---
title: "What is natural language generation (NLG)?"
description: "Natural language generation turns structured, non-linguistic output into text a person can read. Learn more about the technology and common models that you encounter every day. "
image: "https://images.ctfassets.net/lzny33ho1g45/7665JecLHJPvfybteryQtI/fc1ea0368395c618b6a4d73915bbf703/language-hero.jpg"
---

# What is natural language generation (NLG)?

Natural language generation turns structured, non-linguistic output into text a person can read. Learn more about the technology and common models that you encounter every day. 

A few days ago, my fitness app congratulated me on a "great week of activity" with a poetic recap of where I "crushed it." I had spent most of that week visiting family, moving from the couch to the kitchen to the guest bedroom at the speed of continental drift. The app took my step counts, clearly rounded to the nearest thousand, and phrased them _generously_.

Once you start noticing those AI quips, it's hard to stop; and it's not just on your fitness apps. There's the recap at the top of your analytics dashboard, the summary line in a bank alert, and the first reply you get from support. Spoiler alert: none of that was typed by a person.

Data in, readable language out is what natural language generation does, and it's been doing it for decades—long before the recent AI hype. Knowing how it works makes it easier to tell which AI tools are worth your time, and which are basic features wrapped in fancy sales jargon. 

**Table of contents:**

- [Natural language generation (NLG) definition](#definition)
- [NLG vs. NLP vs. NLU](#nlg-nlp-nlu)
- [How does natural language generation work?](#how-nlg-works)
- [Extractive vs. abstractive summarization](#extractive-vs-abstractive)
- [Common NLG methods](#nlg-methods)
- [Natural language generation use cases](#nlg-use-cases)
- [Use natural language generation with Zapier](#zapier)

## Natural language generation (NLG) definition

Natural language generation (NLG) is a subfield of artificial intelligence and computational linguistics that turns structured, usually non-linguistic input—like data, numbers, and semantic representations—into text a person can read. Some NLG jobs start from text instead, but the workflow is the same: something goes in that isn't a finished sentence, and a finished sentence comes out.

If it helps you understand, you can think of NLG as a friendly coworker. You hand them a spreadsheet of last quarter's numbers and ask them to explain it in a meeting. They'd likely pick the figures that matter, order them sensibly, and say them out loud in plain English. NLG does that job at a large scale, whether the job is a mail merge or a conversation with [ChatGPT](https://zapier.com/blog/how-does-chatgpt-work/).

## NLG vs. NLP vs. NLU

There are a lot of "N"s and "L"s in there, so it's easy to get these terms mixed up, but here's the gist: 

- [**Natural language processing**](https://zapier.com/blog/natural-language-processing/)** (NLP):** This is the head honcho, covering all the techniques that let software work with human language, like [tokenizing](https://zapier.com/blog/what-is-a-token-in-ai/), parsing, classifying, and translating.
- **Natural language understanding (NLU):** This is the comprehension half, which converts messy human input into a structured representation of intent and meaning that a machine can act on. It's what lets your phone work out that "text mom I'm running late" is a command with a recipient and a message, not a sentence to read back to you.
- **Natural language generation (NLG):** This is the production half, running that same trip in reverse, from structured meaning to fluent, grammatical sentences. It's what turned my four-figure step count into a "great week of activity."

If you're using a [chatbot](https://zapier.com/blog/best-ai-chatbot/), you're seeing a combination of these. NLU deciphers that your "my order still hasn't shown up" message is a complaint about shipping status, and NLG writes the sentence that answers it. You need both halves working, or there isn't really a conversation happening.

**What it handles**

**Core question	**

**Example tasks**

**NLP**

Everything computers do with human language

How does a machine handle language?

Tokenization, translation, text classification

**NLU**

Input and comprehension

What did this person mean?

Intent detection, semantic analysis, entity extraction, syntactic parsing, sentiment analysis

**NLG**

Output and production

How do I say this back?

Report writing, summarization, chatbot replies

## How does natural language generation work?

The classic account comes from [Ehud Reiter and Robert Dale](https://dl.acm.org/doi/10.1017/S1351324997001502), who broke NLG into six tasks in 1997. Those tasks answer two questions in order: what to say, and then how to say it. 

- **Content determination:** Picking which facts belong in the text and which don't.
- **Discourse planning (later renamed document structuring): **Organizing those facts into a structure, consisting of what leads, what supports, and what follows.
- **Sentence aggregation:** Combining related facts into one sentence instead of three little choppy ones.
- **Lexicalization:** Choosing the actual words. "The train departed" and "the train left" carry the same fact at a different level of pretentiousness.
- **Referring expression generation:** Deciding whether an entity is "Acme Corp," "the company," or "it."
- **Linguistic realization:** Applying grammar, morphology, and punctuation so the output reads correctly.

Older systems ran these as a literal pipeline, each stage handing its output to the next. Modern models don't: all six collapse into one learned process, rather than being handed off between stages. The six tasks are still a useful map of what the model is doing, but you just can't pop open the hood and find them all labeled one by one.

## Extractive vs. abstractive summarization

I'd wager you've already read some type of AI summary today, whether that's the Google recap in your search results, or the email catching you up on a Slack thread you stopped caring about weeks ago. Those summaries come from one of two approaches: extractive or abstractive.

- **Extractive** systems work like a colleague with a highlighter; they pick the most important sentences and hand them back word-for-word.
- **Abstractive** systems, on the other hand, read the whole thing and write something new, using phrasing that may appear nowhere in the original.

That makes extractive output safe yet robotic. Every sentence is reproduced verbatim, but reading them back-to-back is choppy because those sentences were written to sit in different parts of the document. Abstractive output reads as if a person wrote it, but rewriting involves recombining details from across the source material (which is a common cause of [AI hallucinations](https://zapier.com/blog/ai-hallucinations/)). 

## Common NLG methods

If you've ever sent a mail merge that dropped {FirstName} into a greeting, you've used the oldest NLG method in the book. Beyond that classic, here are some of the common NLG methods you'll see in the wild:

- **Templates:** A sentence written once with blanks where the data goes, filled every time it runs. Newsrooms have used them for years to publish earthquake and election results within minutes of the numbers arriving. You'll never be caught off guard by the output, which is kind of the whole point.
- **Rule-based systems:** A step past templates, with explicit rules for each stage of writing, covering what to include, how to order it, and which words to choose. The results are predictable and consistently stiff, like an email from that friend who spent one year at Oxford and won't stop talking about it.
- **Statistical models:** Rather than following rules, these pick each next word based on what usually follows in a large body of text. It's the same idea behind your phone's keyboard finishing your sentence for you.
- **Deep learning: **Neural networks, including recurrent architectures like LSTMs, that learn language patterns from examples instead of from rules. They handle short passages well and tend to completely lose the plot across long ones.
- **Transformers and LLMs:** In 2017, a new architecture replaced step-by-step processing with self-attention, letting a model weigh each word against the rest of the text it can see, in a single pass instead of one word at a time. That made training far faster and long-range context tractable, and it's the architecture under most [major LLMs](https://zapier.com/blog/best-llm/) in use today.

Most natural language generation software you can buy today sits on one of the last two rungs, even when the vendor doesn't say so on the pricing page. And plenty of it sits on the first rung and calls itself AI; here's a tip: if a tool's output never surprises you, you're probably dealing with a template.

## Natural language generation use cases

Natural language generation usually shows up anywhere there's more data than anyone wants to write about by hand. That covers dashboards, support queues, product catalogs, and most of the [generative AI tools](https://zapier.com/blog/generative-ai-tools/) you already use.

- **Chatbots and AI agents:** A support bot builds each reply word by word, shaped around the question in front of it. This leads to better responses than older bots, which picked the closest match from a list of prewritten answers. The catch is that writing the reply and acting on it are two different jobs; an agent can draft the refund confirmation on its own, but it needs a connection to your billing system to actually issue the refund. That's what a [model context protocol (MCP)](https://zapier.com/blog/mcp/) is for.
- **Voice assistants:** Ask about the weather, and the assistant turns the forecast data into a sentence, then reads it aloud. It's the reason you hear Siri tell you that rain is starting around four instead of a table of hourly precipitation percentages.
- **Automated reporting:** Sales figures go in, a written summary comes out telling you what grew, what dropped, and what's worth sending to your manager.
- **Business intelligence:** Dashboards show you the graphs, while NLG writes the sentence explaining them. You get an answer to why your signups had a huge jump last Tuesday without waiting for an analyst to get back from their lunch break.
- **Personalized communication and outreach:** With NLG, every recipient gets their own email or product description, tailored to what your CRM knows about them. It beats blasting 4,000 people with one generic message.

## Use natural language generation with Zapier

NLG is the part of AI that turns structured data into sentences, and it was doing that long before LLMs made it look effortless. Knowing which method you're actually looking at tells you what to expect. A template isn't going to surprise you, but an LLM might. That's useful both when you're picking a tool and when you're trying to work out why one went off the rails.

[Zapier](https://zapier.com/mcp) can help you put NLG to work. It connects Claude, ChatGPT, Cursor, and other AI tools to  across your tech stack, so the assistant that just drafted a summary can also file it in your team's docs, update your database, and send it over Slack. And because every connection runs through Zapier, your credentials are never exposed to the model, and you control which apps and actions are on the table.

**Related reading:**

- [What is an AI agent?](https://zapier.com/blog/ai-agent/)
- [AI workflow automation: What it is + how to get started](https://zapier.com/blog/ai-workflows/)
- [How to build safe and trustworthy AI agents with Zapier](https://zapier.com/blog/safe-trustworthy-ai-agents/)
- [Zapier MCP: Perform tens of thousands of actions in your AI tool](https://zapier.com/blog/zapier-mcp-guide/)