OCR skill

New export formats in OCR skill and Output activity

OCR skill and Output activity have new export formats:

• HTML – preserving text order for use in OCR in further extraction scenarios
• ALTO – with Text Only and Preserve document structure options
• PPTX

New PDF export options

PDF export has received two new options:

• Smaller size. This is the default option for export to PDF. With this option selected, image compression is used to reduce PDF size for faster opening and long-term preservation scenarios. This setting provides optimal balance between size and quality of images.
• Maximum quality. Using this option selected, the output PDF will contain an image of better visual quality, but the file size will be bigger.

Arabic support

Vantage 2.4 offers improved support for the Arabic language. The following Vantage components fully

support documents in Arabic:

• Classification skill and Classify activity
• Extraction Rules activity
• Fast Learning activity and Online Learning
• Manual Review client

Technology Core Version 2.4

Technology Core is a set of technology-related workers for OCR, classification, extraction, and NLP.

This version of Vantage is the first to use Technology Core version 2.4. Each skill is compatible with only one Technology Core version, which is used to train the skill. At the same time, Vantage can operate using several Technology Cores at once. For example, this version of Vantage uses both Technology Core versions 2 and 2.4, meaning it can run skills intended to work with both Technology Core versions.

Skills developed in previous Vantage versions will continue to work in Vantage 2.4 using Technology Core version 2 by default. However, these skills will not benefit from the new functionality and technology advancements that have been added in the Technology Core version 2.4 version until they are upgraded to a new Technology Core version.

Manual Review

Show documents with errors on manual review

Documents with errors are now designated with special icons depending on the type of error. Documents with validation errors are marked with a red flag, and documents that contain low-confidence characters are marked with a red dot. When a task is received for verification, the first document containing errors will be opened.

Cursor mode for unstructured documents

There are three methods to select a region that contains a field value.

The first method is to hover over a word and click on it. This will create a region and copy the text to the field.

The second method is to draw a rectangle around some words. All the words inside this rectangle will be copied to the field.

In Vantage 2.4, a third method has been added. Now you can select an interval of text by clicking on the first word of the interval and, while holding down the left mouse button, dragging the cursor to the last word of the interval and releasing the mouse button. This will create a region, or several regions, containing the selected phrase and copy the text to the field.

This third method will work for documents processed with Document skills that contain NLP activities (e.g., Lease Agreement US).

Mobile Input

Vantage Mobile Input provides a simple and user-friendly mobile experience to capture and send

document images for further processing in Vantage. Depending on the document type expected, Vantage Mobile Input adjusts its appearance and gives users hints on how to capture documents with the best accuracy. Finally, Vantage Mobile Input provides an auto-capture feature for taking a picture of a document automatically to achieve the best quality of the image.

Customizing the interface to capture a specific set of documents

There are scenarios where a specific set of documents is expected to be processed. For example, an application for opening a new bank account requires a set of supporting documents to be uploaded. Earlier, when such documents were uploaded using Vantage Mobile Input, developers had to give users additional instructions on what specific documents should be captured because Vantage Mobile Input allowed uploading any number of any sorts of documents.

Vantage 2.4 allows skill designers to specify what documents should be provided to start a transaction. It is possible to specify the structure of documents to be captured: document types (ID cards, passports, or A4 documents) and names, page numbers, and page names. The Vantage Mobile Input interface will be customized according to these settings to show the user what documents should be captured. Additionally, Vantage Mobile Input will not allow the user to upload any documents and start the transaction until all the required documents have been prepared.

Monitoring and Reporting

Warehouse for Quality Analytics

Vantage 2.4 now includes a Warehouse for Quality Analytics, designed to enhance your document processing efficiency. This feature offers key insights, allowing you to pinpoint and address areas that require manual effort.

Listing tenant transactions using the Vantage API

Some Vantage users build their tools and dashboards for monitoring Vantage operations to integrate them into corporate tools and quickly react to events requiring attention. They need to monitor the system performance (i.e., the number of documents and transactions processed within a period) and the queues in the system (i.e., the number of documents and transactions that are still in progress or require manual review).

Administrating Vantage

Restricting access to a Vantage tenant by IP address

In high-security situations, it may be required to restrict access to Vantage by allowing connections only from a safelist of IP addresses.

Vantage 2.4 allows Tenant Administrators to specify a list of IP addresses that can access their tenant. Users with IP addresses not in the safelist will not be able to access the tenant.

Deep Learning AI

Hypothesis filtering for greater control over AI output

Machine learning allows the program to achieve very high quality of field extraction; however, in some cases, it is necessary to control the neural network’s output.

Output control is essential when the neural network produces entire words, but you only need specific parts, or when you need to filter out accidentally captured noise. It can also be useful for identifying parts of larger fields, such as addresses, which may be overlooked by the neural network. Additionally, it enables you to choose the best hypothesis for different occurrences of the same value. For instance, when a vendor name is printed multiple times on a document, you can select the most accurate extraction from among the various instances.

Vantage 2.4 allows you to control NN output by filtering hypotheses generated by the neural network. It comprises one or more Deep Learning activities and an Extraction Rules activity, which has access to all neural network elements.

Sample scenario of working with hypotheses

Often the neural network will find entire words and you will want to get rid of the redundant word parts. This often happens if a number and unit of measurement are written close together, or if only the part of a number without the postfix is needed. In the example below, the network has found the entire number.

Automatic division of document sets

Deep learning always involves working with large volumes of documents. You can create different collections of documents for testing and training. Large document sets can be automatically divided in a certain ratio.

Importing/exporting labeling to a folder

Another useful feature added in Vantage 2.4 is the ability to export labeled documents to a folder and load them into another or the same skill. This allows you to reuse labeling obtained during the input process or labeling created in a similar skill. If you are a skill developer, you can send your skill to production without documents, storing the set of training documents locally. To export a document set, select it in the list, click on the vertical dots and select Export Set with Labeling. To import documents, use the Upload > Import Labeled Documents from Folder... command.

Importing/exporting semi-structured activities in JSON format

To import or export an Extraction Rules activity in JSON, open the extraction rules editor and click on the vertical dots next to the elements and fields.

Improving classification by company with regular expressions and keywords

The Classify by Company activity has received the ability to complement company searches with regular expressions and keywords. You can search for TaxID, National TaxId, and IBAN fields. Regular expressions and keywords can be added when editing the mapping in this activity. Values for regular expressions can be entered directly, or via a parameter, which makes it easier to edit and reuse search settings.

Deep Learning for Segmentation

Vantage 2.4 includes a new enhanced mode, called “Thorough,” for the Segmentation activity. This mode trains a Deep Learning LLM-equivalent model to find relevant paragraphs.

We recommend using this mode if:

• You have more than 150 documents in the training set.
• There is a lot of variability in how segments look.

This mode is only supported for the English language.

NLP Extraction Rules

Vantage 2.4 now includes NLP Extraction Rules. This feature is designed for handling complex text-based documents and allows combining any of the NLP capabilities available in Vantage Advanced Designer.

NLP Extraction Rules have three primary objectives:

• Put AI under control. By allowing limitations within document contexts, this tool enables skill designers to filter and correct incorrectly extracted values, putting AI under rigorous control.
• Overcoming training limitations. This feature proves useful in scenarios where documents lack clear segments or require grouping of extracted values. It uses previously extracted information as input fields, utilizing their relative positions and textual context to refine grouping logic.
• Efficiency with minimal data. Even with a handful of documents, NLP Extraction Rules allow quick setup of extraction logic, ideal for Proof of Concept (POC) stages and for automatically labeling documents to create training sets.

NLP Extraction Rules offer the following capabilities to set up extraction logic:

• Use previously extracted information as input fields to combine the outputs of various activities. This could include fields extracted through other NLP activities.
• Pre-trained Named Entities (NER) provide a solid foundation for extraction logic, enhancing the accuracy and relevance of the data extracted.
• NLP Extraction Rules can search for static text, values from dictionaries, and patterns defined by Regular Expressions.
• By combining elements relative to each other, a skill designer can group and extract information more effectively.

Run tests with a set of documents to evaluate the effectiveness of your rules. Refine the logic as needed to ensure maximum accuracy of the extracted data.

NLP Language Support

Vantage 2.4 now supports the “Duration” named entity in Dutch documents.

Example: Dit contract heeft een looptijd van 10 jaar.

Vantage offers full NLP support* for:

• English
• French
• Spanish
• German
• Russian
• Japanese (preview)
• Portuguese
• Italian
• Dutch

*Full NLP support means support for Segmentation, Named Entities (NER), NLP Deep Learning, and NLP Extraction Rules activities.

Localization

This release has been localized into the following languages:

1. Portuguese (Brazilian) – Vantage UI and documentation. Advanced Designer localization will be added in upcoming releases.
2. Chinese Traditional – Vantage UI and documentation. Advanced Designer localization will be added in upcoming releases.
3. Italian – Vantage UI only.