What is natural language processing?
Stanford's Institute for Human-Centered AI defines it as a branch of artificial intelligence "focused on enabling computers to understand, interpret, and generate human language in a meaningful way." In plain words, NLP is everything that turns messy human text or speech into something a program can use, and back again.

| Question | Answer | Source |
|---|---|---|
| First public demo | 7 January 1954, more than sixty Russian sentences translated, 250 vocabulary items, six rules | Wikipedia, Georgetown and IBM experiment |
| Statistical turn | 1990, IBM team trains translation on Canadian Parliament text | Wikipedia, NLP history |
| Engine behind today's models | The transformer, June 2017 | Vaswani et al., arXiv |
| Size of GPT-3 | 175 billion parameters, May 2020 | Brown et al., arXiv |
Two halves make up the field. Understanding takes language in and works out what it means. Generation puts language out. A spam filter only understands. A chatbot does both.
Why is human language so hard for computers?
Language is hard because the same words carry different meanings depending on what surrounds them. A person fills the gaps without noticing. A program has to be taught how, and that's the whole job of NLP.
Take three short sentences. "I saw her duck" could be about a bird or a dodge. "Apple fell" could be a share price or a fruit. "Great, another meeting" is praise on the page and a groan in the room. Rule writers in the 1950s tried to list every case. They didn't finish, and neither will anyone else.
That's why the field moved from rules to data. Instead of writing down what a sentence means, researchers let a program read millions of sentences and learn which readings are likely. Every era in the history below is a different answer to the same problem.
Is NLP the same as neuro-linguistic programming?
No. Neuro-linguistic programming is a self help method that Richard Bandler and John Grinder introduced in their 1975 book The Structure of Magic I, and Wikipedia describes it as a pseudoscientific approach to communication and psychotherapy.
It has nothing to do with computers. When this page, a job ad for a data team or a machine learning course says NLP, it means natural language processing. If the surrounding words are therapy, coaching or persuasion, it means the other one.
What are the two types of NLP, NLU and NLG?
The two halves are natural language understanding (NLU) and natural language generation (NLG). NLU takes text or speech in and works out what it means, as in sentiment analysis or entity recognition. NLG puts language out, as in summaries, translations and chatbot replies.
Most products use both. A voice assistant runs speech recognition and NLU on what you said, then NLG to write the answer. Stanford's definition names the same pair: understand, interpret and generate.
What are the main NLP tasks?
Wikipedia's list of NLP tasks runs to dozens of entries, from tokenization to argument mining. Eight of them cover most of what you meet in daily life. The task names come from that list and from Stanford's examples, and the plain examples are ours.
| Task | What it does | Where you meet it |
|---|---|---|
| Tokenization | Splits text into words or word pieces | The first step inside every chatbot |
| Part of speech tagging | Marks each word as a noun, verb or adjective | Grammar checkers |
| Named entity recognition | Finds names, places, companies and dates | Pulling an invoice date out of an email |
| Sentiment analysis | Scores a text as positive, negative or neutral | Sorting thousands of product reviews |
| Machine translation | Turns text in one language into another | Translating a web page |
| Speech recognition | Turns audio into text | Voice assistants and captions |
| Summarization | Shrinks a long text to its main points | A one paragraph digest of a report |
| Question answering | Replies to a question from a text or from memory | Search answers and chatbots |
Older explainers teach these as a pipeline: split, tag, find entities, then judge. That is how systems worked before 2017. Today one pretrained model usually handles several of the jobs at once. BERT's authors, for instance, reported a GLUE score of 80.5 percent, 7.7 points above the previous best, from a single pretrained model.
How did NLP develop from rules to LLMs?
NLP moved through four methods in about seventy years, and each one replaced hand work with more data. Rules came first, then statistics, then neural networks, then very large language models. The chart marks the milestones with a date you can check.
| Era | Years | How it worked | Landmark | Source |
|---|---|---|---|---|
| Rules | 1950s to 1980s | People wrote grammar and word rules by hand | Georgetown demo 1954; ELIZA 1964 to 1967 | Wikipedia |
| Statistics | 1990s | Programs learned patterns from large bodies of text | IBM team, Brown, Jelinek, Mercer and others, 1990 | Wikipedia |
| Neural | 2003 to 2018 | Networks learned word meaning from data | Bengio et al. 2003; transformer 2017; BERT 2018 | Wikipedia, arXiv |
| Large language models | 2020 on | One huge pretrained model does many tasks | GPT-3, 175 billion parameters | arXiv |
The rule era ended with a report. In 1966 the ALPAC report found that a decade of machine translation research had fallen short of expectations, and Wikipedia notes that funding was cut sharply afterward. The statistical era began when researchers stopped writing rules and counted instead. The 1990 IBM paper trained on English and French text from the Canadian Parliament.
Each jump cost more compute and needed more text. That's the pattern to remember. Rules needed linguists, statistics needed corpora, and neural models needed both data and chips. GPT-3's 175 billion parameters were possible only because the earlier steps had already shown that more data beat more hand work.
The neural era brought the design that runs nearly everything now. The transformer paper, posted on 12 June 2017, replaced older recurrent designs. What is a transformer in AI explains how it works. BERT followed in October 2018 and GPT-3 in May 2020. When was AI invented puts these dates inside the wider story.
How does NLP work today?
Modern NLP cuts text into tokens, turns each token into numbers, and lets a neural network predict what should come next or what label fits. What are tokens in AI covers the unit. The network learned its patterns from a very large body of text before you ever typed a word.
A large language model is the current best tool for the job, not the whole field. What is an LLM goes through how one is trained. NLP also covers small, narrow systems that run on a phone, such as a spell checker or a keyword spotter for voice commands.
So the two words answer different questions. NLP names the problem: make computers handle language. An LLM names one way of solving it. ChatGPT, Claude and Gemini are NLP products built on LLMs.
You use this every day without a label on it. A search box that understands a question, a phone that turns speech into a text message and a mail app that moves junk out of your inbox are all NLP. Stanford's list of examples names translation, sentiment analysis, chatbots, voice assistants, summarization, spell checking and information extraction from documents. None of them needs a human to read each item.
What changed in 2017 is how little each system needs built by hand. Before, a team assembled a separate tool for each task. Now a team can start from a pretrained model and adapt it. That's why language features now show up in ordinary apps.
What are real examples of NLP?
Three dated cases show the range, from a stage demo to a model that translates at scale. Each one comes from a primary paper or a reference page, and the numbers are the ones those pages report.
- The 1954 Georgetown demo. On 7 January 1954 a Georgetown and IBM team translated more than sixty Russian sentences into English. The system held 250 vocabulary items and six grammar rules. It looked like a breakthrough, and Wikipedia reports that the field then spent a decade falling short of the hope it raised.
- ELIZA in 1966. Joseph Weizenbaum at MIT built a program that scanned for ranked keywords and rebuilt a reply from a script. Its best known script, DOCTOR, imitated a therapist by turning your words back as questions. Wikipedia describes the method as pattern matching and substitution that gave an illusion of understanding without any. It is the classic reminder that sounding fluent is not understanding.
- The 2017 transformer on translation. The paper reported 28.4 BLEU on WMT 2014 English to German and 41.8 on English to French, the French model after training for 3.5 days on eight GPUs. BLEU scores a translation against human references, so higher is better. The score isn't a measure of meaning, but it showed one design could beat the older systems at the task NLP began with.
What questions should you ask an AI about NLP?
An AI model is a good tutor for NLP if you make it show its work. Paste any of these into Claude, ChatGPT, Gemini or Whizi and fill the brackets.
- Explain at my level: "Explain [NLP task, such as named entity recognition] to a [grade or job] who knows [prior skill]. Give one everyday example, then ask me one question to check I understood."
- Test on my text: "Label the sentiment of each sentence below as positive, negative or neutral, and give a one clause reason for each. Flag any sentence where you are unsure: [paste 6 sentences]."
- Find the entities: "List every person, place, company and date in this text as a table with a column for the type. Mark anything that could be read two ways: [paste text]."
- Old versus new: "Show how a rule based system and a transformer based model would each handle [task]. Say which mistakes each one makes."
- Pick a tool for a job: "I need to [goal, such as sort 5,000 support emails by topic]. Would I use NLP, a large language model or both? Give the steps and the main risk."
- Check a claim: "Is this claim about NLP true: [claim]? Search the web, give the source and its date, and say what would make you wrong."
Where does Whizi fit when you are learning NLP?
Whizi is one app instead of a pile of single purpose tools. You can ask Claude, GPT and Gemini the same question about NLP and see where they differ. Side by side Compare, which shows two models on one prompt, is a Pro feature. Web search is on by default in each chat.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. See Whizi pricing for what each plan holds.
Where does NLP still get language wrong?
NLP still fails on ambiguity, on sarcasm and on facts. The sentence "The bank was closed" has two readings until a model sees the lines around it. A sentiment tool can score a sarcastic review as praise because every word in it is positive.
Fluent output also hides invention. A generation model writes the most likely next words, not the true ones, which is why what is an AI hallucination matters for any NLP product that writes. ELIZA showed the trap in 1966: a program can sound attentive and understand nothing.
For anything with money, health or law in it, treat an NLP output as a draft. The 1966 ALPAC report is the oldest warning in the field about promising more than the method delivers.
Language coverage is uneven too. A system trained mostly on English text does worse on a language with less written data, and slang or dialect can trip a model that handles formal writing well. When you test a tool, test it on your own kind of text, not on the demo sentence the vendor chose.
Finally, a good score isn't a good product. BLEU and GLUE rank systems on fixed test sets. They don't tell you whether the tool will cope with your customers' typos, your industry's jargon or your team's tone.
- Say which NLP you mean: the AI field, or the neuro-linguistic self help method
- Name the task before you pick a tool: classify, extract, translate, summarize or answer
- Read an NLP number as a score on one test, not a measure of understanding
- Check fluent output against a source when a fact or a figure matters
- Try two models on the same text and read only where they disagree
Frequently asked questions
Is natural language processing a type of AI?
Yes. Stanford's Institute for Human-Centered AI calls NLP a branch of artificial intelligence, and it draws on computational linguistics, machine learning and deep learning. A spell checker built from fixed rules is NLP too, though few call it AI today.
Is ChatGPT natural language processing?
Yes. ChatGPT is an NLP product built on a large language model, and it does both halves of the field: it understands your prompt and generates a reply. NLP is the field, the LLM is the engine, and the chat window is the product.
What is the difference between NLP and an LLM?
NLP is the whole field of making computers handle language, and an LLM is one tool inside it. The field started with hand written rules in the 1950s. LLMs such as GPT-3, with 175 billion parameters, arrived in 2020.
What is NLP used for?
NLP is used for translation, sentiment analysis, chatbots, voice assistants, summarization, spell checking and pulling structured information out of documents. Those are the applications Stanford's Institute for Human-Centered AI lists. Search boxes and spam filters rely on it too.
What does NLP stand for?
In AI and computing it stands for natural language processing. In self help and coaching it stands for neuro-linguistic programming, a 1975 method that Wikipedia describes as pseudoscientific. The two share only the letters.