Sources Sought-Information Extraction.pdf

Cleared for Public Release

Notice Information

Notice type: Sources Sought

NAICS Code: 541 – Professional, Scientific, and Technical Services/541511 – Custom Computer Programming Services

Agency/Office: Defense Threat Reduction Agency (DTRA)

Title: Information Extraction from Scientific Documents

Description: This is a Sources Sought Notice; there is no solicitation available at this time. No response will be provided to requests for a solicitation. THIS SOURCES SOUGHT IS PUBLISHED FOR

MARKET RESEARCH PURPOSES ONLY.

DTRA requires Artificial Intelligence and Machine Learning (AI/ML) models and pipelines capable of identifying and extracting elements of scanned technical and scientific documents (e.g., tables, plots, photos, etc.). These elements are typically present among structured and unstructured text and in single-or multi-column documents. Further, some elements may need additional processing beyond extraction (e.g., transposition of tables, identifying explicit and implicit axis values in plots, etc.). Every step in the pipelines – identification, extraction, and any additional processing – must be automated, system agnostic, containerized, and capable of being hardened according to DoD and NIST guidelines. DTRA requires the models and pipelines to be immediately deployable or very close to deployable; the Agency is not interested in early research efforts.

Examples of Desired Outputs

1. Tables: Identification and extraction of entire tables, including their associated titles, captions, and all entries within the table. Subsequent processing on the extracted table should identify and extract the title, caption, headers, entries and their numerical or textual formats, etc., separately and store them in a machine readable and structured format as well as described in a markup language to best allow the recreation of the original table using commonly available software.

2. Plots: Identification and extraction of entire plots, including the title, caption, axes, key(s), and data (curves, points, bars, etc.) within the plot. In all cases, subsequent processing should identify and extract the title, caption, axes, and key(s) separately, and make them available in a machine readable and structured format as well as described in a markup language (where appropriate) for use with commonly available software. Additional processing will depend on the nature of the plot. For example:

Histograms or Scatter plots: Additional processing will identify and extract information to allow recreation of the histogram or scatterplot using commonly available software.

Plots with a single curve: Additional processing should identify and extract a sampling of the curve sufficient to recreate its features.

Plots with multiple curves: Additional processing should identify and extract samples of each curve in the plot sufficient to recreate the curve’s features.

The output of any additional processing should be in machine readable and structured format as well as described in a markup language (where appropriate) and be sufficient to allow the recreation of the original table or plot using commonly available software.

DTRA welcomes information on other immediately available AI/ML models and pipelines capable of similar processing on other elements of technical and scientific documents.

Additional Requirements

In addition to the capabilities listed above, the AI/ML models and pipelines shall have the following attributes:

1. Can ingest documents in common formats (e.g., Portable Document Format, pdf).

2. Provide outputs in standard formats suitable for use in a modern software pipeline (e.g., json, xml, markdown, etc.).

3. Can work with digitized documents that have artifacts that degrade readability.

4. Be written in a common modern computer language (e.g., python) and run on common computer hardware (GPU-accelerated architectures, included).

5. Are designed and developed for interoperability with other software written in modern computer languages (e.g., python).

6. Are modular in design and pipeline and system agnostic.

7. Are stable for sustained use over long periods of time/many documents.

8. Are well documented, with clear and detailed explanations of their design, implementation, and use, and are suitable for and accessible to software engineers and data scientists.

9. Perform all computations locally; no access to any off-premises resources shall be allowed.

10. Able to be independently operated and fine-tuned; no ongoing relationship with the offeror or any other entity shall be required.

11. The model training, validation, and testing data must be available for inspection by the govern-ment.

Intellectual Property

If a solicitation is made, the government would prefer unlimited data rights to the code, models, docu-mentation, and data used for training testing, and validation, but proposals and an awarded contract that deliver these items with fewer rights would be considered.

Additional questions

As part of your capability submittal, please provide responses to the following:

1. Describe your history of developing AI/ML models and pipelines.

2. Describe the current state of the AI/ML model(s) and pipeline(s) that you believe meet the criteria of this Sources Sought notice.

3. Provide measures of extraction accuracy for the AI/ML model(s) and pipeline(s) you believe meet the criteria of this Sources Sought notice.

Sources having the ability to meet the requirements described above are invited to respond to this notice.

Responses should be limited to a ten (10) page white paper in 12-point Times New Roman font, not in-cluding cover page and cover letter. Any proprietary information should be clearly identified. Submitted data and information will not be returned. Any information submitted in response to this Sources Sought Notice is strictly voluntary. This notice is for information for planning purposes and the Government will not pay or otherwise reimburse respondents for information submitted.

Electronic responses are required no later than 11 July 2024 at 16:00 ET and shall be submitted to Ms. Mary Kay Chase, Contracting Officer, at mary.k.chase2.civ@mail.mil. All emails should include in the subject line the Sources Sought Number and the submitting organization’s name. All documentation submitted in response to this sources sought shall become the property of the Government.

Unless otherwise stated herein, no additional information is available. Requests for same will be disregarded. There is no commitment by the Government to issue a solicitation, to make award or awards, or to be responsible for any monies expended by any interested parties in support of any effort in response to this Sources Sought Notice. Information provided herein is subject to change and in no way binds the Government to solicit or award a task order.

Contracting Office Address:

DTRA AL-ACS

8725 John J. Kingman Road Fort Belvoir, Virginia 22060-6201

Secondary Point of Contact:

Mary Kay Chase mary.k.chase2.civ@mail.mil