An Introduction to Intelligent Document Processing

Generative Pre-trained Transformers (GPT) have brought powerful AI capabilities to the common user, making a significant impact on how enterprises handle unstructured data. One such area of interest is Intelligent Document Processing (IDP), a concept that has been around for a while now. In this article, we will introduce IDP, discussing its origins, core concepts, potential value, and practical considerations for investment as a digital enabler.

The history of Intelligent Document Processing (IDP) dates from the 19th century, with the development of Optical Character Recognition (OCR) by Emanuel Goldberg. This early form of OCR was designed to convert characters into telegraph code and had a specialized purpose.

In the 1970s, the first omni-font OCR system was created, leading to widespread use of traditional computer vision algorithms in the 1980s due to advances in scanning and optical technology. From the early 2000s, OCR capabilities evolved with the integration of recognition algorithms and machine learning models, now available through major cloud services. This represents a significant advancement in OCR technology.

Recent developments in natural language processing (NLP) and deep learning Transformer models, like Large Language Models (LLM), indicate a shift towards more sophisticated document processing tools capable of understanding data meaning beyond text and structure. 

IDP is a systematic approach to handling unstructured data within documents on large scale. It yields structured and trustworthy data that can be effectively leveraged in downstream process and reporting activities.

From a technological point of view, we consider IDP as an umbrella term for various activities related to the automated extraction, interpretation, classification, and handling of digitized documents. IDP solutions typically use machine learning (ML), deep learning, computer vision (OCR), and natural language processing (NLP) technologies to extract and interpret data from structured and unstructured content, to support automation and augmentation of business processes. 

A high-level IDP processing flow example is as follows;

• Data Ingestion: The data that needs to be processed is analyzed, explored, and ingested for processing. The initial steps of the project will start at this stage and might require some manual iterations to get collect all the right data. However, going forward this step mostly becomes an integration that sends the files straight into processing.
• Pre-processing & OCR: After the collection of the documents, they are sent through various pre-processing steps that clean the imperfections from the documents, preparing them for extracting the content via OCR and computer vision. These extract the content on character level, leveraging heuristics and models to keep high accuracy under non-ideal circumstances.
• Document Classification: Diverse document types might come in and require different handling. The classification of documents is a challenge on its own as this isn’t always clear cut. Classifying the documents helps to organize the data and makes it easier to work with downstream. For example, we might want to classify based on the language, the layout, or the content of the document.
• Information Extraction: Information is extracted from documents in what we normally refer to as “entities”. They might involve things like specific numbers, dates, or identifying key phrases or topics within a document, or even as ‘complex’ entities, mapping to ontologies within a specific domain. The goal is to turn the documents into structured data that can be used to automate or to gain insights.
• Validation: After data is extracted, it's important to make sure that it is accurate, consistent, and complete. In edge cases this might require a ‘Human-in-the-loop’ check, but overall, this consists of configurable business rules that are executed on the now structured data set. We want to ensure that we can trust data before it gets ingested in downstream systems for insights and automation.
• Data Enrichment: As an advanced case of validation, it can also be enriched and crosschecked with live internal systems, business rule engines or models. This typically involves sending the extracted data to an internal or external API endpoint, including the response in the final processing result.
• Integration: The result of the structured and ‘interpreted’ document is finally sent to the relevant system(s) for further processing. It may trigger further (automated) business processes, or feed reporting and insight flows. Information that was previously manually used for a very specific purpose, can now be leveraged throughout the company.

For some of the steps mentioned, a ‘human-in-the-loop’ flow might be required. This is typically the case for activities that require expert decision-making or interpretation. Machine learning models are typically used here, validating and correcting these models requires a close feedback loop that feeds those validations and corrections back into the model. This enables continuous learning, required for higher accuracy in the model, which should lead to a higher level of trust in the solution.