NCA-GENM - Data Analysis and Visualization (10% of the exam) - Section 5.3
Apply PDF extraction using OCR.
Apply optical character recognition (OCR) to extract text and structured data from PDF documents and scanned images. Recognise the additional preprocessing steps - such as deskewing, binarisation, and layout detection - that improve OCR accuracy on complex document layouts.
OCR
Practice question for this objective
lock_openFree sampleData Analysis and Visualizationeasy
A data engineer receives a batch of scanned insurance forms saved as PDF files. The text layer is absent from all files. Which processing step is required before these documents can be passed to a large language model for information extraction?
ARun a diffusion-based image enhancement model to upscale the scanned pages before ingestion.
BApply optical character recognition to convert the rasterised page images into a plain-text representation.check_circle Correct
CUse a CLIP-based embedding model to index the page images by semantic similarity before retrieval.
DParse the PDF cross-reference table to locate embedded font descriptors and reconstruct the character stream.
Identify when OCR is required to extract usable text from scanned PDF documents for LLM processing. Scanned PDFs are essentially image containers - each page is stored as a raster bitmap with no underlying text layer. Optical character recognition analyses pixel patterns to identify characters and outputs a text string. Only after OCR produces that text can a downstream LLM receive, tokenise, and reason over the document content. Digitally-created PDFs, by contrast, embed a text stream that can be extracted directly without OCR.
Why A is wrong: Upscaling improves visual resolution but does not produce a machine-readable text layer; the LLM still receives no extractable text without an OCR step.
Why B is correct: Scanned PDFs contain only pixel data with no embedded text. OCR analyses the image and outputs a character sequence, producing the text layer an LLM can consume.
Why C is wrong: CLIP enables image-text similarity search but does not extract the document text itself; downstream LLM prompting still lacks the raw text content from the forms.
Why D is wrong: Font descriptor parsing applies to digitally-created PDFs that embed a text layer; scanned documents store pages as raster images and have no font or character stream to parse.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
Apply OCR to every file unconditionally, because OCR output is always more reliable than an embedded text layer.
errorWhy it is wrong: This is the over-application the task wants to avoid; an embedded text layer from a word processor is exact, whereas re-OCRing it wastes compute and can inject recognition errors, so blanket OCR is the wrong default.
Both document types require OCR because every PDF stores its content purely as page images.
errorWhy it is wrong: Tempting because OCR is strongly associated with PDFs, but born-digital exports already carry encoded text, so not every PDF stores content as images.
Rotating each page by 90 degrees so the text runs along the longer image axis before recognition begins.
errorWhy it is wrong: Reorienting a page only matters when the text is sideways or upside down; it does nothing to separate faint strokes from a mottled background, which is the contrast problem actually described.