Facundo Iglesia 
Every day for 40 years, a grandmother bent over a desk, patiently cutting out news articles with a pair of scissors. Raquel Radío de Mazcurra was scouring the newspapers for hints about her grandchildren. She carefully pasted and stored the clippings in folders.
Raquel was one of the twelve women who, in 1977, founded the Grandmothers of Plaza de Mayo. The mission of the Grandmothers (Abuelas) is to find the children of people kidnapped during the 1976-1983 civic-military dictatorship in Argentina. These children, many of whom were born in horrific conditions in clandestine detention centers, were taken from their parents and illegally adopted under false names. They are the Abuelas’ grandchildren.
Raquel passed away in 2017 without ever being reunited with her grandchildren. But her now-famous folders were the basis of the Grandmothers’ journalistic archive, which is stored in the ex-ESMA — a former illegal detention center used by the dictatorship, which is now a memory site and home to various human rights organizations. The archive is made up of more than 10,000 clippings related to the disappeared (desaparecidos), the human rights movement in Argentina, and the military dictatorship.
However, this valuable archive is not easily accessible to the public. The folders are scattered in piled-up boxes in a room in the ESMA. How could it be made available to anyone and everyone who wants to consult it?
Enter artificial intelligence, and a team of three 23-year-old college students.
“This effectively democratizes people’s access to the information,” Julieta Goria, a student at the Exact and Natural Sciences School at Buenos Aires University told the Herald. Goria was part of the team that won the “AI for Identity” award, a competition organized by the Sadosky Foundation, the Science and Technology Ministry and the Grandmothers of Plaza de Mayo. The goal of the contest was to develop software that could make the Abuelas’ archive accessible to anyone.
The team, made up of Goria, Sofía Goy and Francisco Sandalinas — a group of friends and students at the university — created software that easily indexes the transcribed scans of Raquel’s clippings. It is now available on a public website.
“Imagine you are researching, for example, the repercussions of the dictatorship in [the suburb of] Olivos and you want to read every article available. You can now search through every article by just typing ‘Olivos’ into the website, instead of physically going through every clipping,” Sandalinas added.
The team — named “The Cicadas” in honor of a popular 1970s song that became a hymn for democracy in Argentina — were able to visit the actual archive in the ESMA. “It was a beautiful experience being so close to the Grandmothers’ archive,” Sofía said. “There are boxes and boxes and boxes piling up. It was a very handmade archive. We were told about Raquel, who sat at her desk every morning cutting out newspapers and pasting them into her little folder. That became a gigantic journalistic archive. But it’s paper — anything could happen to it.”
How did you approach the problem?
Julieta Goria: First of all, we researched. We read a lot about what techniques were used to solve this kind of problem and what we ended up deciding was to train an AI — more properly, a neural network — that knows how to segment an article. Think of a newspaper as an image. We would like to be able to separate it into its semantic parts: the date, the headline, the body of the text, captions, et cetera. When you give it to transcription software, it’s going to be all jumbled up — the title on one side, then a photo that cuts off the text… So there is a task, identifying the parts of the text and being able to read them in a logical order. So our work consisted first of all of training an object recognition engine. We trained it and put the database together.
How does the training work?
Sofía Goy: By hand. We had to select a little square around each title, each photo, each date… I don’t know, we must have done it with 400 or so images.
JG: With that we trained the network that draws a little square around each part and says which part.
Do you remember when it worked for the first time?
JG: Yes. We couldn’t believe it! AI in general works like a black box — you feed it with a lot of data and then maybe something happens. You don’t know until you try. So, when we saw that it worked, it was like “ok, we weren’t delusional with our plan”.
SG: We didn’t have a plan B.
Francisco Sandalinas: Generally you keep a couple of cases that you see are complicated: one that is all smudged, one that has a stamp in the middle, one that is so badly scanned that you can’t read a single letter. You save them and when you finish training the system, you try them out. The crazy thing was how well it gave us scans that we said, “this is horrible, this is going to break everything”. And it worked.
JG: When we went [to the archive], people from Abuelas told us that they always had a very strong connection with the forefront of technology. The genetic procedure to determine whether a person is someone’s grandchild did not exist until they contacted scientists and were able to put together the “grandparental index”. For us, this experience is very comforting and also a bit of a personal reason why I am studying what I am studying.
What can you tell us from the awards ceremony?
JG: It was very emotional. [Abuelas’ Vice-President] Buscarita Roa spoke about the uncertainty about the future of science, about the fear of what would happen in the future when the Abuelas may no longer be around…. But she was happy that there are still people with the drive and motivation to continue the work that they started.
