Natural language processing (NLP) is the field of computer science and AI that builds software to read, interpret, and produce human language, in text or speech. It covers tasks such as classifying a message's intent, extracting names and order numbers, detecting sentiment, translating, summarizing, and generating replies.
NLP is the parent field. Large language models are its most recent and most capable tool, but the field is decades older and the tasks it defined still structure how customer service software is built: work out what the customer wants, pull out the details, judge the mood, and respond.
What changed with LLMs is the method. Older systems needed a separate, hand-built component for each task. A modern model can do most of them from an instruction and a few examples.5
From rules to statistics to language models
NLP has gone through four broad approaches. Each one is still in use somewhere, often inside the same product.
Rules
Knowledge comes from: linguists and developersHand-written patterns and grammars. If the message contains "refund" and "order," route it to billing. Predictable and brittle: a synonym or a typo breaks it.
Statistical NLP
Knowledge comes from: labeled examplesN-gram language models and classifiers such as naive Bayes and logistic regression learn from counts in data instead of rules.1 Better coverage, but each task needs its own labeled dataset.
Neural embeddings
Knowledge comes from: large unlabeled textWords become vectors that capture meaning, so "parcel" and "package" sit close together. Word2vec learned such vectors from 1.6 billion words in under a day.2
Pretrained transformers
Knowledge comes from: pretraining, then fine-tuningThe transformer architecture relies on attention alone.3 Models such as BERT are pretrained once and fine-tuned per task with one extra output layer.4
Large language modelsToday
Knowledge comes from: pretraining at scale, instructionsOne model handles many tasks from a text instruction and a few examples, without retraining.5
What NLP extracts from one message
A customer writes to an electronics retailer. Before any reply is drafted, NLP turns the free text into structured data the support system can act on.
Extracted: intent damaged item, replacement · entities product: laptop, order: 48213 · sentiment negative, repeat issue · constraint no refund
{
"intent": "damaged_item.replacement",
"entities": {"order_id": "48213", "product": "laptop", "damage": "cracked screen"},
"sentiment": "negative",
"signals": ["repeat_issue", "declines_refund"],
"route": "returns_agent",
"priority": "high"
}Illustrative example of structured output. Field names are placeholders; each team defines its own schema.
NLP tasks that matter in customer service
Intent recognition
Classifying what the customer wants: order status, return, cancellation, complaint. See intent recognition. It decides routing and which agent or workflow takes over.
Entity extraction
Pulling out order numbers, product names, dates, amounts, and addresses. See named entity recognition. Entities become the inputs to tool calls.
Sentiment analysis
Judging tone and frustration. See sentiment analysis. Used to prioritize, to adjust replies, and to hand angry customers to a person sooner.
Summarization
Condensing a long conversation into a handover note or a ticket summary, so a human agent does not have to reread the whole thread.
Generation
Writing the reply itself, in the brand's tone and the customer's language. Modern agents generate from retrieved policies rather than from templates.
Translation
Serving customers in several languages from one knowledge base, which matters for brands across Poland, CEE, Germany, and the UK.
NLP, NLU, LLMs, and conversational AI
These terms are often used as synonyms in vendor material. They sit at different levels.
| Term | What it names | Relationship to NLP |
|---|---|---|
| Natural language processing | The whole field of making software work with language | The parent field |
| Natural language understanding (NLU) | The interpretation side: intent, entities, meaning | A subfield of NLP |
| Natural language generation (NLG) | The production side: writing text | A subfield of NLP |
| Large language model | A pretrained model that does many NLP tasks from instructions5 | Today's main NLP tool |
| Conversational AI | Products that hold a dialogue: chatbots, voice agents | An application built with NLP |
| Machine learning | Learning patterns from data, for any kind of data | The broader method most modern NLP uses |
Where NLP shows up in a support operation
Most of these run before or after the conversation, not only inside it.
Triage every message
A telecom classifies incoming messages by intent and urgency, extracts the account number, and sends each to the right agent or queue.
message → intent + entities → routeRead documents
An insurer extracts claim numbers, dates, and amounts from emails and attachments and fills the claims system.
email → extract → post to systemAnalyze every conversation
A marketplace tags all conversations by topic and sentiment to find what drives contacts, instead of reading a small sample.
transcripts → tag → report driversUnderstand what shoppers mean
A retailer's shopping agent maps "something for a small balcony, under 300 złoty" to catalog filters for size and price.
need → entities → search catalogLimits and the usual controls
Language and dialect gaps
ControlQuality varies by language. Test intent and entity accuracy in each language you serve, on real messages, before launch.
Misread intent
ControlAsk a clarifying question when confidence is low instead of guessing. Track misroutes as a metric.
Sarcasm and context
ControlSentiment models misjudge irony and short replies. Use sentiment to prioritize, not to make decisions on its own.
Personal data in text
ControlFree text contains names, addresses, and card numbers. Detect and mask personal data before storage or analysis.
Common questions
Is NLP the same as AI?
No. NLP is one field within AI, focused on language. AI also covers vision, robotics, planning, and more.
Is ChatGPT NLP?
Yes. ChatGPT is built on a large language model, which is the current main technique in NLP. It performs NLP tasks such as understanding, summarizing, translating, and generating text.
What is the difference between NLP and NLU?
NLU, natural language understanding, is the part of NLP concerned with interpreting meaning, such as intent and entities. NLP also includes generating language, translation, and speech.
Do support teams still need separate NLP models if they use an LLM?
Sometimes. An LLM can classify intent and extract entities from instructions, but a small dedicated classifier or a rule can be cheaper, faster, and easier to audit for high-volume, stable tasks.
References
- Jurafsky, D. and Martin, J. H. Speech and Language Processing, 3rd edition draft. Stanford University, 2026.
- Mikolov, T. et al. Efficient Estimation of Word Representations in Vector Space. arXiv, 2013.
- Vaswani, A. et al. Attention Is All You Need. NeurIPS 2017.
- Devlin, J. et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL 2019.
- Brown, T. B. et al. Language Models are Few-Shot Learners. NeurIPS 2020.