A Digitally Trained Classifier correctly identifies the document type (invoice), while Anchor-Based extraction is ideal for semi-structured documents. It finds key labels (like ‘Invoice Number:’) and extracts data relative to them, making it resilient to positional shifts. This balances accuracy and maintainability better than rigid templates (C), a fully generic AI Skill (A), or text-only rules (D) which may fail with similar labels elsewhere in the document.

A core principle of the Document Understanding ML lifecycle is continuous learning. Human validations serve as ground truth data. When these corrections are exported and used to retrain the extraction model (e.g., a Digitally Trained classifier or an Anchor-Based model), the system’s accuracy improves over time, reducing future validation effort.

Scanned documents are images to a computer. Optical Character Recognition (OCR) is the essential step that analyzes these images and outputs the text and its position. If OCR is not performed or fails, the subsequent extraction techniques (text-based, anchor-based) have no text data to work with, leading to poor accuracy regardless of how well the extraction scope is configured.

The Digitally Trained Document Classifier’s role is to analyze an input document and categorize it (e.g., as ‘Digital Claim Form’ vs. ‘Scanned Handwritten Claim Form’). This classification can then be used in the workflow (e.g., within a Switch activity) to route the document to different processing branches, perhaps applying OCR only to the scanned branch or using different extraction strategies.

The ‘Number’ data type in a Taxonomy field is designed for this scenario. During extraction, the underlying parsing logic is robust and strips out common non-numeric prefixes and suffixes (like ‘$’, ‘USD’, ‘Dollars’, commas) to isolate the core numeric value. This is more efficient and maintainable than manual string parsing (A) or complex RegEx (C).

Anchor-Based extraction relies on finding a stable ‘anchor’ text and then locating the target relative to it. If the original anchor (e.g., old logo text) is gone, the rule breaks. The solution is to update the rule’s configuration to use a new, reliable anchor (e.g., the text ‘Contract Party:’ which is still present). This is a quick, surgical fix in the Data Extraction Scope, aligning with the maintainability of anchor-based rules.

The Document Understanding framework centers around the `DocumentData` object. The `Extract Data` activity processes this object and populates its `ExtractionResults` property with the structured results (fields, values, confidences). The subsequent `Validate Data` activity reads from this same property to display the results for human validation.

The `Validate Data` activity can be configured with confidence thresholds. By setting a high threshold, you implement a ‘human-in-the-loop’ only where needed. Documents where the ML model is highly confident (above 95%) are auto-approved, drastically reducing validator workload. Documents with low-confidence extractions (below 95%) are still sent for review. This optimizes efficiency for high-volume, high-accuracy batches.

The `Load Document` activity (or `Read PDF Text` for simpler cases) is the typical entry point. It creates the `DocumentData` object, loads the physical document file (PDF, image, etc.), and performs initial OCR if needed. This object is then passed sequentially to `Classify Document`, `Extract Data`, and `Validate Data` activities.

A Digitally Trained Classifier correctly identifies the document type (invoice), while Anchor-Based extraction is ideal for semi-structured documents. It finds key labels (like ‘Invoice Number:’) and extracts data relative to them, making it resilient to positional shifts. This balances accuracy and maintainability better than rigid templates (C), a fully generic AI Skill (A), or text-only rules (D) which may fail with similar labels elsewhere in the document.0