Natural language processing (NLP) is the field that teaches computers to process what people write or say. Almost all NLP today relies on machine learning. Typical tasks range from classifying text and recognising names to translating and summarising.
Natural language processing (NLP) teaches computers to handle text and speech: tokens, embeddings, typical tasks and its role in search.
What natural language processing means
Google's machine learning glossary defines NLP as the field of teaching computers to process what a user said or typed using linguistic rules. The next sentence adds that almost all modern natural language processing relies on machine learning.
Google also separates out natural language understanding (NLU), a subset of NLP that determines the intentions behind something said or typed. According to the glossary, NLU can go further and consider context, sarcasm and sentiment.
NLP is a field, not a type of model. Google defines a language model as a model that estimates the probability of a token or sequence of tokens occurring in a longer sequence. An LLM is, at a minimum, a language model with a very high number of parameters. LLMs are therefore one family of models within NLP. Many tasks in the field were handled before they existed.
How it works
The first step is tokenisation. Dan Jurafsky and James H. Martin describe it in their textbook "Speech and Language Processing" as the first stage of NLP: the input text is segmented into tokens. A token can be a word, a character or part of a word. Modern tokenisers learn their own sets of tokens, including units smaller than words, called subword tokens. According to the textbook, the two algorithms most widely used in current language models are byte-pair encoding (BPE) and unigram language modelling. An illustrative example, not the output of any particular tokeniser: "unhelpfulness" might be split into "un", "help", "ful" and "ness".
Next, tokens are turned into numbers. In 2003, Yoshua Bengio and colleagues proposed learning a distributed representation for each word, so that one training sentence tells the model something about many sentences with a similar meaning. In 2013, Tomas Mikolov and his team presented architectures that compute continuous word vectors from very large data sets. These vectors are embeddings: lists of numbers the model uses to compare meanings.
With that representation, NLP solves concrete tasks. Jurafsky and Martin write that classification lies at the heart of language processing; spam detection and sentiment analysis are examples. Named entity recognition assigns words or phrases tags such as PERSON, LOCATION or ORGANIZATION. Machine translation is, according to the textbook, probably one of the most common uses of NLP technology. The same book lists summarisation as a task without a single right answer.
Getting here took three phases. According to Jurafsky and Martin, symbolic, rule-based approaches dominated roughly from 1965 until the early 1990s. One example is ELIZA, which Joseph Weizenbaum presented in 1966: the program analysed sentences with decomposition rules triggered by key words. The field then moved back towards statistical and neural models. IBM researchers presented a statistical approach to machine translation in 1990, and Bengio's neural language model followed in 2003. In 2017, Ashish Vaswani and colleagues introduced the Transformer, a neural network architecture that, the textbook notes, was invented for machine translation and quickly became a general-purpose architecture.
From text to task
Text“unhelpfulness”Illustrative example: the word might be split like this, not the output of any particular tokeniser
→
Tokensunhelpfulnessa word, a character or part of a word
→
Embeddingslists of numbers the model uses to compare meaningsSchematic, no real values
→
Modeltoday almost always based on machine learning
Tasks built on top
Classificationspam detection, sentiment analysis
Named entity recognitionPERSON · LOCATION · ORGANIZATION
Machine translationone of the most common uses
Summarisationwithout a single right answer
Before a model sorts anything, the text has already become numbers.
Why it matters
Google Search uses NLP. On 25 October 2019, Pandu Nayak announced on Google's blog that the company was applying BERT to ranking and featured snippets. Google describes BERT as a neural network-based technique for NLP pre-training. According to the announcement, BERT would help Search better understand one in 10 searches in the U.S. in English. The figure dates from 2019 and covers only that market and that language. In its guide to Search ranking systems, updated on 10 December 2025, Google describes BERT as an AI system that allows it to understand how combinations of words express different meanings and intent. The same guide says MUM can both understand and generate language; Google uses it for specific applications and not for general ranking.
Google does not say how much weight these systems carry in ranking. What can be read from its statements is the purpose: the search engine interprets the combination of words in a query, not only the individual words.
In marketing, NLP sits inside many tools: sentiment analysis of reviews, sorting of customer enquiries, chatbots and machine translation. Google's glossary counts context and sarcasm among the complex aspects of language, and that is where these tools are worth testing.
Best practices
Keep the terms apart: NLP is the field, a language model is a type of model, and an LLM is a very large language model.
Test a sentiment tool on a sample of your own, checked by hand, before trusting its numbers.
Check machine translations in the target language before publishing, especially product names and legal terms.
Write pages as answers to the complete question, not for a list of keyword variants.
Common mistakes
Treating NLP and LLM as synonyms.
Reading the 2019 BERT figure as a current value or as valid for other languages.
Assuming a tool handles irony reliably because it gets simple sentences right.
Trying to "optimise for BERT", although neither the 2019 announcement nor the ranking systems guide describes any specific optimisation for that system.
Believing NLP understands text the way a person does. It computes with tokens and vectors.
Manuel Riveiro RodriguezCEO & Digital Strategist
A technical audit covers this and everything else in one pass.
NLP is the field that teaches computers to process what a person writes or says. Google defines it this way in its machine learning glossary and adds that almost all modern NLP relies on machine learning. Typical tasks are classifying text, recognising the names of people or companies, translating and summarising.
What is the difference between NLP and an LLM?
NLP is the field, and an LLM is a type of model within it. Google defines a language model as a model that estimates the probability of a token within a sequence, and an LLM as a language model with a very high number of parameters. Many NLP tasks were handled before LLMs existed, for example with statistical models.
What is tokenisation?
Tokenisation splits a text into units called tokens. Jurafsky and Martin describe it as the first stage of natural language processing. A token can be a word, a character or part of a word. Current language models mostly use algorithms such as byte-pair encoding, which learn by themselves which parts of words are worth splitting off.
Does Google use natural language processing?
Yes. In October 2019, Google announced that it was applying BERT, a neural network-based NLP technique, to ranking and featured snippets. In its guide to ranking systems, Google describes BERT as an AI system for understanding how combinations of words express meanings and intent. Google does not say how much weight BERT carries in ranking.
How long has natural language processing existed?
According to Jurafsky and Martin, rule-based approaches dominated from the mid-1960s. An early example is ELIZA by Joseph Weizenbaum, published in 1966. Statistical models followed, such as IBM's approach to machine translation in 1990, then neural language models from 2003 and, in 2017, the Transformer that today's LLMs build on.
