What is natural language processing (NLP)?

In plain English

“Natural language” just means the languages people speak and write, as opposed to programming languages. Natural language processing, or NLP, is the long-running branch of AI that tries to get computers to work with it: spotting spam, translating between languages, turning speech into text, finding names and dates in documents, judging whether a review is positive or negative, and, more recently, holding a conversation.

For decades NLP used hand-written grammar rules and then smaller statistical models, each built for one job. Since around 2018, large language models built on the transformer have taken over most of these tasks, because one big model can do many of them well.

Why it matters

NLP is in the tools you use every day: autocorrect and predictive text, email spam filters, voice assistants, machine translation, search engines that understand your question, and chatbots like ChatGPT. When people talk about “AI” today, they very often mean a form of NLP.

An example

You ask a voice assistant, “Remind me to call Mum when I get home.” Several NLP steps happen in a second: speech recognition turns your voice into text, the system works out that this is a reminder, picks out “call Mum” as the task and “when I get home” as a location trigger, then confirms in a natural-sounding voice.

What to watch out for

Language is full of sarcasm, slang, local dialects and ambiguity, and NLP systems still trip over them. Most of the text they learn from is in English and a handful of other big languages, so they often work noticeably worse in smaller languages, and they can absorb the stereotypes in the text they are trained on.