Extraction

⏱ ~15 minutes📖 Theory🎯 Beginner–Intermediate

🎯 Learning objectives

  • Explain how data extraction retrieves field values from a classified document
  • Describe how confidence scores and thresholds drive human-in-the-loop validation
  • Walk through the full end-to-end DU pipeline from digitization to export

Data extraction

Extraction retrieves specific field values from the classified document. The Data Extraction Scope activity in Studio orchestrates the extraction process using the configured extractors.

Validation & human-in-the-loop

Not all extractions will be 100% accurate. UiPath DU uses confidence scores to determine when human review is needed.

How confidence scores work

  • Each extracted field gets a confidence score from 0.0 to 1.0 (0% to 100%)
  • You configure threshold values — e.g., "automatically accept if confidence > 80%"
  • Fields below threshold are sent to UiPath Action Center for human validation
  • Human corrections are fed back to retrain the model (continuous learning)
✅ Success
Human-in-the-Loop Best Practice: Start with a lower automatic acceptance threshold (e.g., 90%) to catch more errors early. As the model improves through retraining, gradually raise the threshold to reduce human workload.

End-to-end processing pipeline

A complete Document Understanding automation follows this logical flow:

1
Pre-digitization Processing
Any processing before OCR — grayscale conversion, skew correction, noise removal
2
Digitize Document
Convert document to machine-readable text using the configured OCR engine
3
Classify Document
Identify the document type using a classifier
4
Classification Validation
Check if classification confidence meets threshold; send to human if not
5
Pre-extraction Processing
Document splitting (multi-document PDFs), page selection
6
Data Extraction
Run configured extractors to pull field values
7
Human Validation
Route low-confidence extractions to Action Center for review
8
Extractor Feedback (Learning)
Use corrections to retrain the model
9
Post-extraction Processing
Data formatting, business rule application
10
Data Export
Send extracted data to downstream systems (ERP, database, API)
📌 Note
Not all steps are required for every use case. Simple processes may skip classification, human validation, or feedback steps. Design your flow based on your specific business requirements.

📝 Recap

  • Document Understanding handles both structured (fixed layout) and semi-structured (varying layout) documents, using rule-based, ML-based, or hybrid extraction depending on complexity.
  • Taxonomy is a three-level hierarchy — Taxonomy → Document Types → Fields — that defines what documents to expect and what data to extract.
  • Classification identifies the document type (keyword, ML, or Helix classifiers) before extraction begins.
  • Confidence scores (0.0–1.0) measure the model's certainty for each extracted field; scores below the configured threshold route to Action Center for human validation.
  • The end-to-end pipeline follows Digitize → Classify → Extract → Validate → Export, with human corrections feeding continuous learning.

✅ Knowledge check

1. A client sends invoices from 50 different vendors. Each invoice has the same fields (vendor name, amount, date) but different layouts. What is the best extraction approach?

2. Which of the following correctly describes the three-level hierarchy of a Taxonomy in Document Understanding?

3. An extracted "Total Amount" field comes back with a confidence score of 65%, and the project's automatic acceptance threshold is set to 80%. What happens next?