March 12, 2021

Introducing TILT

Good news world! 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. This time, our pioneering contribution applies to Applica’s proprietary template-free layout-aware technology to structured text, for unprecedented ease, speed, and accuracy in the automation of these types of documents.

When our company was formed, we set out to harness the potential of deep learning to automate unstructured and semi-structured documents, such as contracts and financial reports, which had long been roadblocks to machine learning-based solutions. For decades, this had been the area no one wanted to touch, because the returns seemed far off and the technology seemed impossible. But then the team at Applica decided it was time to challenge the predominant rule-based and machine learning-based AI paradigm.

Future-Proof Technology

Thanks to a tireless, insightful, and driven R&D team – and owing much to the way we swore allegiance to one specialized area – we achieved our goal of giving clients a disruptive advantage where there hadn’t been one before. Today Applica makes it possible to extract information from documents that feature unpredictable content and running text. More importantly, we make this process fast, reliable, lucrative, scalable, and operable by end users who don’t need to be engineers.

In fact, the Applica solution to the unstructured text problem has proven so worthwhile and usable, that it has cast a dark shadow on the existing models for automating the processing of structured documents – forms, mainly, and other text-based data sources that tightly define what goes where. While so-called AI-based technologies claiming to do this have existed for years, they typically deploy rather simple template-based solutions. Among the drawbacks to such approaches are the ways these systems (a) rely on IT staff to handle every new form and update and (b) require prohibitively large sets of training data. For Applica, getting our R&D team to develop a future-proof alternative to these antiquated methods was a natural next step.

Currently, our new model is described by the acronym TILT – short for Text-Image-Layout-Transformer – which highlights the way this new approach has an unprecedented ability to “read” graphical embeddings. Having long eluded other AI solutions to text scrutiny, graphical embeddings (like signatures), have been the holy grail along the path to full-spectrum document automation. Thus, it is headline news that Applica’s team managed to scale this peak. Another distinctive advantage of TILT is the model’s generativity, which maximizes the utility of each piece of training data. This is what allows the machine to learn on a minimal number of samples. And the less data the solution needs to see, the less time, effort, and cost is required for setting up every operation. Fittingly, the acronym TILT also communicates a shifting of perspective. After all, we really have “tilted” the technology to strategically expand the field of possibility for the entire sector.

Engineers, Linguits, Polymaths

If our business model had us selling high-priced labor in addition to priceless tech, we could offer the advantages of TILT right away, because skilled engineers on our end are already able to set up the AI to perform any required task. But our mission includes putting the tech directly in the hands of our clients. This is why we’re still hard at work integrating all of the benefits of our new solution into Applica’s product – so that when we launch, we launch correct.

For a tech company that is simply riding the wave of a market trend, R&D is a box to tick, but never the heart and soul of the operation. For Applica however, research and development is the most crucial part of shaping an entirely new category of next-age services. Our researchers, engineers, linguists, coders, and creative polymaths are the central nervous system of our company. Figuring out how to package and sell what they do is an important second step, but make no mistake about it: that’s step two. In the case of TILT, getting the front end to match the back end is now a matter of a couple swift months. The solution you won’t believe will be live in no time. We’ll keep you posted.

Outcompeting the Competition

And in terms of external validation, Applica’s TILT model beat out all other AI solutions (including tech industry behemoths) in the recent Document Visual Question Answering Challenge and also took the top spot in the Infographics VQA Challenge.

To get an in-depth understanding how TILT simultaneously learns layout, visual aspects, and textual semantics without relying on human-made heuristics, read our latest paper. And then watch this short video to see how TILT can answer user questions without any prior training.

Interested in learning how Applica is solving business problems that no one else can? Contact one of our experts today.

Continue reading

Articles
Gaining the Loss Run Advantage
Streamline and expedite analysis of loss runs across the insurance industry.
Research
MLOps - Real-World Application
In the first part of this article, I will try to present and discuss the steps needed to prepare a machine learning model for deployment in production. The second part will cover deployment and automation of the ML system creation process.

MLOps - Real-World Application