Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Bài kiểm tra hiểu tài liệu không có OCR (bánh rán)
Kiểm tra kiến trúc không có OCR của Donut, các tác vụ được nghiên cứu, phạm vi điểm chuẩn và các biện pháp bảo vệ để kiểm tra các trường tài liệu được tạo so với hình ảnh nguồn.
Câu hỏi 1 của 8
What makes Donut OCR-free?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- What makes Donut OCR-free?
- It avoids reading text and only classifies page color
- It uses a scanner instead of a document image
- It requires an OCR transcript as the model’s first input
- It maps a document image to task output without first calling a separate OCR engine
- Which architecture does Donut use according to its documentation?
- A vision encoder and decoder that also receives a separate OCR transcript
- A vision encoder and an autoregressive text decoder
- A text encoder-decoder that reads tokens emitted by an external OCR engine
- A visual feature extractor followed by an OCR engine and a field classifier
- Which problem did the Donut paper identify as a motivation for OCR-free processing?
- Avoiding the need to define a task-specific output format
- Reducing document-processing latency by removing every decoder stage
- Errors from a separate OCR stage propagating into document understanding
- Generalizing to every language without training or evaluation
- What does a benchmark result in Donut’s original paper establish?
- Performance on the evaluated model, tasks, datasets, and test setup
- Accuracy on every document type and language
- That every fine-tuned checkpoint performs identically
- Deployment readiness for any organization’s forms
- Why verify a structured field produced from a document image?
- The decoder cannot produce unsupported text
- The model may omit, confuse, or generate a plausible value not supported by the page
- OCR-free architecture proves every field is visible
- A valid JSON schema proves values are correct
- Why can an OCR-free output be harder to debug than an OCR-plus-model pipeline?
- It does not process visual pixels
- It cannot be trained for a document task
- It always separates OCR errors into their own output file
- It may not provide a separate OCR transcript showing where text recognition failed
- Which evaluation design best supports a fair comparison and a receipt-use decision?
- Test it and an OCR-plus-model baseline on held-out representative receipts
- Evaluate Donut on representative held-out receipts and check field-level accuracy
- Compare Donut and a baseline on the same receipt images used for training
- Use results from a public benchmark with unrelated documents as deployment evidence
- What does OCR-free not mean in this context?
- That the model processes an image directly
- That it can be evaluated on document tasks
- That the document output is error-free or needs no validation
- That no separate OCR engine is required in the architecture