What Is an Extraction Schema: The Specification for What to Pull from a Document and in What Form, A Complete Guide

An extraction schema is the specification for which items to pull from a document, under what names, formats, and conditions. It is the standard that aligns extraction results with a data structure that business systems and downstream automation can use directly.

Why an extraction schema became necessary

The same value was called different names in each department.

Even as the technology for pulling values from documents improves, the results often cannot be used as they are. One department calls it the representative's name, another calls it the representative's full name. Dates mix dots and hyphens, and amounts arrive mixing won units and thousand-won units. If people must re-format the data before it can go into a business system, the savings from automation shrink.

Another problem is that the items to extract differ from document to document. A business registration certificate requires the registration number and the opening date, while a corporate registry requires the names of officers and the dates of changes. If the knowledge of what to extract from which form stays only in one staff member's experience, there is no standard against which to verify the results.

An extraction schema fixes this agreement in a document. By extracting only the required items according to defined fields and normalizing them into item-value pairs, the results become the foundation for downstream processing and automation.

The precise definition of an extraction schema: how it differs from a coordinate template

Three axes that separate a coordinate template and an extraction schema

First, what they are based on differs. A coordinate template specifies which value sits at which position on a form. An extraction schema specifies what meaning of value to receive and in what format, not where it is.

Second, the impact of a form change differs. A coordinate template must be rebuilt if a box moves even slightly. An extraction schema can keep being used as long as the meaning of the item stays the same, even after the form is revised.

Third, the scope of information held differs. A coordinate template holds only the way to find a value. An extraction schema holds the item name, value format, whether it is required, and validation conditions together, making it the standard for checking results.

With the spread of extraction based on vision-language models, values can be found by meaning rather than by specified position. That has increased the weight of the schema, which defines what to receive and in what form. Finding values has become something models do better, but how the organization receives the values found still has to be decided by the organization.

Five elements that make up an extraction schema

Five elements to record when defining one item

First, the item name. Match it to the field names in the business system, and unify the names that different departments used. Even when the same item is labeled company name, corporate name, or business name depending on the form, the result is received under one name.

Second, the value format. Define the format the result must follow: for dates, the order of year, month, and day and the separators; for amounts, the units and whether commas are used; for numbers, whether hyphens are included.

Third, whether it is required. Distinguish items that should go to exception handling when empty from items that may be left blank. When a required item is missing, also record whether to request a correction or to look for it in another document.

Fourth, normalization rules. Record the rules that convert the notation in the original into a standard form, such as converting Korean-language amount notation into numbers, or removing the corporate suffix from a company name.

Fifth, validation conditions. Include the conditions that check whether a value is correct, such as the digit count of a business registration number, the order of dates, or items that must match another document.

The document that organizes these five elements into a single table is commonly called a field definition sheet. It becomes the starting point from which developers and business staff read the results on the same standard.

How to apply an extraction schema in practice

Work backward from the input items of the business system

Rather than extracting every item visible on a document, start by checking which fields the business system that receives the results has. Extracting items the system does not accept only adds more things to verify.

Conversely, items that the system does not hold but judgment requires should be marked separately, so they are not overlooked in the later condition-judgment stage. Creating a table that maps each system field to a schema item line by line lets you find format errors early, at the integration stage.

Separate by form and share common items

Create a schema for each form, but share the same definition for items that appear across several forms, such as the business registration number or the representative. A common definition is what makes it possible to write cross-document checks as rules.

Changing the definition of a common item affects every form that uses it, so check the scope before making the change.

Manage it by version

When items are added or formats change because of a change in law or in business standards, the schema changes too. Record what changed and when as a version, compare the revised schema against the same evaluation document as before, and then roll it out to operations. Also record who requested the change and who approved it along with the version.

Keeping versions is what lets you explain by what standard past results were produced.

Extraction schema in the Korean context

Korean documents often contain numbers with a fixed format, such as resident registration numbers, business registration numbers, and corporate registration numbers, along with amounts written in Korean and dates written with dots. Items like these must have their normalization and validation rules stated explicitly in the schema for results to be consistent.

For items that are personal information, such as resident registration numbers, you must decide at the schema stage whether to de-identify them at the time of extraction or to keep the original. Only then will storage and access control remain consistent.