Belge Yapay Zekası Sınavı

Testi çözdüğünüzde cevaplar ve açıklamalar görünür.

1. A scanned invoice has no usable text layer. Which step can make its characters machine-readable?
   • Only changing the PDF filename
   • Only adding a category label
   • Only searching its metadata
   • Optical character recognition, with later verification
2. A PDF has a text layer but two columns read in the wrong order. What is still needed?
   • Layout-aware interpretation of reading order
   • Another OCR pass guarantees the order
   • A larger font in the output
   • Only the document’s page count
3. A table column says “Revenue, thousands of USD” and the row value is 250. What is the interpreted USD amount?
   • 250 USD regardless of heading
   • 25,000 USD by default
   • 250,000 USD if the heading applies to the row
   • The number cannot have units
4. An extracted total is a valid number but has a misplaced decimal point. Which check can catch it?
   • Check only the numeric JSON type
   • Compare with the source and related arithmetic
   • Trust OCR confidence as proof
   • Ignore the source once parsing succeeds
5. A required field is absent from a form. How should the system represent it?
   • As missing or explicitly unknown, not a plausible invented value
   • Guess from a similar prior form
   • Fill it with the most common value silently
   • Treat absence as evidence of zero
6. Which record helps a reviewer check an extracted value?
   • Only a final summary with no source link
   • Only the model checkpoint size
   • Only the file upload timestamp
   • Page number, source text and bounding location
7. Why does high OCR confidence not prove a complete extracted record is correct?
   • OCR confidence guarantees all table structure
   • Layout, row association and meaning may still be wrong
   • Every recognized word is a verified fact
   • The record has no fields
8. What privacy scope should a document AI pipeline include?
   • Only the original PDF, not extracted data
   • Only the screen used to view the file
   • Original files plus derived text, embeddings and retention paths
   • Only the person who uploaded it