The Beginner’s Guide to Natural Language Generation (NLG)
You may have heard about the potential of artificial intelligence (AI), machine learning, and related technologies.
Indeed they can improve marketing performance, productivity and personalization.
But getting started with these technologies isn't always easy.
This post describes how to apply AI technologies to your marketing efforts.
In it, we'll give you a 30,000-foot view of an AI technology called natural language generation (NLG) that is used by companies like the Associated Press to scale content creation and general natural language using machines.
Then, we'll teach you how to get started with that technology today.
What Is Artificial Intelligence (AI)?
Artificial intelligence is the science of making machines smart.
Traditional marketing technology is built on algorithms that are told what to do by humans. With AI, the machine creates its own algorithms, determines new paths, and unlocks increased potential to advance marketing and business goals.
But "artificial intelligence" isn't just one thing.
It's a group of technologies, like machine learning, image recognition, deep learning, neural networks, natural language processing (NLP), and natural language generation (NLG).
Image Credit: Medium, John Koetsier
Each of these technologies can play a role in uncovering insights from massive datasets and achieving marketing and business goals using those insights.
And they're used widely already today by companies like Amazon, Google, and Apple to power product recommendations, voice commands on your phone, online search, and much, much more.
What Is NLG?
Natural language generation is a type of AI that generates natural language from structured data.
With the right data in the right format, a natural language generation system can automatically turn numbers in a spreadsheet into data-driven narratives or even use associations between words to create partially or fully machine-written text. Natural language generation systems use machine learning, deep learning, and neural networks (all forms of AI) to generate natural language.
These systems can generate natural language in a variety of formats. They can turn numbers into narratives based on pre-set templates. They can predict which words need to be generated next (in, say, an email you're actively typing). Or, the most sophisticated systems can formulate entire summaries, articles, or responses.
NLG technology and NLG models are used across a range of contexts. Today, NLG systems can write email subject lines, snippets of blog posts, summaries of in-depth reports, short-form advertising copy, chatbot responses, and much, much more.
NLG has been a field of AI research for decades, but the most impressive, and useful, advances in the technology have only happened recently.
As far back as 1986, research was published on possible NLG use cases. Ten years later, researchers at the University of Aberdeen were publishing about how to use the technology for text and sentence planning. As late as 2006, obstacles to NLG adoption were still being defined and discussed among leaders in the field.
Yet today you use NLG every day whether you know it or not.
A system like Gmail's Smart Compose now offers suggestions on what you should type next in an email. Then, it learns each time you pick one of its suggestions. It's using advanced NLG and NLP to read your emails while you write them, then predict what you are trying to say next-in real-time.
The Associated Press uses NLG to automatically create machine-written corporate earnings reports. In the past, a human reporter reviewed earnings data when it was released, then wrote a report on it. Today, the AP uses NLG to ingest that data, then produce a narrative in seconds, freeing up reporters to pursue higher-value tasks within the organization. All the AP earnings reports follow the same exact format. So, once NLG is set up, their system can produce thousands of stories at scale.
Even weirder, in 2019 OpenAI, a non-profit AI research company, announced they built an AI model that essentially writes coherent paragraphs of text at scale. The model is called GPT-2 and it learned how to write this well by analyzing eight million web pages. In 2020, OpenAI updated the model and released GPT-3.
In early experiments, GPT-3 has been used to produce everything from coherent blog posts to press releases to technical manuals, often with a high degree of accuracy. To do that, GPT-3 uses 175 billion parameters in its language model, compared to GPT-2's 1.5 billion.
It's still early days for GPT-3, and the validity of the model hasn't been fully explored. But one thing should give marketers pause:
The speed of improvement in OpenAI's language models.
The first GPT model came out in 2018. GPT-2 was released with greatly expanded capabilities in 2019. Just a year later, GPT-3 uses 100x as much data as its predecessor and is beginning to display incredible content creation capabilities, including turning text into code and evaluating investment memos.
This type of machine-powered language generation has serious implications for marketers.
With properly structured data, marketers can use NLG to automate some narratives that are based on data. That might include analytics reports, data-driven sections of blog posts, product descriptions with standardized formats and lots of specifications, or even partial or entire articles.
With the right data and a properly configured NLG template, you can leave some types of content creation to the machines while humans create other types of content and double down on promotion.
NLG vs. NLP vs. Natural Language Understanding (NLU)
NLG is one piece of the AI-generated content puzzle. When you look into NLG solutions, you'll also encounter other related technologies, notably natural language processing (NLP) and natural language understanding (NLU).
We will devote longer articles to these topics in the future. But, for now, here are the basics you need to know.
Natural language processing (NLP) is the technology that turns your words (written or spoken) into data for the NLG system to understand. NLP is the step that occurs before language can be generated through NLG. As put by NLG provider Narrative Science, "Natural language processing reads while NLG writes" (or speaks).
Natural language understanding (NLU) is a subset of NLP. NLP is focused on getting the actual data into a structure that a machine can then use for NLG. NLU is the step that comes before this, by extracting the meaning of and relationships between words.
First, you need NLU. Then, you can do NLP. Once you do NLP, you have the ability to use NLG.
So how do you actually use NLG in your life, in business, and in marketing? There are a number of major applications for the technology.
Text Generation and Content Production
NLG systems can produce short-form, long-form, or summarized text.
In the case of Gmail Smart Compose, NLG produces short-form text snippets based on what the system thinks you type next. Other NLG systems can produce short-form ad copy (including Facebook and search ads) or email subject lines.
NLG can generate long-form text snippets to supplement blog posts or create more substantial written narratives, including written interpretations of performance data.
The technology can also take long-form text and extract shorter summaries of the content, producing briefs and outlines.
Major voice assistants like Apple's Siri and Amazon's Alexa rely on NLG to answer your questions and queries. Any software program or device that speaks back when you speak to it is using some form of NLG.
NLG is used heavily in many chatbot technologies to partially or fully generate natural language responses to use queries. Sophisticated chatbots use NLG to interpret what you type, then generate responses based on what you say.
This may include customizing language based on your query or tone or offering specific messaging depending on the topic. It may also include answering specific questions you ask by extracting content from areas of the website and packaging it into a response.
Automated Lead Nurturing
Combining some of the capabilities above, some automated lead nurturing tools use NLG. These tools are AI-powered solutions that automatically respond to leads via email and chat using NLG. This ensures all leads coming into a pipeline receive a response 24/7, whether or not human reps are available.
Hundreds of natural language tools exist. But several of them are ready for use today by marketers across a variety of use cases.
Arria uses NLG to extract and summarize data into readable narratives.
Automated Insights uses NLG to help organizations create earnings reports at scale, sports articles (box scores, results, etc.), and data-driven narratives.
Clickvoyant uses NLG to dramatically reduce the time it takes to extract insights from analytics and create presentations based on those insights.
Drift uses conversational NLG to remove friction from the buying process with chat, email, video, and automation products.
Exceed.ai uses AI to engage with every sales lead that enters your pipeline, using human-like, two-way conversations by email and chat.
MarketMuse is an AI-driven assistant for building content strategies and creating content. It uses NLG to create summarized briefs telling you how to write posts for maximum impact and will even automatically generate text for you.
Pencil uses NLG to automatically generate better-performing Facebook ads.
Persado uses NLG to generate marketing copy and creative to create the best possible messaging for your individual prospects across channels.
Phrasee uses NLG to automatically write email subject lines better than humans, resulting in higher open rates.
Snazzy AI uses NLG like GPT-3 and the company's own machine learning layer to create compelling content for ads, landing pages, and product descriptions.
Yseop uses NLG to automatically generate narratives from data across financial and medical reporting.
How Do You Get Started with NLG?
So how do marketers actually get started using natural language generation? There are a ton of exciting use cases for natural language technology that might now possible thanks to recent developments like GPT-3. We're just beginning to explore what might be done with fully machine-generated articles. This work can be highly experimental and quite technical—and may not be appropriate for businesses just getting started with a natural language generation tool.
But, you can get started in simpler, yet powerful, ways with this technology—without hiring a data scientists or revamping your entire operation. Here are some initial steps that any marketer can take, starting today.
1. Determine if you have a use case for basic NLG.
Look first to the stories you're already manually telling with numbers. Think about "stories" as any narratives that make sense of data. This might include external-facing or internal-facing reports, summaries, fact sheets, etc.
Do you produce these types of stories regularly? Are any of these narratives in a consistent, repeatable formats (i.e. you're reporting on or telling a story about the same types of numbers each week or month)? These might be candidates for natural language generation.
For instance, PR 20/20 , the marketing agency behind the Marketing Artificial Intelligence Institute, has used NLG to cut down analysis and production time of Google Analytics reports by 80%. This isn't an advanced NLG use case that leverages something as robust as GPT-3, but it is a valuable one.
It just goes to show that low-hanging fruit may create value faster for your organization, while teaching you the basics of natural language generation.
2. Look at how your data is structured.
Even with a use case, natural language generation needs structured data to work.
Are your datasets organized in ordered columns and rows? The current NLG solution we use requires a CSV upload, so data needs to be clean and relatively consistent to get value out of this technology.
Or, you may need to invest time into cleaning up your data before uploading it to a system that uses natural language generation.
3. Be realistic about your ROI.
NLG solutions, even basic ones, typically require substantial time to set up. You also need to pay for a solution, and possibly related NLG services. You'll want to take a realistic look at the technology, what it can do for you, and how much you can scale using it.
Start by analyzing how long reports, articles or narratives currently take, then see how much time NLG can potentially shave off.
Finally, apply those time savings to all staff whom NLG would affect. An hour saved per week per employee may make financial sense for your organization.
About Mike Kaput
Mike Kaput is Chief Content Officer at Marketing AI Institute and a senior consultant at PR 20/20. He writes and speaks about how marketers can understand, adopt, and pilot artificial intelligence to increase revenue and reduce costs. Full bio.