The Company

Skilja (pronounced skill-yah) takes its name from the Icelandic word meaning to understand or to separate. This is an apt title for a cognitive capture software company specializing in what it calls “the 4th generation of document understanding”; that is, the application of document classification and data extraction to read and understand scanned and textual documents. The company’s goal is to continue to develop AI that automates repetitive human tasks, specifically where documents are involved.

Started in 2012 and headquartered in Freiburg, Germany, the company is led by founder Alexander Goerke, who has over 20 years’ experience in the development of intelligent capture technologies. Goerke worked for early capture innovators such as Brainware and Learning Computers International, acquired by Kofax in 2006, where he used AI technology now known as “fast machine learning” back in 2002.

Skilja has technical competencies in the following areas of cognitive capture:

• AI algorithms, classification, feature extraction, and deep learning, and their application to document understanding.
• An open, service-oriented document processing platform for high volume and critical enterprise applications. The company says it has over 20 installations with processing volumes of over 10 million pages per year, the largest of which processes 180 million documents per year.

The company’s business model is to sell its software tools through partners such as OEMs, VARs, and system integrators. For example, a large IT services company in Germany uses Skilja in document capture projects; an OEM customer embeds Skilja software within an ultra-high speed scanning process; and other customers include well-known ECM and capture vendors.

The Technology

Skilja’s product line consists of three offerings that can be purchased separately or combined: Laera cognitive recognition SDKs, the Vinna document processing platform, and the new Tegra web services.

Laera SDKs

Laera comes in two flavors: Classifier and Extraction. Laera is sold as a software development kit (SDK) for other developers to embed within their products.

Laera Classifier is an image- and text-based classifier of documents and pages using semantic analysis and machine learning (ML) tools. It includes layout classification, content classification, ID card classification, photo detection, and document separation.

Laera Extraction provides content recognition and extraction for forms, semi-structured documents, and completely unstructured documents based on trainable semantic and syntactic analysis.

For data extraction, Skilja favors ML over deep learning (DL) because for business documents, the number of samples available is much smaller than what DL requires. DL also ignores standard business conventions and doesn’t allow for retraining by the customer.

For the rapid learning of the patterns found within business documents, Skilja developed a cascaded ML method it calls “Fast Machine Learning.” This works well because there is a body of pre-existing knowledge, standard syntax, and grammar, and because the semantics of business documents – e.g., headers, labels, line breaks, fonts – are already known. Skilja’s training algorithms are fast enough that they can run continuously in a background service and, based on user corrections, can improve the models on the fly. Skilja calls this “Online Learning.”

Laera Extraction uses a pretrained language model with semantic elements and entity detection. Each entity is classified using role detection, which is based on the context where the entity was found. To assign entities to fields, Skilja uses trained probability models where each entity can be assigned to one or multiple roles. It looks for logical patterns like labels and headers, semantic patterns that appear within the same sentence, and geometric patterns within semi-structured documents.
Skilja developed its own NLP methods and uses deep learning neural network methods for visual classification of graphic elements. Everything is fed back through the Online Learning service for exception handling by humans in the loop and to improve the system’s performance.