Table of Contents
- Key takeaways
- What is an NLP chatbot?
- How does NLP work in a chatbot?
- Intents: what the user wants
- Entities: the details inside the message
- Context: what was already said
- Retrieval: where the facts come from
- NLP vs NLU vs NLG: what is the difference?
- What are the types of NLP chatbots?
- Intents and entities vs LLMs: which NLP approach should you use?
- Worked example: one message through three kinds of chatbot
- What NLP tasks does a chatbot use?
- What is an example of an NLP chatbot?
- How do you build an NLP chatbot?
- Common NLP chatbot problems and how to fix them
- Frequently asked questions
- Basics
- What is NLP in a chatbot?
- What is an NLP chatbot?
- How does NLP work in a chatbot?
- Comparisons
- What is the difference between NLP and NLU?
- What is the difference between NLP and an LLM?
- What is the difference between a rule-based chatbot and an NLP chatbot?
- What is the difference between machine learning and NLP in a chatbot?
- Building
- What are the best frameworks for an NLP chatbot?
- How many training phrases does an intent need?
- Does an LLM chatbot still need intents?
- Accuracy and limits
- Why does an NLP chatbot give wrong answers?
- Can an NLP chatbot understand different languages?
- Can an NLP chatbot detect emotions?
- Sources
Created by
Do not index
Created time
Sep 27, 2023 11:13 PM
An NLP chatbot is a chatbot that uses natural language processing (NLP) to read a message written in everyday language, work out what the person wants, and reply in natural language. Older NLP chatbots do this with intent classification and entity extraction. Newer ones use a large language model (LLM) that answers from the business's own content through retrieval-augmented generation (RAG).
TL;DR
- NLP is the part of a chatbot that turns free text into meaning, and meaning back into a reply.
- Intent-based chatbots match each message to a list of intents, such as "track order", and pull out entities, such as an order number or a date. Google's Dialogflow recommends 10-20 example phrases per intent.
- LLM chatbots skip the intent list. The model reads any phrasing, and retrieval supplies the facts from your pages and documents.
- Use intents and scripted flows for fixed tasks like refunds or bookings. Use an LLM with retrieval for open questions about products, policies, and how-to steps.
- When an LLM chatbot gets an answer wrong, check the source content first. Missing or outdated pages are a common cause.
Key takeaways
Approach | How it understands a message | Best for | Main risk |
Keyword matching | Looks for fixed words such as "refund" | A handful of menu-style questions | Misreads any message that uses other words |
Intent-based NLU | Classifies the message into a known intent and extracts entities | Fixed tasks with a few outcomes | Falls back on every topic nobody scripted |
LLM only | A language model reads the message and writes a reply from what it learned in training | General conversation | Invents business facts it was never given |
LLM with retrieval (RAG) | The model reads the message, retrieves passages from your content, and answers from them | Open questions about your products and policies | Repeats outdated or missing content |
What is an NLP chatbot?
An NLP chatbot is a chatbot that understands free text instead of only buttons or exact commands. IBM defines NLP as a subfield of computer science and AI "that uses machine learning to enable computers to understand and communicate with human language."
In a chatbot, NLP does two jobs. It reads the visitor's message and turns it into something the system can act on. Then it turns the system's answer back into a readable sentence.
A chatbot without NLP can only follow a menu or match exact keywords. A chatbot with NLP can handle "where's my parcel?", "has my order shipped yet", and "tracking pls" as the same request.
NLP chatbots are one kind of conversational AI. The guide to conversational AI examples shows how voice assistants, banking assistants, and customer service bots use the same ideas.
How does NLP work in a chatbot?
NLP works in a chatbot as a pipeline of five steps: read the message, work out the meaning, track the conversation, find the facts, and write the reply. Intent-based chatbots run each step as a separate component. LLM chatbots handle most of them inside one model.
- Read the message. The text is split into tokens, which are words or word pieces, and the language is detected.
- Work out the meaning. Intent classification names the goal of the message. Entity extraction pulls out the details, such as a product, a date, or an order number.
- Track the conversation. Dialogue state keeps what was already said, so a follow-up such as "what about the red one?" still makes sense.
- Find the facts. A scripted chatbot runs the flow tied to the intent. An LLM chatbot with retrieval searches the business's content for passages that answer the question.
- Write the reply. Natural language generation turns the result into a sentence, either from a template a person wrote or from a language model.
Intents: what the user wants
An intent is the goal behind a message, such as "order pizza" or "track order". Amazon's Lex documentation describes an intent as an action the user wants to perform, with sample utterances that show how a user might express it.
The team writes example phrases for every intent, and a model learns to match new messages to the closest one. Google's Dialogflow documentation says to create at least 10-20 training phrases per intent, depending on how complex the intent is. Its machine learning then expands the list with similar phrases.
Entities: the details inside the message
An entity is a piece of structured data inside a message, such as a date, a city, or a product name. Dialogflow's example is the phrase "What is the forecast tomorrow for Tokyo?", where "tomorrow" becomes a date parameter and "Tokyo" becomes a location parameter.
Frameworks ship system entities for common types such as dates, times, and email addresses. Teams add custom entities for their own terms, such as product names. Amazon Lex calls the required details for an intent "slots" and prompts the user until every required slot has a value.
Context: what was already said
Context is the record of earlier turns in the same conversation. Without it, a chatbot asks again for details the visitor already gave.
Intent frameworks store context as variables and active states. An LLM chatbot passes the recent conversation back to the model with each new message, so the model reads the whole exchange.
Retrieval: where the facts come from
Retrieval is the step that finds the right passage in the business's own content before the model writes an answer. The 2020 paper by Lewis and colleagues that named retrieval-augmented generation describes models that combine a language model with an external store of documents.
Retrieval matters because a language model only knows what it saw in training. It has never seen a company's returns policy or price list. The guide to what RAG is explains the retrieval loop in more detail.
NLP vs NLU vs NLG: what is the difference?
NLP is the whole field of processing human language. Natural language understanding (NLU) is the part that reads input and extracts meaning. Natural language generation (NLG) is the part that writes output.
Term | What it covers | In a chatbot | Example |
NLP (natural language processing) | The full field: reading, understanding, and producing language | Every language step in the pipeline | Tokenizing a message, detecting its language |
NLU (natural language understanding) | Turning text into meaning | Intent classification and entity extraction | "Can I send back my jacket?" becomes intent: return, product: jacket |
NLG (natural language generation) | Turning meaning into text | Writing the reply | "Yes. Returns are accepted within 30 days of delivery." |
In an intent-based chatbot, NLU and NLG are separate parts. A classifier does the understanding and a template does the writing. In an LLM chatbot, one model does both.
What are the types of NLP chatbots?
There are four types of NLP chatbots: keyword-matching bots, intent-based NLU bots, LLM-only bots, and LLM bots with retrieval. They differ in how they read a message and where their answers come from.
Type | Understanding | Answers come from | Setup work | Handles new phrasing |
Keyword matching | Fixed words and rules | Canned replies | Write rules for every keyword | Poorly |
Intent-based NLU | Trained intent classifier plus entities | Templates and scripted flows | 10-20 example phrases per intent, plus flows | Within each intent |
LLM only | Pretrained language model | The model's training data | Write instructions | Well |
LLM with retrieval | Pretrained language model | Passages retrieved from your content | Connect content and write instructions | Well |
Most current large language models are transformers. The 2017 paper that introduced the transformer described an architecture "based solely on attention mechanisms". IBM explains that transformers use tokenization and self-attention to work out how the parts of a sentence relate to each other.
For a ranked list of tools in this category, see the best conversational AI chatbots.
Intents and entities vs LLMs: which NLP approach should you use?
Use intents and scripted flows for fixed tasks that must run the same way every time. Use an LLM with retrieval for open questions where people phrase things in many ways. Many chatbots use both: the LLM answers questions, and an intent or a button starts a scripted flow for transactions.
If you need... | Pick... | Why |
Answers to open questions about products, policies, or how-to steps | LLM with retrieval | Nobody can script every phrasing, and the facts live in your content |
A booking, refund, or account change that must follow fixed steps | Intent-based flow | Every step is predictable and auditable |
A bot that stays current when pages change weekly | LLM with retrieval | Updating content updates answers, with no retraining |
Structured data from a message, such as a date or order number | Entity extraction | Entities give the system clean values to act on |
Small talk with no business facts | LLM only | No retrieval needed when nothing factual is at stake |
A question outside the trained content | Neither. Decline and offer a person | Guessing is how chatbots give wrong answers |
The machine learning side of these choices, including how to score an intent classifier, is covered in the guide to a customer service chatbot using machine learning.
Worked example: one message through three kinds of chatbot
This example follows one customer message through three chatbots to show what NLP does at each step. The store and its policy are hypothetical.
The message, sent on March 20: "Hi, I bought the blue rain jacket on March 3. Can I still send it back?"
The store's returns page says: "Items can be returned within 30 days of delivery."
Chatbot type | What it understands | What it replies | Result |
Keyword matching | No keyword "return" or "refund" appears, because the customer wrote "send it back" | "Sorry, I didn't understand. Choose an option from the menu." | Fails |
Intent-based NLU | Intent: return request. Entities: product = blue rain jacket, date = March 3 | Starts the returns flow and asks for the order number | Works, but only because someone built a returns intent and flow |
LLM with retrieval | Reads the question, retrieves the returns policy passage | "Returns are accepted within 30 days of delivery. If the jacket arrived on or after March 3, at least 13 days are left. Here is the returns page." | Works, and links the source |
The math behind the LLM reply: March 3 to March 20 is 17 days. The window is 30 days, so 30 - 17 = 13 days remain if the jacket was delivered on March 3. A later delivery date leaves more time.
The LLM reply also shows a good habit. The policy counts from delivery, and the customer gave a purchase date. A grounded chatbot states the rule it found instead of guessing the delivery date.
What NLP tasks does a chatbot use?
A chatbot uses six common NLP tasks: tokenization, intent classification, entity recognition, sentiment analysis, language detection, and question answering. Not every chatbot runs all six.
NLP task | What it does | Example in a chatbot |
Tokenization | Splits text into words or word pieces | Each word, or piece of a word, becomes a token the model can process |
Intent classification | Names the goal of the message | "Where's my parcel?" maps to track order |
Named entity recognition | Finds names, places, dates, and other entities | IBM's example: "London" is a location and "Maria" is a person |
Sentiment analysis | Detects attitude and emotion in text | An angry message is routed to a person sooner |
Language detection and translation | Identifies the language and replies in it | A question in Spanish gets a Spanish answer from English content |
Question answering | Finds or writes the answer to a question | Retrieves the shipping policy and answers from it |
IBM notes that sentiment analysis is often used to route communications to the system or person most likely to handle them. For chatbots that answer in many languages, the list of multilingual AI chatbots compares the options.
What is an example of an NLP chatbot?
SiteGPT is an example of an LLM chatbot with retrieval. A business enters its website URL, uploads files, or pastes raw text, and the chatbot answers visitors from that content. The home page lists 95+ languages and an escalate-to-human button.
SiteGPT also applies NLP after the conversation. On Scale and Enterprise, AI conversation insights classify each conversation's topic, sentiment, and resolution outcome.
SiteGPT has limits worth knowing. It has no visual flow builder for scripted, step-by-step flows. Starter refreshes content manually, and automatic refresh starts on Growth (monthly). Chat integrations such as Slack and Crisp also start on Growth.
Starter costs $59 a month billed monthly, or $39 a month billed yearly ($468 a year). Every self-serve plan has a 7-day free trial, and a credit card is required.
How do you build an NLP chatbot?
To build an NLP chatbot, pick the approach first, then a framework or platform that fits it. Intent-based frameworks need intents, example phrases, entities, and flows. LLM platforms need content, instructions, and tests.
Framework or route | Approach | What you define | Fits |
Google Dialogflow ES | Intent-based NLU | Intents, training phrases, entities, contexts | Teams on Google Cloud building fixed flows |
Amazon Lex V2 | Intent-based NLU | Intents, sample utterances, slots, a fallback intent | Teams on AWS building fixed flows |
Microsoft CLU (Azure AI Language) | Intent-based NLU, with an LLM-based training option | Intents and entities, or intent descriptions for the LLM option | Teams on Azure |
Rasa | Agent platform with language models and RAG guides | Flows, custom actions, and prompts | Developer teams that want control over flows and code |
Model API plus your own retrieval | LLM with retrieval | Indexing, retrieval code, prompts, and the chat interface | Developer teams with time to maintain it |
No-code RAG platform | LLM with retrieval | Content sources, instructions, and the widget's look | Teams without developers that need a website chatbot |
The build itself has the same core steps on any route:
- List the questions and tasks the chatbot must handle, using real transcripts or emails.
- Split them into open questions and fixed tasks.
- Connect the content that answers the open questions, and build flows for the fixed tasks.
- Write instructions for tone, scope, and when to hand off to a person.
- Test with a fixed set of real questions, including some the content cannot answer.
- Launch, read transcripts weekly, and fill the content gaps they show.
Each step is covered in more depth in how to make a chatbot and chatbot training.
Common NLP chatbot problems and how to fix them
The most common NLP chatbot problems are ambiguity, missing content, invented answers, slang, and lost context. Most have a content or design fix.
- Ambiguous messages. "It doesn't work" names no product. Have the chatbot ask one clarifying question instead of guessing.
- Missing content. An LLM chatbot with retrieval cannot answer from a page that does not exist. Read the questions it declined and add the missing pages.
- Invented answers. LLMs can state facts that appear nowhere in your content. Ground answers in retrieval and instruct the bot to decline when no passage matches. The guide to stopping chatbot hallucinations lists more guardrails.
- Slang, typos, and dialects. Intent classifiers miss phrasings far from their examples. Add real customer phrasings to the training phrases, or use an LLM, which handles informal text better.
- Lost context. A bot that forgets the last turn asks again for the order number. Test multi-turn conversations before launch.
- No route to a person. NLP will misread some messages. Every chatbot needs a visible way to reach a human.
Try SiteGPT on your own content: the 7-day trial shows how an LLM chatbot with retrieval answers your real customer questions.
Frequently asked questions
Basics
What is NLP in a chatbot?
NLP in a chatbot is the technology that reads a visitor's message, works out what it means, and writes a reply in natural language. It covers tokenization, intent classification, entity extraction, context tracking, and response generation.
What is an NLP chatbot?
An NLP chatbot is a chatbot that understands free text instead of only buttons or exact keywords. It can treat "where's my parcel?" and "has my order shipped?" as the same request.
How does NLP work in a chatbot?
NLP works in a chatbot in five steps. It reads and tokenizes the message, works out the intent and entities, tracks earlier turns, finds the facts in a flow or in retrieved content, and writes the reply.
Comparisons
What is the difference between NLP and NLU?
NLP is the whole field of processing language. NLU is the part of NLP that reads input and extracts meaning, such as the intent and the entities in a message.
What is the difference between NLP and an LLM?
NLP is the field. A large language model is one tool inside it. An LLM can do several NLP tasks at once, such as understanding a question and writing the answer, which older chatbots split across separate models.
What is the difference between a rule-based chatbot and an NLP chatbot?
A rule-based chatbot follows fixed menus or exact keywords. An NLP chatbot interprets free text, so it handles phrasings nobody wrote a rule for.
What is the difference between machine learning and NLP in a chatbot?
Machine learning is the general method of learning patterns from data. NLP applies it to language. Intent classifiers and language models are both machine learning models used for NLP.
Building
What are the best frameworks for an NLP chatbot?
Common intent-based frameworks are Google Dialogflow ES, Amazon Lex V2, and Microsoft CLU. Rasa is a developer platform with guides for language models and RAG. Teams without developers often use a no-code RAG platform instead.
How many training phrases does an intent need?
Google's Dialogflow documentation recommends at least 10-20 training phrases per intent, depending on how complex the intent is.
Does an LLM chatbot still need intents?
An LLM chatbot does not need intents to understand questions. Many teams still use an intent or a button to start scripted flows for refunds, bookings, and other fixed tasks.
Accuracy and limits
Why does an NLP chatbot give wrong answers?
An NLP chatbot gives wrong answers when its content is missing or outdated, when it guesses instead of declining, or when a message is ambiguous. Retrieval from current content and a rule to decline when no passage matches reduce these errors.
Can an NLP chatbot understand different languages?
Yes. LLM chatbots detect the language of a message and can reply in it, even when the source content is in English. Intent-based bots need training phrases in each language.
Can an NLP chatbot detect emotions?
Yes, through sentiment analysis. The chatbot scores the attitude of a message, and an upset customer can be routed to a person sooner.
Sources
Sources checked on 18 October 2026.
Written by
