Applied AI Task Force

As a researcher who searches in Arabic, Persian, or Chinese, I want the catalog to understand my query regardless of romanization or diacritics, so I can find items I already know exist.

When a query returns no results, I assume that the library does not have an item.  

Discovery

To meet federal requirements, all public video content must have captioning that is 99% accurate. 

Captioning

A graduate student with dyslexia wants to load primary-source documents into text-to-speech software so they can access handwritten and historical materials.

A faculty member listens to a book from our collections on their drive to Princeton from Brooklyn 

Text Recognition

- The quality of text from our current OCR tool (Tesseract) is low. 

- Use vision-language models (VLMs) can significantly  improve extracted text in Figgy

- We allow staff to request text improvement, which provides data on need and delivery

- Patrons get readable text

OCR and

Handwritten Text Recognition

Letter from Hugh Simm to Andrew Simm, October 9, 1778

Tesseract:

AV ee |
CHCA Ala, |

f Sa :
<a f . : LA

m ef
Sle ed
hgh
fea 2?

Qwen3-VL:

 William Varney Dr.

    1802
    William Varney Dr.
    June 1 — to 1 Cow @ 25 Dollars
    Settled in full — Cost 15 Dollars
    [Marginal note above, partially crossed out:]
    Ct. by Cash 5 Dollars July 7
    August 19 by Cash 5 Dollars

—

 Calf Skin

    1802
    June 10 — to one Calf Skin from my father — Sent to Joseph
    — 2 shillings to him

Cook Almy Daybook

Create a system to measure the relevance of search results.

Evaluate aspects of semantic search.

- Simple embeddings

- Multilingual embeddings

- NLP classification and expansion

Implement those that lead to improved results.  

Discovery

Current AI solutions for automatic speech recognition (ASR) are ~95% accurate.

This is sufficient to facilitate discovery, but does not meet federal accessibility requirements.

Vendors provide 99% and continue to be the preferable way to meet this need.

Captioning

$11,600

• Hardware for production delivery and research/development.
• Small budget for renting the hardware first to verify its capacity before purchase.
• 2,500 images a month of cutting-edge HTR for our most complex content. ($50/mo ongoing)

Budget