1 of 12

Using AI to format Alma item descriptions

WRLC’s First Artificial Intelligence Project

2 of 12

A Holy Grail of Shared Print

Other tools like OpenRefine might work, but…
AI is teachable and iterative
This is the only way we’ll all eventually get on the same page for shared print periodicals
It’s a project that can be done centrally to benefit all partners

If you’ve worked around shared print for any period of time, you will know that serials holdings statements are problematic. There is so much variability in the way people have cataloged volume, number, part, and date information over the years. And that makes it extremely difficult to quickly determine that the volume we have on the stacks or in a storage facility in Washington DC is unambiguously the same volume sitting at the University of Washington in Seattle even if mine says v.1 and their’s says vol. 1. And it’s not just a problem across consortia, it’s a problem within consortia and within individual institutions. So that also makes it a problem for discovery and access. WRLC is a founding member of the Rosemont Shared Print Alliance and the Partnerhship for Shared Book Collections. Getting standardization of these fields will help the collaboration between regional programs interested in coordinating their efforts on a larger scale to ensure the retention of and access to print journal backfiles.

3 of 12

Getting Started

Last year WRLC’s recommendations for print periodicals were written by the Serial Holdings Subgroup and approved by the WRLC Metadata Committee.

4 of 12

The Proof of Concept

Slowly refined a query with ChatGPT
Item description templates were very helpful
Great opportunity to begin learning to use AI
ChatGPT free tier allows 100 lines of input which

is plenty for testing a query statement

Back in the spring of 2023, OpenAI released its ChatGPT AI platform to the public for free trials. It gained popularity and public awareness very fast and this summer I starting using it like everyone else to slowly refine prompts to generate the one perfect library robot picture. One great thing about artificial intelligence is that it mimics organic intelligence in learning through plain language. So I applied this same idea to see if we can use AI to apply the standardization for those 900,000 existing item descriptions. I started with a basic prompt to reprint a single description and it worked. Lucky for me, we librarians love to create workflows and guidelines to help preserve institutional knowledge and those are perfect documents for teaching AI also. The item description templates created by our Metadata Committee are a good example of that. They created at least a dozen examples of what faulty input and ideal output should look like and how to correctly express statements for volume, part, page, etc etc. Perfect for training AI.

5 of 12

6 of 12

7 of 12

8 of 12

9 of 12

What it can fix…

and what it can’t.

Original Item Description	Reformatted Item Description
v.30 apr.-june 1994 pp.663-1350	v.30 (1994: Apr.-June) pp.663-1350
v.30, no.1 suppl. 1940	v.30: no.1 (1940) suppl.
Index v. 72 Jan.-June 1998	v.72 (1998: Jan.-June) Index

Original Item Description	Reformatted Item Description
v.39, pp. 1623-3606 Oct-Dec 1963	v.39 (1623-3606: Oct.-Dec.) pp. 1623-3606
January 1958 - November 1964 January 1958 - November 1964	(1958: Jan.-Jan.-Nov.)
v.28, nos. 1+12 1981	v.28: no.1 (1981)

This solution has been very successful in fixing variable cataloging. If the information is there, but it’s in a format that is out of spec, you should have great results getting it into spec. It can fix bad or non-standard item descriptions. However, it’s still somewhat of a garbage in/garbage out proposition. It cannot obviously fix blank item descriptions or situations where there were input errors like barcodes that accidentally got scanned into the middle of a string of otherwise correct information. After refining our queries over a few weeks and identifying the most common standardization errors, we were able to coach ChatGPT into 80-90% accuracy. Jackie Saavedra on my team worked with ChatGPT to develop a way for the queries to do verification and data validation on their own work. So we know when an item description element that appeared in the original metadata cannot be reformatted in the new metadata and those items are not changed in Alma.

10 of 12

Checking our work…

Original Description	Reformatted Description	Missing information
v.92 pg.961-1944 1996	v.92 (1944) pg.961-1944	1996
Jan. 1981 - June 1982	(1981: Jan.-June)	1982
v.33:no.9,11 2020	v.33: no.9 (2020)	11
nos.3-5, 10	no.3-5	10

11 of 12

Next Steps…

From normalized item descriptions, to 866s to coded fields
National/international policies and templates for standardized 866s
Or not!
What other periodical problems can this solve?

I’ll admit that I’m not a cataloger and as part of this project had to learn how MARC 863 structured fields relate to the enumeration and chronology fields in an Alma item record. My use case was only fixing the item description field in Alma and improving summary holdings. However, Jackie on my team is a great cataloger and was able to also use ChatGPT to break these item descriptions into their individual enumeration and chronology fields so that the MARC structure would also be corrected. So we’ve actually been able to do two things at once. Fix these descriptions and also the enumeration and chronology fields that they are composed of.

The biggest problem in working with item description fields and serials in general in the past has been that there’s so much variation, it’s impossible to write code to manipulate them. Sometimes the only way to know for sure that two items were the same volume of a title was to manually compare them. But now that we are applying some standardization, we can actually start to use code to do things like break them down into component parts and put the various enumeration and chronology bits into item records and MARC records. So maybe groups like the Rosemont Shared Print Alliance now need to work on national standards so that everyone’s item descriptions look the same. But then again, maybe they don’t. Maybe we can all continue to have our own holdings statements and when we need to compare them to each other, an entity like CRL will us AI to enforce its own normalization. I’m not sure. But at the very least this can make our own metadata much better. We’ve already changed all of the item descriptions in the SCF holdings and have received permission from at least two of our institutions to do this same formatting on their local records also.

12 of 12

Q&A

Ask me some Qs and I’ll try to A them