PubTables-1M: One Million Tables for Document Extraction and Structure Recognition

PubTables-1M is a dataset for evaluating object detection and image-to-text models on table extraction from unstructured documents. It was introduced by Smock et al. and converted to Hugging Face format with an Optimized Table Structure Language (OTSL) by the docling-project. The dataset page was last updated on 2023-08-31.

Use Cases

• Train object detection models to locate tables within document images based on the dataset's purpose.
• Evaluate image-to-text methods for converting table images into structured formats based on the OTSL representation.
• Benchmark table extraction pipelines that combine detection and structure recognition based on the dataset's dual-purpose design.

Strengths

• Designed for the specific, dual-purpose evaluation of both object detection and image-to-text methods.
• Based on a published academic paper, 'PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents'.
• Includes an Optimized Table Structure Language (OTSL) format, as described in a related paper.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2023-08-31 16:00:24; freshness should be verified.

Provenance

Source

docling-project on Hugging Face, based on the original PubTables-1M dataset by Smock et al.

Collection Method

Likely extracted and annotated from unstructured documents, then converted to Hugging Face and OTSL formats.

Freshness

2023-08-31