences; ordinary text would have broken it. Rule-based systems kept running into the same wall: every rule needed exceptions, and the exceptions needed exceptions. Statistics got around it by counting what actually happens in real text rather than writing down what should. Neural networks went further, learning their own representations of words and sentences. The wider history of AI followed the same arc.

Where do people meet NLP every day?

More often than they notice. A few common examples, and what each one does with language:

• Spam filters classify a message as wanted or not from its words.
• Autocomplete and predictive text guess the next word from the words before.
• Translation maps a sentence in one language to one in another.
• Voice assistants turn speech into text, work out the request, and answer, often by turning text back into speech.
• Search engines match the words of a query to the meaning of pages, not just to the same letters.
• Chatbots and writing tools generate whole passages in reply.

Spam filters and search mostly need understanding; autocomplete and chatbots mostly generate; translation and voice assistants do both. The field calls the two halves natural language understanding and natural language generation.

What can't NLP do well yet?

Fluency is no longer the main limit. Problems that remain include stating false things confidently, which is why AI hallucinates; doing much worse in languages with little text online than in English; and inheriting the biases of the text it learned from. Whether a system that handles language this well understands it, as a person does, is a separate and older argument, the one Alan Turing's imitation game and later the Chinese room were meant to settle. The wider field all this belongs to is artificial intelligence.

Check yourself

• Walk the sentence "The bank will not lend her money" through the classic pipeline, and say where the ambiguity is.
• What does a language model do instead of tagging parts of speech and parsing?
• Why did rule-based translation keep failing, and what did statistical methods do differently?

Sources

• Speech and Language Processing, 3rd edition draft — Daniel Jurafsky and James H. Martin, Stanford University, 2026
• The first public demonstration of machine translation: the Georgetown-IBM system, 7th January 1954 — John Hutchins, 2006
• A Plan for Spam — Paul Graham, 2002
• A Neural Network for Machine Translation, at Production Scale — Google Research, 2016
• Attention Is All You Need — Ashish Vaswani and others, 2017
• Natural language processing — Wikipedia