From our inbox…
“So, I read What Marketers Need to Know About Natural Language Generation (NLG), and I want to use it now! Give me some starter tips!
One of the easiest places to begin – and frankly, to see firsthand how NLG works in your business – is to review your existing content. Make a list of all the things you are writing each month, and then identify the #1 thing NLG could help you with immediately. Perhaps a monthly analytics summary for management? Technical updates to your customers? FAQs for your new products? Updated questions for your chatbot? I know. I know. None of those ideas are as sexy as the “we’ll write all the copy for your 2,000 new Holiday items in less than an hour” pitches promised by vendors these days but the easiest way to ensure successful projects is to start with something manageable so you can learn how things work in real-time. But I digress…
Take ONE piece of content and figure out how much time it takes to put it together on your own, unassisted. Next, figure out how much time it will take to do it assisted. (Several NLG vendors have trials and demos or can help you figure this out.)
Once you’ve identified your piece of content, outline the existing process. When starting with NLG, some of this may seem like an exercise in futility but it’s critical that you establish a solid foundation, train the system correctly and, most importantly, know how to identify the elements and their relationships when things go awry. And they will run amok at some point, that’s for sure.
Then, create the template. The template will be your finished product, so include what you want/need and eliminate everything you don’t. (It’s easier to start small and then add.) Many NLG vendors have pre-set templates you can refer to and/or use. Your template can be Mad Libs-like, or it can be more open-ended.
Depending on what program/vendor you’re using, the road branches out here. Some services do all the heavy lifting for you. Others ask you to identify your data points and write out your conditionals. Either way, sketch out what you want and how you want things to look. Writing a hierarchical list of what’s most important to you may also be helpful. For example: I want a 10-point summary of all this year’s email results. From a presentation perspective, if Revenue is what’s most important, list that first. Then, the number of new customers converted/acquired; the number of inactive customers reactivated, and so on. Executive Summary on top. Supporting graphs, tables, etc. below.
Execute your test. How much time will you save? Can the NLG program(s) write most of a first draft so you can just go in and tweak a few things here and there? Can it write half? Will it produce something you can work with and/or feel good about? How/where is the NLG solution adding value? Please remember that you will need to review any/all content before it’s used/posted.
“Should I buy an NLG package, or should I build one in-house?”
Depends on the project.
I know. I know. Garbage answer…. and it depends on the size of your project; how much time you have to build it, to get it done and babysit it; and how much money you have.
My general rule of thumb is that if it’s something super tasty that you don’t want anyone else to have the recipe for or to even learn about, you need to build it in-house. Some folks will say that’s a short-sighted response, but I’ve found most of them have skin in the game (either as a vendor or a consultant getting kickbacks) and that the majority of the time, your KFC special recipe, Big Mac sauce, etc., should be in your vault. And yes, there are a few fantastic vendors I trust and to whom I am willing to give my secret recipes. However, all but one of them plans to sell in the next five years, and to be frank; I don’t want my secret recipes at the players they’ll likely sell to. Been There. Done That. Didn’t even get a T-Shirt.
Will my NLG models get better over time?
Almost every NLG vendor I know immediately tells you, “Yes.”
From my experience, it’s “the majority of time with proper care.” Right now, this stuff is way less set-it-and-forget-it than it’s being sold.
Your models will adapt as you bring new data into the mix.
If your data foundation is solid and you test, train, and babysit your models properly, things will often improve in ways you can’t imagine. (Good ways…)