Technology and architecture
KIE Explained: Key Information Extraction for Pulling Business Data From Documents, Complete Guide
Korea Deep LearningPublished
KIE (Key Information Extraction) is a technology that picks out only the items a business needs from the content of a document, structures them as key-value pairs, and converts them into a form that can be entered directly into an existing business system.
Why KIE is needed
Reading alone is not enough.
Technology that converts document images into text has existed for a long time, and accuracy for printed text is already high enough. Yet staff workload has not dropped much. When recognition results come out as one block of characters, finding where the business registration number is and where the representative's name is still falls to a person, who must look at the page and copy the values into the system. In one overseas corporate credit review operation, reviewing, entering and cross-checking a single application took a long time for the same reason. KIE is a technology aimed at this point.
The exact definition of KIE: how it differs from OCR
Three axes of difference between OCR and KIE
First, the target differs. OCR targets the characters on the image. KIE targets the recognized characters together with the position where those characters sit and the surrounding item labels.
Second, the goal differs. The goal of OCR is to transcribe the written characters completely and accurately. The goal of KIE is to pick out the items the business needs and pair them with their values for output. On the same document, OCR gives a business registration number and the digits after it as one string, while KIE gives data in which those digits are linked as the value of a key called business registration number.
Third, the performance indicators differ. OCR is measured by character-level accuracy. KIE is measured by item-level exact match rate and the proportion corrected by people. Even with high character accuracy, if item linking is wrong, it counts as an error in business terms, so the two indicators are better viewed separately.
The two technologies are not substitutes. Characters OCR cannot read cannot be extracted by KIE either, so the right order is to build the extraction structure on top of recognition quality.
Five conditions for putting KIE into business use
Five conditions for KIE in real operations
First, the extraction items must be defined on the business's terms. The standard is not the items the model can extract, but the items staff must enter into the system.
Second, it must hold up when forms change. Even the same application has different layouts depending on branch and year, so methods that fix coordinates do not last long.
Third, it must be able to handle documents that mix handwriting, checkboxes and stamps. In one card issuer's validation, high key-value extraction accuracy was confirmed for both printed and handwritten text, which means this condition was met.
Fourth, it must be possible to trace where an extracted value came from. If you cannot confirm which location in the original a value came from, the reviewer ends up reading the whole document again.
Fifth, validation rules must run alongside. Conditions the business requires, such as the order of dates, the total of amounts, and matching names across documents, are set as rules rather than left to the model.
Practical application of KIE
Item definition comes first
The first deliverable of a project is not the model but the item definition sheet. In one bank's damage-relief automation, defining which items to extract, and in what format, for each document came first. The documents included the damage-relief application, the payment suspension request, and the incident confirmation certificate. Without this document, the criteria for evaluating performance do not exist at all.
Separate classification and extraction in the design
In an environment where several documents arrive merged into one file, the stage of splitting documents and identifying forms is needed first. The reason one card issuer's case set up the initial document groups and classification code system before moving on to extraction lies here. Item extraction works stably only after classification accuracy has been secured.
Run validation rules alongside extraction
If extracted values go straight into the system, errors are found later, downstream. Filtering them with rules at the moment the values come out reduces the number of cases passed to staff.
Rules fall into three types. The first is format validation. It checks items whose shape is fixed, such as the digit count of a business registration number, the format of a date, or the separator in an account number. The second is relationship validation. It checks whether the application date is not earlier than the issue date, and whether the total of amounts matches the sum of the item values. The third is cross-document validation. It checks whether the representative's name on the application matches the one on the supporting documents, and whether the account number matches the bankbook copy.
The third type is especially important. Financial screening work is not finished with a single document, because a case is completed only when several documents are matched against each other. In one bank's damage-relief automation, cases where inconsistencies between documents were found automatically through this matching were confirmed.
Manage extraction items by version
The item definition sheet is not finished once it is written. When forms are revised or business rules change, the extraction items change too. If the sheet is overwritten at that point, what changed, when, and why disappears.
The practice used is to create a new version each time a change occurs and keep the previous versions. Record whether items were added, whether format hints changed, and whether subitems were grouped, together with a note. Then compare performance with the previous version on the same evaluation documents, and apply it to operations after an approval process.
With this procedure in place, if a change lowers performance it can be reverted immediately, and even when staff change, the reasons for the settings remain in the document.
KIE in a domestic environment
Korean financial and public organizations are often subject to network separation rules that limit external API calls. So whether an on-premises deployment is possible becomes the first gate of the technical review. In addition, you must handle HWP (Hangul word processor) documents, low-resolution scanned copies received by fax, and the table structures peculiar to government forms. Vendors that have repeatedly trained and validated on domestic documents have an advantage on these points.
Frequently asked questions
If OCR accuracy is high, is KIE automatically accurate too?
No. OCR accuracy only sets the upper limit for KIE. Even if characters are read correctly, linking them to the wrong item counts as an error in business terms.
Can handwritten items be extracted?
Yes. In one card issuer's validation, key-value extraction accuracy for handwritten text was confirmed to be high. However, for items with large variation in handwriting, it is safer to also set validation rules.
Do we have to retrain every time the form changes?
No. Coordinate-based rule methods are fragile to form changes. Methods that interpret the structure and context of the document, however, find values by item name even when the layout changes.
How does the extracted data get into existing systems?
Data that has been extracted and validated is loaded into the electronic document management system, then registered in the core business system through an automation tool. Without this connection, staff would have to enter the results again.
What happens when the number of extraction items grows later?
Add them to the item definition sheet and save as a new version. Compare performance with the previous version on the same evaluation documents, and apply it to operations after an approval process.
Can items spanning several documents be validated?
Yes. Matching the representative name or account number between the application and supporting documents automatically is used in actual screening work.
Can it be used in an air-gapped environment?
Yes. With an on-premises deployment installed on a customer's internal server, it operates without external communication.
Send just one document, and we will reply with the actual extraction result as it is.
Contact us



