August 8, 2022

TILT Can Go Where Others Can’t

If you’re following the news on deep learning natural language processing models, you’ve most likely heard of GPT-3, the general transformer language model released by OpenAI in 2020. And if you’re following the news on the best document automation solutions on the market, you might have wondered if—and how—Applica’s TILT differs from GPT-3. In this post, we’ll describe the crucial differences that set apart our proprietary technology.

Applica’s uniqueness has its origins in our company’s roots and mission. We are a service company first, a tech shop second. Automating documents was not something we stumbled into while playing around with code for something else, and it is not something we’re retro-fitting into a wider plan to dominate the language automation sphere in general. In fact, our company originally set out to create a solution tailored to the way organizations must engage with and act upon business documents. We’ve dedicated ourselves to this exclusive specialization from day one. For nearly nine years, we’ve been refining our unprecedented solution to a specific business need. Do we differ from the general language model that’s making the headlines? You bet. And this is a good thing.

One very significant difference has to do with what our model knows vs. what other models know. For Applica’s TILT, language is the grammar and vocabulary of business documents—this is what our technology is pre-trained on and it is what it has perfected, both as input and as output. By this we mean words, sentences, layout, and information structures relevant to contracts, loans, customer surveys, court papers, official documents, affidavits, and so on. Our model is not simultaneously versed in rap lyrics and political trolling on social media, say, so there’s nothing it needs to “unlearn” when it’s deployed to handle a company’s paperwork. Other models, understandably concerned with natural language in general, needlessly recognize and produce communication that has no place in the kinds of documents our clients are automating.

One of the breakthroughs we achieved with TILT concerns the way it can comprehend documents of various types, regardless of whether they contain structured data, unstructured text, semi-structured information, or a combination of the above. This is thanks to features that allow for the simultaneous recognition of linguistic, graphical, and structural elements. In contrast, models that apply deep learning to predictive text generation alone are typically weak or inapplicable when it comes to the non-linguistic information present in most business documents, such as infographics, checkboxes, and signatures.

Another important feature unique to TILT versus other deep learning language models concerns its agility and usability. A much, much smaller model—owing to its specialized, laser-focused applications—TILT is significantly less expensive to run in terms of computational power. This affects cost of service, as well as accessibility and speed. It has always been the Applica mission to create a solution precisely calibrated to what our clients need. Quite simply, we set out to create document automation technology that performs every function that’s necessary and sufficient in our field of expertise—and none that aren’t. Training TILT on business documents only is one way we narrow the scope of our engine to stay specialized. Enhancing real-world performance with each wave of user feedback is another. Many rival companies are in deep learning R&D for a lot of different industries, reasons, and need states, so naturally they can’t afford to narrow their search to vital breakthroughs in document automation.

And speaking of R&D, for many players in our sector the scales are significantly tilted in favor of development, with research and afterthought or outright a non-starter. Our company, on the other hand, has always attracted inquisitive and analytical IT minds whose early careers were often in academia. And while we agree that pragmatism is necessary to problem solving of any kind, we’re also convinced that theory matters. Because without deep analytical research, how do you know that the problem you’re solving is the right one? Or that your solution really is the best in its class? That’s why we make sure our R&D team has dyed-in-the-wool researchers in addition to enthusiastic, imaginative developers—our developers move fast, and our researchers make sure they don’t break things. And we’ve structured the day-to-day work at our company in a way that allows people to share ideas and requires that every viable idea is explored by the people whose thinking it ignites.

The research directions we are currently pursuing include multimodality, zero-shot deep domain adaptation, chain-of-thought prompt engineering for emerging large language models, and the always current matter of data accuracy and quantity. And because we consider document processing to encompass every step from data extraction and document sorting all the way up to the decision making required for business tasks, we’re busy researching the research possibilities in a range of areas. Anything that helps solve existing and emerging document-related problems for organizations—that is what we’re following, studying, questioning, and helping to refine.

Interested in finding out how Applica can apply deep learning to solve your organization’s document-related challenges? Contact us for a demo today.

Continue reading

A Deeper Look at TILT – Applica’s Revolutionary Deep Learning Tech
We’ve figured out a way to give our clients more speed, more precision, more control, and more options, while again setting an entirely new standard in the document automation game.
Improving Work-Life Balance With Extra Paid Leave
Applica’s team members are everything to us, which is why we trust them to take the time that they need.
Overcoming Limitations With Legacy Automation Technology
In the world of document automation, legacy solutions don’t fulfill present-day data needs. If you want progress, you need to think differently.