Natural language processing (NLP) has been around at least since the early 1980s, under various names. Back then, I remember machine translations of Russian language materials, for example. Those translations were just plain awful; my Navy compadres and I couldn’t make head nor tail of them.
Back then, before the current crop of data scientists were born, we also had Optical Character Recognition. OCR, by contrast with the machine translation of the age, at least had some utility. The IBM Selectrics in the office had special OCR balls you could use, and special forms you could type on.
If you managed to type in the right places in the right blocks with the right type of text on the form, you could send secure messages through official channels via the communications center in the building. The comms. center people scanned the OCR forms for transmission. It was a comms. process that was very fussy to use, and the comms. people often kicked back typed, formatted messages that couldn’t be accurately scanned because of some lack of readability in one block or another. You’d have to redo them. But if you were careful, the process worked.
Today’s process automation: A more capable version of macros
In 2022, NLP and OCR are both still around and a lot more useful than they used to be. But they’re still difficult to use, because we’re trying to use them to solve harder problems now than we used to.
Take the current process automation challenge, for instance. One trend big banks and other information-intensive organizations have tried to latch onto over the last five or ten years has been “Robotic Process Automation,” or RPA. RPA as implemented learns steps that an office worker with a PC takes to complete a rote process, By studying the user’s actions in the user interface, or presentation layer, RPA records those steps and makes it possible to partially automate the process.
A major difference between RPA and the Excel macros that have been around for decades is that RPA works between applications, including those from different vendors. In today’s large enterprises, many of the major applications in broad use are de facto standards–Salesforce, Sharepoint, Office, and so on. When a worker does a few steps in Office, then Salesforce, then Sharepoint for another step or two, then back into Office to complete the series of tasks, RPA can record the whole process between the different applications and provide the means of reusing that recording so that the worker doesn’t have to take steps manually. The scripts the recording generates can be modified and repurposed, similar to the way the Visual Basic used in Office macros is extensible.
RPA versus Intelligent process automation, or IPA
If RPA on the face of it sounds a bit Rube Goldberg-ish to you, the software equivalent of a brittle, unwieldy contraption, you’re right to be suspicious. “RPA scripts are highly coupled with the UI structure, making them very fragile to subtle changes in the UI,” Nicolas D’Ippolito, Head of Engineering and Associate Professor at Buenos Aires University, points out. If the challenge is to automate more of the workflow of customer service representatives in a contact center, for instance, more of the context of the workflow and the purpose of the business unit has to be modeled to make what the machine side does adaptive to UI changes.
D’Ippolito’s company ASAPP focuses on contact center automation. When it comes to more intelligent process automation, ASAPP takes a more decoupled approach. “Since we have a behavior model of every system,” he says, “we can detect deviations from standard usage patterns due to changes in navigation flows in the UI. We can also detect new states in the system that we didn’t observe before due to changes in the UI or agents’ usage patterns.”
As a result, “these deviations are considered during the automatic retraining cycles. Our models adapt to the changes in the UI, and navigation tools are re-generated, ensuring agent access to needed information is always current.”
The commoditization of RPA and the move to IPA
For a number of years, startups such as UIPat, Automation Anywhere and Blue Prism captured most enterprise RPA market share. But more recently, Microsoft has gained steam with its Power Automate product line, leveraging its installed base of Office in the process, as well as other large SaaS providers who have their own RPA product lines. These can leverage their own installed bases to gain share.
So now the pure plays in the newer IPA portion of the market are scrambling for NLP advantage.In an August 2022 piece, Esther Ajao of TechTarget noted UIPath’s acquisition of NLP vendor Re:infer. “UIPath showed it is trying to move beyond being an RPA specialist and compete more effectively with the likes of tech giants such as Microsoft,” citing Leslie Joseph, a Forrester Research analyst.
Automation fabrics, fact vs. fiction
Forrester points to the unclaimed territory of an “automation fabric,” and NLP is one of the means that can help. But any sort of true fabric at enterprise scale implies the need for more than a point-to-point integration capability, and more than a modeling of domains one by one that can’t then easily talk to one another.
Once again, the lack of a knowledge graph-enabled foundation designed for many-to-many interconnection and interoperability becomes the inhibitor to scalable, vendor-agnostic automation.