Blog 1

How pre-trained AI models can help with data extraction

There are many situations where structured and unstructured data from documents need to be put into a system or used in a process when firms adopt automation solutions to digitize their workflows and processes. Typically, this is accomplished through manual transcribing or the use of an extraction tool that relies on the creation and maintenance of document layout templates that map out essential information locations.


Think about the visible barriers

Let’s take a look at an invoice and break it down into visual chunks. The invoice will usually include a variety of items, such as:

  • Addresses
  • Dates
  • Order numbers
  • Graphics
  • Tables
  • Table elements

We can immediately recognize these blocks and discover distinct patterns linked with each of the items when looking at an invoice. The format of a “date” block, for example, differs from that of an “address” block. These factors must be dealt with differently. Table elements, for example, require a separate action than number fields or pictures.

This identification is required and raises the need of applying certain implied logic and rules for the elements to be appropriately extracted, grouped, and the information organized. Consider a situation where a description spans numerous lines or serial numbers are included in the columns. To properly arrange and retrieve information in a usable manner, the first step is to grasp what an object is.

We can locate and identify each of the different elements on bills by using a multi-layered deep learning model to train. Each layer of the model feeds into the next based on the block’s and contents’ properties until the block type is determined. The benefit of this technique is that the blocks may be moved about and placed wherever on a document, regardless of their position.

Other approaches, particularly those based on document coordinate maps, may be able to adjust for minor alterations in an object’s position caused by scanning skews and shifts. However, as a vendor updates a layout to make it look more current and relocates field positions, the maps become obsolete and must be rebuilt. The scope of maintaining those document maps is substantial, given that clients can have hundreds or thousands of providers.

That difficulty is solved by computer vision, which is document agnostic. Once a computer vision model has been sufficiently trained on an element, it can be placed anywhere on the page. Because it isn’t linked to any layout, the model can be used with a variety of document formats.


Extracting Value

Let’s have a look at the problems using an invoice field: invoice date. We get basic optical character extraction results and probability scores for the output when the date field is submitted to a commercial off-the-shelf OCR engine. The numerous returning result combinations can then be submitted into an AI model that determines the final output because the element block is a “date” field.

What is the subset of valid alpha-numeric characters? There are a finite number of methods for generating a date field. This information is “learned” by the model through model training, and it can determine the valid possibilities with a high degree of accuracy.

Then accomplish this type of analysis, employing a machine learning (ML) model is preferable to using a set of simple rules or applying regex patterns and validation tests. If we were to implement equivalent capabilities using rules and patterns, there might easily be hundreds of checks to guarantee that American date conventions are properly covered.

The number of rules and patterns quickly grows when the issue set is expanded to include additional languages, countries, and norms. As a result, the strategy of employing them is unsuitable for dealing with the issue. Using AI, on the other hand, all that is required to cover additional languages and nations is to update the model with fresh samples from those languages and countries.

When it comes to AI model training, the bigger and more complete the data set, the better the findings or outcomes. This task is frequently very technical and comprehensive, and it necessitates the use of a data scientist or data analyst to guarantee that the data set is accurate, impartial, balanced, and properly labeled. To achieve robust results, the models employed in the IQ Bot auto-extraction capability are trained using over a million data points.

Obstacles are reduced by having pre-trained models packaged for users, which speeds both data and business flow.


Share This

Maxime Barber

Maxime Barber

Share your details and we’d get back to you.

1About You
2About Company