Skip to main content
In this lesson we introduce information extraction: the process of converting useful data from unstructured content into structured, machine-readable formats that applications can query, validate, and analyze. Organizations receive many forms of unstructured content every day: reports and PDFs, invoices and receipts, business cards and identity documents, photographs or scanned notes, and even audio or video. Humans can read these easily, but unless the content is transformed into structured fields and values, software systems cannot search, filter, or compute over it.
A presentation slide titled "What Is Information Extraction?" listing five types of unstructured content—documents/PDFs, receipts/invoices/forms, business cards/ID, images with text, and audio/video—that can be transformed into structured data. A footer notes this content can't be queried, sorted, or totaled while it remains unstructured.
Why this matters
  • Structured data enables querying (e.g., “show invoices over $1,000”), automation (routing, approvals), analytics (totals, trends), and integration with back-end systems.
  • Information extraction recovers the fields and values that matter to business processes so they can be stored in databases or forwarded to workflows.
Concrete example — a business card Imagine receiving a photo of a business card. A human reads it and extracts name, company, phone, and email. An information-extraction system analyzes the image and outputs structured fields (JSON), for example:
In this JSON, the fields map visual data (the printed card) to named properties applications expect. The confidence value indicates how certain the model is about the extraction. Reading vs. understanding There are two distinct stages between the original content and the structured result:
  1. Reading the content — for documents and images, this is typically OCR (Optical Character Recognition): converting visual glyphs into characters.
  2. Understanding the content — mapping extracted text to business fields, normalizing formats, validating values, and exporting to target systems.
A simple example: OCR might return the tokens “Northwind Traders”, “2026 0101”, and “20”. Mapping those tokens to meaningful fields (vendor, date, total) is the responsibility of the extraction and business-logic layer, not OCR alone.
Extraction must not invent data. It should recover and express only the information present in the source, with appropriate confidence and validation checks.
The OCR pipeline OCR is the foundational technology that converts text appearing in images into machine-readable text. A typical OCR workflow has five stages:
  1. Capture the image — obtain the document, scan, or photograph.
  2. Enhance the image — deskew, denoise, and adjust contrast to improve recognition.
  3. Detect text regions — find lines, words, and character areas.
  4. Recognize the text — convert glyphs into characters and words.
  5. Return structured results — output recognized text together with layout information (lines, paragraphs, bounding boxes).
An infographic titled "Optical Character Recognition (OCR)" showing a five-step workflow: Capture the Image, Enhance the Image, Detect Text Regions, Recognize Text, and Return Structured Results. Each step is presented in a column with a short description of the action.
Why layout and bounding boxes matter The position of text on a page often provides semantic cues — for example, totals are usually found near the bottom-right of invoices. OCR returns text + layout (bounding boxes and line order), which downstream extraction models use to associate values with fields reliably. OCR stages at a glance Follow a page through OCR
  • Captured image — may be skewed, noisy, or low contrast.
  • Enhanced image — preprocessing corrects orientation and improves legibility.
  • Text detection — boxes identify regions containing text.
  • Text recognition — characters are converted to machine-readable text.
  • Structured result — OCR returns the text and layout data needed for downstream field extraction.
From OCR text to business data After OCR produces machine-readable text and layout, field extraction gives business meaning and maps values to a schema. Typical stages:
  1. Read OCR results — use text and layout information.
  2. Identify business values — detect names, dates, totals, invoice numbers, addresses, etc.
  3. Map fields to a schema — assign values to defined fields (for example, map 20 to amount).
  4. Normalize values — convert dates, currencies, and phone numbers into consistent formats.
  5. Export structured data — send normalized data to databases, APIs, or automation workflows.
Common field types and normalization examples Example: sentence → business data Document text: “Mark spent $20 on January 1st, 2026.”
  • OCR output: “Mark spent $20 on January 1st, 2026.”
  • Identify key fields: payee = “Mark”, amount = “$20”, date = “January 1st, 2026”.
  • Map to schema: payee = "Mark", amount = 20.00, date = "2026-01-01".
  • Normalize: amount as decimal, date in ISO format.
  • Export: store or forward the structured record.
Putting it together: typical system architecture
  • Input: images, PDFs, or multimedia.
  • OCR layer: converts visual text to strings + layout.
  • Extraction layer: ML rules or models identify fields and entities.
  • Normalization & validation: format conversion, currency detection, phone normalization, confidence thresholds.
  • Export/integration: database, CRM, ERP, or automation pipeline.
Key idea Information extraction converts content that is difficult for machines to process into structured data that applications and databases can understand and use. It combines OCR (reading characters and layout) with field extraction and normalization (giving business meaning).
OCR extracts characters and layout. Field extraction assigns business meaning, maps values to a schema, and normalizes formats so applications can consume the data.
Links and references Now that you understand the concepts behind information extraction, you can explore practical tools and implementations that perform OCR, field extraction, and data normalization for real-world documents.

Watch Video