Implement information extraction solutions 5.2 Extract content from documents
A company must extract labeled fields from a mix of document types, including structured…
Study guide skill: Extract information by using multimodal pipelines that combine OCR, layout analysis, and field extraction
Question B1-049
A company must extract labeled fields from a mix of document types, including structured surveys, semistructured purchase orders and unstructured letters. Which Document Intelligence model type should it train?
Select one answer.
Answer: A
Custom neural models extract labeled data from structured, semistructured and unstructured documents, so they handle mixed types. Custom template models expect static or highly structured layouts, the read model only extracts text, and the receipt model is specific to receipts.
Deep dive
How it works
Document Intelligence custom extraction models are trained on your own labeled documents. You label a dataset with the values you want extracted, train on it, and only need five examples of the same form or document type to get started. There are two types of custom extraction model, and the choice depends on how much the layouts vary:
- Custom neural: Built on deep learning models with a base model trained on a large collection of documents, then adapted to your labeled data. It supports extracting key fields from structured, semi-structured and unstructured documents, and the overview's one-line summary is "Extract data from mixed-type documents". With v4.0 it also supports signature detection, table confidence and overlapping fields.
- Custom template: This model relies on a consistent visual template. Variations in layout reduce its accuracy, so the training set should be structured documents whose formatting and layout stay the same from one document to the next.
A mix of surveys, purchase orders and letters is the textbook neural case. The docs also advise, when you're choosing between the two, to start with a neural model to see if it meets your needs.
The other options, and when you'd use them
- Prebuilt read: This is the OCR model. It extracts printed and handwritten text and detects paragraphs, lines, words, locations and languages, and it's the underlying OCR engine for other prebuilt models. Use it when you need text, not labeled fields.
- Prebuilt receipt: This model extracts key information from sales receipts, such as merchant name, transaction date, tax and transaction total. Use it for receipts only.
- Custom template: Template models suit highly structured documents with defined visual templates, such as questionnaires or applications. They support more field types and languages than neural models. Use one when every document shares the same fixed layout.
Exam tip
"Mixed or varying layouts" means custom neural, and "one fixed form layout" means custom template. "Need to know which document type first" means a custom classification model in front of the extractor.
References
Last reviewed: