The next step is to send the detailed descriptions on notecards to Chreston Miller, data and informatics consultant at University Libraries, where he and his student team oversee the natural language processing, also known as machine learning. There, they compare the cards and predict what Costume Core terms should be used to describe the item.
“This project is pretty interesting,” said Miller. “I’m a computer scientist, so when I tell people I’m working with fashion, they’re like, what connection is that? The people in Fashion Merchandising and Design have data they want to work with.”
Given the challenges of describing artifacts, especially ones of historic nature, Miller said one reason to have a controlled vocabulary is the descriptions that come with an artifact were written in a certain time period, so they have certain phrases for different aspects of the artifacts. “We read it today and we said, ‘What does this mean?'” said Miller.
Graduate student and University Libraries student employee, Madhuvanti Muralikrishnan ’23, is assisting Miller on the project while working on her master’s degree in computer science and applications. Her role is implementing the algorithms that map a given description to the Costume Core vocabulary.
“In machine learning, we’re essentially trying to teach the computer to do something that humans do, and that’s pretty complex,” said Muralikrishnan. “We know what multicolored is, but how do you teach that to a program and the intricacies involved in that?” That was surprising for me. I read about these things in class, and it’s very gratifying being able to apply what I learned and seeing the result of it.”
Muralikrishnan said after receiving her graduate degree she wants to work in machine learning, and this project is directly related to her career goals. “I believe that computer science is domain agnostic. So today we’re doing this for fashion. Tomorrow, I can take the same algorithm and do it for medicine or something else. Each domain has its own problems and challenges and that’s been fascinating.”
So far, Miller has received more than 5,000 notecards with descriptions that need digitizing and organizing into a database so everything is accessible and searchable. The notecards start off on paper, then Miller performs optical character recognition on the notecards so the team can digitize the descriptions from the notecards, a crucial step for making the natural language processing possible.
One item or notecard can take up to 45 minutes to process. The team also uses an online database to share information. “You think, ‘OK, color doesn’t seem too hard to describe,'” said Miller. “Well, how many different words of color can you use? Take blue for example. There’s navy blue, baby blue, royal blue, and so on. So you have to think about how you have a control vocabulary for this.”
Terms are inconsistent over time. “I noticed in a lot of our old records that some of the terminology is not quite accurate or is very specific to the time period in which it was recorded,” said Smith-Glavina. “One thing that has stuck out to me in the collection is the term bloomers. Bloomers was actually a term used in the 1850s to describe the first form of female trousers worn by dress reformers and by bicyclists as a sports costume. But over time, it changed to be ruffly, fluffy balloon-like undergarments mainly for babies — the ruffly underwear that they wear over their diapers that people tend to describe as bloomers.”
Virginia Tech’s rare fashion artifacts are usually donated by people cleaning out the homes of their parents or grandparents, and they find trunks with many generations of old garments in them.
“Most of the time they won’t know what the items are and it’s up to us, the experts, to identify what the garments are,” said Smith-Glaviana. “We’ll ask the donors questions like, ‘Who do you think wore this?’ And that will help us figure out if it’s a female or male, older or younger person, or even a child because sometimes it’s hard to tell with historical clothing. We get a sense of the history from the donor but usually their knowledge is quite limited.”
Their ultimate goal is to digitize the historic costume collection and make it accessible to the public anyone who is interested in researching historic costumes. Terms that are accurate, understandable, and consistent allow the average website user or noncostume history expert, to search websites terms that they understand even when the items they are looking for are very specific.
“For example, most people know corsets as corsets,” said Smith-Glaviana. “But they don’t realize that throughout much of history corsets are actually referred to as ‘stays’ and they’re two different garments. So depending on the words they search, they may never find what they’re looking for.”
.