Table
of Contents
- Introduction
- What is NLG?
- Who uses it?
- Data
- Humans
- Isn't it just a content spinner?
- A little background
- What's the difference?
- How do we use it?
- What industries have we helped?
- Is all this effort worth it?
- What does the future look like?
- Are the machines taking our jobs?
- Let's get personal
A team of two within our writing department produced 1,469,398 pages of digital content for one client, in less than a year.
How many hours a week do we work? Forty hours on average, or what used to be a typical work week for many. Do we have hundreds of writers on our team? No. We also aren’t using an army of freelancers (or monkeys with typewriters). It’s mostly two people creating all those thousands of pages. In addition, we’re not writing one version duplicated thousands of times... but we’re also not writing each individual piece.
So, how
is it that a small team of writers can produce so much in such a short period of time?
Natural Language Generation (NLG), aka Ai Text Generator
What is Natural Language Generation (NLG), and how do we use it to help our clients?
Glad you asked. Read on to learn more about artificial intelligence content creation programs, and what the future of content may look like.
You may have heard about Natural Language Generation (NLG) and Natural Language Processing (NLP). Here’s the lowdown: NLPs “read” prose written by humans and turn it into data. NLG (which we will be discussing at length in this piece) is a subcategory of artificial intelligence (AI) that can generate narratives or reports in readable prose by feeding the program a data set.
In short, you put data in, and you get narratives out. You could say that Ai text generators (NLG software) translates the data. Unfortunately, NLGs aren’t to the point where you can just upload a treasure trove of data and get your story.
Not yet anyway.
There are close to 2 billion pages on the internet (at least as of this writing…it could be more now), and it is probably safe to say that the majority of these were written by humans.
Think of all the travel blogs, advice columns, sports recaps, and long-form journalism that you encounter every day. But in those billions of pages, there is a growing amount that is produced by Ai content generators.
.png)
Okay, sure—but you’re not into quarterly earnings reports. Fair enough.
But what if your 1986 Honda Accord finally gives out and you want to do a little research on new and used cars?
You might find yourself on Edmunds reading car descriptions. Chances are the majority of those were written using NLG as well. In fact, after looking through the websites of several local dealers, it’s entirely possible you’ve read nothing but NLG-produced descriptions.
According to an article on KD Nuggets, “Dominion Dealer Solutions use an NLG tool to generate unique vehicle descriptions based on data gathered from a variety of sources, including automotive reviews, Kelley Blue Book, and CARFAX.”
More importantly, is it working? It seems to be. That same article also revealed that cars with NLG-generated descriptions sold, on average, 20 days faster than those without.
Quarterly reports, car descriptions, and weather reports (oh yeah, Ai writers are used to produce some of AccuWeather’s reports, too) aren’t something that most people are going to enjoy reading every day. And in most cases, this type of writing is pretty impersonal.
But can NLG create something that has a more personal touch? What if it could create a narrative just for you?
Well, if you are a Fantasy Football fan and use Yahoo!, then you’ve probably received a recap like this one on your most recent draft.
It’s just one use-case, but this is when Ai text generator software really begins to shine from a UX standpoint. It would have been easy to just slap a letter grade on the draft and call it a day. Instead, the narrative has made the whole experience more immersive by personalizing it.
The personalization goes further than the recap discussing the picks, though. It also analyzes those picks based on the data and delivers the analysis not in a spreadsheet, but with a snarky recap that is ideal for this venue.
So, what’s the point of all this wordplay wizardry? It makes users want to participate in Yahoo! for its Fantasy Football league. Players can look forward to these recaps no matter how much they might disagree with them.
They also help create a deeper experience by making you feel as if you are running an actual NFL franchise. The recaps will celebrate your victories and admonish your failures, much like the beat reporters do for their hometown teams.
What’s more, is that it creates something you can share with others. We shared this example on Facebook with a comment about how the team would prove the doubters wrong. And we weren’t the only ones to do so.
However, the personalization doesn’t end with Fantasy Football. Bodybuilding.com also uses Ai writing in its BodySpace app to send its members reports on their weekly workouts in the form of a fun and engaging narrative.
Besides offering some motivation, their reports provide statistics about the user’s workout..
Moreover, the statistics don’t just cover the transactional record of the user’s most recent exercises—they also show trends. What the end-user gets is not only information, but a digital trainer. But all this doesn’t happen without some very important components.
One component that Ai content software relies heavily on is data. In most cases, the more data you have, the better your output. But it can’t just be data for data’s sake; the information has to be applicable to the narrative you are creating. It also must be structured in a way that allows NLG programs to consume it. So, what kind of data are we talking about? Well, that depends on the story you want to tell.
Let’s take some of the work we have done for PACIFIC’s clients in the travel industry. If you wanted to put content on some hotel pages, what data would you need to produce engaging copy? What does that data look like? You would definitely have the city, state, and country. Maybe what amenities are available at the hotel? Number of rooms at the hotel? Room size? Things to do nearby? You can see where this is going. Each of those pieces of information needs a data point. And the more granular you can get with your data, the more dynamic your narratives.
If you’ve worked in an office environment, you know how daunting large spreadsheet files can be. Now, imagine one that is 50 columns across and hundreds or thousands of rows long. That is what your data could look like. It could even be more. With most NLG software, you don’t need to work with that large of a data set (and you don’t want to, either). You want to work with a smaller subset of data that is a snapshot of the total output.
Before you start writing, you must have the data in place. Unlike a human writer, NLG programs will always trust the data you give them. If your data is inaccurate, your narratives will be inaccurate; the programs are not going to alert you that the Statue of Liberty is not in Topeka, or that the population of San Diego is not 2,000 people. What’s more, this Ai content writing software does not automatically organize the data that you need, nor does it guide you in implementing it. You still need a human to discern what is correct, organize it, and edit when necessary. Which leads us to the next vital component…
Yes, NLG software still needs humans to do the actual work. All that data needs to be organized and analyzed. You can certainly use tools to help you streamline and concatenate all that data, but as previously mentioned, you have to ensure the data you are using is accurate. If not, then you don’t just have content at scale, you also have mistakes at scale. And these checks require a sharp eye and a knowledge of the data being used by the program.
Okay, so you’ve organized your data and vetted it. Now what? Well, you’ll need a data analyst or scientist to figure out how to use that data. But let’s be real.
In many cases, it is the writer who does the analyzing. Which means your writer will have to wear their data scientist hat (which fits surprisingly well over the brooding-artist and helpless-romantic hats we’ve seen our teammates wear). Remember, it can’t be data for data’s sake. You need people on your team who are creative enough to tell a great story with the data. To help Ai text generator programs really shine, you have to layer that information. But it’s more than just sprinkling some data throughout your writing.
You also have to structure your template in such a way that the program knows what to say and when to say it. Many NLG programs are based on Boolean logic (also known as “if-and-then” statements).
Just like the programs won’t tell you that there’s an issue with your data, most also won’t tell you if there is an issue with your formulas or how they are structured. Circling back to our hotel example, you may want to tell the reader something specific about hotel ratings.
The formula you would concoct for that would look something like this:
We get this question a lot. But content spinners and Natural Language Generation programs aren’t the same.
So, what is content spinning? It’s the practice of rewriting or repurposing (“spinning”) a piece of content or information (like an article) with minor changes. In theory, a writer could spin their own content into several variations and post them on multiple sites. Or they could use a content spinning program and alter someone else’s work, re-spinning multiple variations.
Mad Libs is an excellent example of how content spinning works. The players take turns substituting different adjectives, nouns, or verbs into the story’s blanks. Your player sheet could read something like:
“Once there was a [adjective] man named [name] that lived on a hill. His favorite pastime was [verb].”
The game usually produces an absurd and humorous story, and many of the worst content spinners do much of the same. Moreover, this type of content isn’t going to come close to passing the Turing Test (which distinguishes machine from human writers).
If you want to tell the same story with some phrasing differences (and the possibility of it just being nonsense), a content spinner is perfect. However, NLG programs do much more than this, and they are more complex.
Content spinning software isn’t anything new—the practice has been around for over a decade. The issue with many of these content spinning programs is that the algorithms will pick terms from the program’s thesaurus and make unidiomatic decisions that no human writer—well, no decent human writer—would ever make.
In fact, Burger King previously created an ad campaign that pokes fun at how bad content written by artificial intelligence can be.
The content produced by early content spinning and Ai article spinner applications weren’t much better than Burger King’s satire. Beyond the wonky sentences, the content usually had very little originality, and it was not location or even page-specific.
Content spinning has a few more knocks against it as well.
While content spinning is a quick way to generate a massive amount of varied content, it also falls into the practice of black hat SEO, which attempts to take advantage of search engine guidelines. Because of this, content spinning is now considered the perfect way to destroy your rankings and your reputation in one fell swoop.
Think about all the times you may have scrambled to get your homework finished before class, or how you would re-write answers that were perfectly laid out in the book. We all had someone in our class who would copy answers verbatim, and they were almost always caught. These amateur content spinning techniques were called out and penalized by Google’s 2012 Penguin update. Happily, there are better ways to tell the same story multiple times, and when used properly, Ai writers can do just that.
Let’s say you’re looking for an answer to a question like, “When was the Boston Tea Party?” and the textbook states, “The Boston Tea Party took place before the Revolutionary War.” A student decides to write as their answer, “The Boston Tea Party happened before the Revolutionary War.” Not very original. Do you feel like you basically just read the same sentence twice? So do we.
Consider how a more complex, data-filled application would answer this—instead of an irresponsible high schooler pretending to be original. The textbook might have told us when this historical event took place, but we can dive deeper and say something like: “The Boston Tea Party, carried out by the Sons of Liberty in Boston, Massachusetts, issued public offense against the Loyalists and British government. This historic moment took place on December 16, 1773, in Boston Harbor, and it eventually led to the Revolutionary War.”
Text generator Ai applications take data—in this case, the event and the date—and with the writer’s guidance, create multiple meaningful alternative answers with more information… that is if data is included.
With NLGs creating content like this, the audience (consumers) will be able to get more information without sifting through dozens of websites to gather everything they need.
If we were to use a real-world example, we can pull directly from the pages of one of our travel clients. This client needs location-specific boilerplate with plenty of vacation ideas and hotel information. Many similar organizations in the industry will have boilerplate, but it may all read the same, with only the name of the city to set the pages apart. Besides this small detail, the pages would likely all have some version of this sentence: “When you come to San Diego, you’ll find things to do close to your hotel.” If we write the same boilerplate with NLG, we can say something like:
This sounds more natural and provides more information. With a content spinner, you often get a taste of what you need, but you may not have all your questions answered. With the right data, NLGs produce content that answers audience questions and reads more like custom content. This allows the content to be more authoritative, which leads to higher organic traffic.
Does your company require a lot of internal and external reporting? Many Ai text generators can see patterns in a given data set and help you create narratives based on those. This can save your internal team hours of reporting time.
What kind of industries can use automated content? Pretty much all of them. If your company has a large website, those pages can benefit from optimized content. We’ve already discussed how it can help travel, but just about any industry that uses a large website to host thousands of pages or products is where NLGs are most beneficial. For example, a good NLG template can help you create engaging and informative content for each product listing.
No. There is a reason we keep bringing up the travel industry in our examples: It’s because we’ve been successful using Ai content generation to boost organic traffic to our travel clients’ websites. One page type we wrote had a +97% pre/ post in visitors, along with a +28% improvement in year-over-year delta. In case you’re wondering, that is really good. We also helped another travel client increase their share of voice by 33%. In addition, we’ve successfully used NLG with several of our clients outside the travel industry, including healthcare and retail.
Like the best content, a lot of strategy goes into launching a successful template. Our SEO team finds high-ranking keywords, “People Also Ask” questions, and other factors that help us create content that will rank. Our writers determine how we can differentiate our client’s page types from the competition. We find information that can either answer questions or say something unique about a place or product. Our editors have to not only check our writing for mistakes, but also make sure that every iteration of each sentence is grammatically correct and reads naturally. We also have to take formatting for page types into consideration, as well as ensure any necessary HTML can be scaled.
Is all this effort worth it?
Our clients seem to think so, and it is a lot faster than writing and editing 194,000 250-word pieces of content. How much faster? Well…
Without NLG = About 6 years
With NLG = 20-25 days
The shocking thing? It may already have happened. A few years ago, a robot-written novel passed the first round of a literary competition. Which leads us to the question that just about everyone has asked for the last 200 years...
This fear isn’t anything new. In the 1800s, British textile workers raged against sewing machines, and this movement helped usher in the term “Luddite.” More recently, Elon Musk’s dire warnings about our artificial overlords have made headlines, even as Tesla’s own factories seem to rely on automation to produce zero-emission vehicles.
But one job was thought to be above automation: writing. Yet here we are, waiting for the cold robotic claw to snatch the pen from our hand.
PACIFIC employs a few full-time AI writers, plus others that help analyze data and edit the content. But will all these jobs become automated too? Probably not. Here’s why...
Will there be a time when NLPs and NLGs can easily create something that matches human-written prose? Most likely. Ai copywriting has proven invaluable in producing large volumes of informational content. Besides producing it at scale, it can also be more easily edited at scale, which means a refresh of thousands of pages is only a few clicks away.
But until then, it will need a human hand to guide content at scale. And for blog posts, long-form articles, white papers, etc., a great (human) writer is still your best bet.