artefactory/Argimi-Ardian-Finance-10k-text-image

The ArGiMI Ardian datasets : text and images

The ArGiMi project is committed to open-source principles and data sharing. Thanks to our generous partners, we are releasing several valuable datasets to the public.

Dataset description

This dataset comprises 34,000 financial annual reports, written in English, meticulously extracted from their original PDF format to provide a valuable resource for researchers and developers in financial analysis and natural language processing (NLP). These reports were published from the late 90s to 2023.

This dataset provides images of each document pages. A lighter, text-only version, is also available at artefactory/Argimi-Ardian-Finance-10k-text.

You can load the dataset with:

from datasets import load_dataset

ds = load_dataset("artefactory/Argimi-Ardian-Finance-10k-text-image", split="train")

# Or you can stream the dataset to save memory space :

ds = load_dataset("artefactory/Argimi-Ardian-Finance-10k-text-image", split="train", streaming=True)

Dataset composition:

• Each PDF was divided into individual pages to facilitate granular analysis.

• For each page, the following data points were extracted:

  • Raw Text: The complete textual content of the page, capturing all textual information present.
  • Screenshot: A high-resolution image of the page, preserving the visual layout and formatting.
  • Cells: Each cell within tables was identified and represented as a Cell object within the docling framework. Each Cell object encapsulates:
    • id: A unique identifier assigned to each cell, ensuring unambiguous referencing.
    • text: The textual content contained within the cell.
    • bbox: The precise bounding box coordinates of the cell, defining its location and dimensions on the page.
    • When OCR is employed, cells are further represented as OcrCell objects, which include an additional confidence attribute. This attribute quantifies the confidence level of the OCR process in accurately recognizing the cell's textual content.
  • Segments: Beyond individual cells, docling segments each page into distinct content units, each represented as a Segment object. These segments provide a structured representation of the document's layout and content, encompassing elements such as tables, headers, paragraphs, and other structural components. Each Segment object contains:
    • text: The textual content of the segment.
    • bbox: The bounding box coordinates, specifying the segment's position and size on the page.
    • label: A categorical label indicating the type of content the segment represents (e.g., "table," "header," "paragraph").

• To guarantee unique identification, each document is assigned a distinct identifier derived from the hash of its content.

Parsing description:

The dataset creation involved a systematic process using the docling library (Documentation).

• PDFs were processed using the DocumentConverter class, employing the PyPdfiumDocumentBackend for handling of the PDF format.
• To ensure high-quality extraction, the following PdfPipelineOptions were configured:

  pipeline_options = PdfPipelineOptions(ocr_options=EasyOcrOptions(use_gpu=True))
  pipeline_options.images_scale = 2.0  # Scale image resolution by a factor of 2
  pipeline_options.generate_page_images = True  # Generate page images
  pipeline_options.do_ocr = True  # Perform OCR
  pipeline_options.do_table_structure = True  # Extract table structure
  pipeline_options.table_structure_options.do_cell_matching = True  # Perform cell matching in tables
  pipeline_options.table_structure_options.mode = TableFormerMode.ACCURATE  # Use accurate mode for table structure extraction
  

• These options collectively enable:
  • GPU-accelerated Optical Character Recognition (OCR) via EasyOcr.
  • Upscaling of image resolution by a factor of 2, enhancing the clarity of visual elements.
  • Generation of page images, providing a visual representation of each page within the document.
  • Comprehensive table structure extraction, including cell matching, to accurately capture tabular data within the reports.
  • The "accurate" mode for table structure extraction, prioritizing precision in identifying and delineating tables.

Disclaimer:

This dataset, made available for experimental purposes as part of the ArGiMi research project, is provided "as is" for informational purposes only. The original publicly available data was provided by Ardian. Artefact has processed this dataset and now publicly releases it through Ardian, with Ardian's agreement. None of ArGiMi, Artefact, or Ardian make any representations or warranties of any kind (express or implied) regarding the completeness, accuracy, reliability, suitability, or availability of the dataset or its contents. Any reliance you place on such information is strictly at your own risk. In no event shall ArGiMi, Artefact, or Ardian be liable for any loss or damage, including without limitation, indirect or consequential loss or damage, or any loss or damage whatsoever arising from loss of data or profits arising out of, or in connection with, the use of this dataset. This disclaimer includes, but is not limited to, claims relating to intellectual property infringement, negligence, breach of contract, and defamation.

Acknowledgement:

The ArGiMi consortium gratefully acknowledges Ardian for their invaluable contribution in gathering the documents that comprise this dataset. Their effort and collaboration were essential in enabling the creation and release of this dataset for public use. The ArGiMi project is a collaborative project with Giskard, Mistral, INA and BnF, and is sponsored by the France 2030 program of the French Government.

Citation:

If you find our datasets useful for your research, consider citing us in your works:

@misc{argimi2024Datasets,
  title={The ArGiMi datasets},
  author={Hicham Randrianarivo, Charles Moslonka, Arthur Garnier and Emmanuel Malherbe},
  year={2024},
}