Form Understanding In Noisy Scanned Documents Dataset

The FUNSD dataset provides 199 real, fully annotated scanned documents for form understanding tasks. It includes annotations for semantic entities and relationships, supporting layout analysis and information extraction research.

Use Cases

• Train a LayoutLMv2 model to perform semantic entity labeling on scanned document images.
• Develop models for relationship prediction between annotated text fields in noisy scanned forms.
• Benchmark information extraction systems on real-world, noisy scanned documents with full annotation.

Strengths

• Contains 199 fully annotated real-world scanned documents.
• Provides annotations for semantic entities and relationships, enabling complex form understanding tasks.

Limitations

• Limited to 199 documents, which may be insufficient for training large deep learning models without augmentation.
• Focuses on scanned forms, which may not generalize to other document types like reports or invoices.

Provenance

Source

Originally created by Guillaume Jaume and available via the official FUNSD website.

Collection Method

Comprises real, noisy scanned documents with manual annotations for semantic entities and relationships.

Freshness

Last updated on Hugging Face in October 2022.