The future of natural language processing (NLP)

NLP has come a long way since its beginnings. As the amount of unstructured data grows, it keeps getting better at the nuance, context and ambiguity of human language. It has already given rise to voice assistants, chatbots, translators and many other tools that we use every day.

When this post first appeared in September 2021, the field was on the eve of its biggest change. ChatGPT launched on November 30, 2022, and in October 2025 OpenAI said it had more than 800 million weekly users. The applications below are still the core of NLP, but most of them now run on the same kind of model.

How does natural language processing work?

NLP went through three approaches. The first systems used hand-written rules and dictionaries, which broke on anything their authors had not foreseen. Statistical methods followed, assigning probabilities to possible meanings; they powered early spell checkers and machine translation. Today deep learning dominates: neural networks trained on huge amounts of text and speech.

A modern system works in a few steps. It splits the text into tokens, words or pieces of words that the model works with. It turns each token into an embedding, a list of numbers that places words with similar meanings close together. A transformer, the architecture introduced by Google researchers in 2017, then weighs every token against the others to capture context, so that “bank” in “the bank approved the loan” means a lender, not a river bank. Finally the model produces an output: a label, an answer, a summary or a translation.

Large language models are one part of NLP, not a synonym for it: routing support tickets or tagging sentiment is often done more cheaply by small, specialized models.

Natural language processing applications

Natural Language Understanding (NLU)

NLU goes beyond processing words as data: it is designed for computers to understand what a person is saying. Instead of turning words into codes, NLU takes human syntax, sentences and structure and makes decisions based on the way the person speaks.

Sentiment analysis

Sentiment analysis takes into account not only what was said but also the emotions behind it. It is especially useful for B2C businesses, for example to follow social media discussions about a product.

Machine translation

Machine translation is the process of using a machine to convert text from one language to another, for example in a translation app. Older apps translated word for word, and the structure of the sentence or its context could get lost in translation. Neural models translate whole sentences in context, and coverage has grown fast: in July 2022 Meta released NLLB-200, a single model that translates between 200 languages, many of them with little training data.

Chatbots

If you’ve been browsing or shopping online, you’ve probably interacted with chatbots. Chatbots are customer service programs that use natural language processing to respond to your request in real time. In 2021 most followed scripted flows; many now run on large language models that hold an open conversation and hand over to a person when needed.

Email assistant

Autocomplete, spell check and autocorrect are examples of NLP features that help you write emails. Email spam filters use natural language processing too, to decide whether a message is spam or worth keeping. Newer assistants also draft replies and summarize long threads.

Semantic parsing

Just as humans subconsciously analyze words to understand their meaning, machines do the same. Semantic parsing is the ability to transform natural language into something that a machine can understand on its own terms, such as a database query or a command.

Applications of NLP in Data Science

NLP has many applications in data science, which then carry over to other fields, especially in terms of business value.

Speech recognition

Behind speech recognition is NLP. By analyzing speech patterns, meaning, relationships and word classes, the algorithm assembles what it hears into complete sentences. With deep learning, a system can also learn to recognize an accent or a speech disorder and become more accurate. OpenAI’s Whisper, released in September 2022, was trained on 680,000 hours of multilingual audio from the web. The technology also lets people with disabilities communicate with machines much more easily.

Market analysis

NLP lets companies spot trends in large amounts of data. Using topic classification, the machine finds the most common categories. For example, social media analytics can provide information about your industry, product or brand straight from the consumers’ point of view, improving your business intelligence.

Prediction of diseases

NLP is widely used in healthcare to predict possible diseases. By interpreting speech patterns, NLP algorithms can give clinicians information about progressive conditions such as depression or schizophrenia. Medical records are another huge source of information, and practitioners use NLP to detect disease, understand patients better, support care delivery and cut costs.

Search results

Using NLP, search engines can determine the intent of any query. Google uses this technology to give you the best possible results. With the introduction of BERT in 2019, Google greatly improved its understanding of intent and context. This matters for voice search, as spoken queries are more conversational. When it announced BERT, Google noted that as many as 15% of the queries it sees every day are new, so it cannot rely on matching phrases it has seen before. Since 2024 Google also answers many queries itself, with AI Overviews and, since May 2025, the conversational AI Mode.

The trends we watched in 2021

These are the trends this post expected in 2021, and how each of them turned out.

Social media sentiment analysis

Social media generate a huge amount of data at any moment, too much to read manually. Used for sentiment analysis, NLP helps companies quickly understand how customers feel about their brand and how urgent a conversation is.

In 2021 NLP still struggled with the different meanings of a word, with context, and with sarcasm, irony and contradictory statements. Large language models read context far better than the classifiers of 2021, so mixed opinions and sarcasm are easier to catch, but irony, slang and dialects still cause errors.

Multilingual NLP

In 2021 most advances in NLP focused on English, and companies such as Google and Facebook, now Meta, were introducing multilingual models whose performance came close to monolingual ones. That trend held: today’s large models work in dozens of languages, and Meta’s NLLB-200 showed one model could cover 200.

The gap has not closed completely. A NeurIPS 2023 study found that the same text can need up to 15 times more tokens in one language than in another, which makes the same work slower and more expensive.

Transfer Learning

Transfer learning is a machine learning technique in which a model is trained for one task and then reused for another, similar task. Instead of training a new model from scratch, you adapt an existing one.

It saves time and resources, and in 2021 we expected it to become a major NLP trend. It became the foundation of the whole field: every large language model is pre-trained once on a huge amount of text and then adapted to specific jobs by fine-tuning or simply by instructions.

Collaboration between supervised and unsupervised learning

Supervised learning maps inputs to outputs based on example input-output pairs. Unsupervised learning works without such pairs: the model finds patterns on its own.

Supervised learning is precise and helpful in classification problems, but labeled data is expensive; unsupervised learning needs no labels, but cannot map inputs to known outcomes. In 2021 NLP providers were combining the two.

The combination won in the form of self-supervised learning. Language models learn from raw text by predicting the next word, which needs no human labels, and are then tuned with smaller sets of supervised examples and human feedback.

Detection of fake news and cyber violence

NLP has become an essential tool for reducing the time and effort needed to detect and stop the spread of fake news and misinformation. The flood of fake news about Covid-19 in 2020 and 2021 brought new approaches to automated detection.

NLP is also used to detect cyber violence: classifiers flag offensive language and hate speech on social media.

Generative AI has made the problem harder, because convincing fake text is now cheap, and regulation has followed. The EU AI Act has applied rules for general-purpose AI models since August 2025, and from August 2026 people must be told when they talk to a chatbot and AI-generated content must be identifiable.

What changed after 2021: large language models

The transformer of 2017 and the models built on it, BERT in 2018 and GPT-3 in 2020, were known when this post was written. What few expected was how quickly they would become general tools. Three numbers from Stanford’s AI Index 2025 show the pace:

• The cost of running a model as good as GPT-3.5 fell more than 280-fold between November 2022 and October 2024.
• In 2022 the smallest model scoring above 60% on the MMLU knowledge test was PaLM, with 540 billion parameters. In 2024 Microsoft’s Phi-3-mini did it with 3.8 billion, 142 times fewer.
• 78% of organizations reported using AI in 2024, up from 55% a year earlier.

NLP trends for 2026 and beyond

The trends that will shape NLP next follow from that shift:

• Agents. Models that plan, call tools and carry out multi-step tasks, so language becomes the interface to software. In companies this shows up as agentic process automation in accounting, HR and customer service.
• Small and on-device models. Cheaper, faster models for one job, running on a phone or a machine on the factory floor, without sending data to the cloud.
• Retrieval-augmented generation (RAG). The model looks up trusted documents before it answers, which makes answers more current and easier to check.
• Multimodal models. Text, speech, images and video handled by one system, so a model can read a scanned contract or answer a question about a photo.

Is NLP reliable?

For well-defined tasks such as classification, extraction, translation of common languages or transcription of clear speech, modern NLP is reliable enough to run in production. Generative models are a different case: they can produce fluent, confident text that is wrong. In 2023 a US federal judge fined two lawyers $5,000 after they filed a brief with court decisions that ChatGPT had made up. The AI Index 2025 also notes that complex reasoning remains a challenge even for the strongest models.

The practical rule is to match the checks to the stakes: ground answers in your own documents, test the model on your own data and keep a person in the loop wherever an error is costly.

Will AI replace language interpreters?

Official projections do not expect that. The US Bureau of Labor Statistics projects that employment of interpreters and translators will grow 2% from 2025 to 2035, with about 6,000 openings a year. It notes that AI tools make translators more efficient, but that many of these jobs cannot be fully automated, because machines cannot yet match human work in most cases. The work is shifting toward reviewing machine output, with human interpreters where the stakes are high.

The future of natural language processing

With more information available and ever more accurate algorithms, NLP keeps growing and changes the way people and machines interact.

Most of the information around us is unstructured text: emails, contracts, reports, chats and reviews. That is why NLP is one of the largest fields of data science.

Natural language processing (NLP) is evolving rapidly, and so are its applications. With so much data available, it is important to understand it, monitor it and, in some cases, moderate it.

In 2021 we expected NLP to spread through ready-to-use, pre-trained models. That happened faster than predicted. The next question is less whether machines can handle language and more where companies trust them to act on it.