Interview
Nothing about us without us

Over the course of two years, eleven partner organisations will work together with minority communities to develop an AI-powered tool that will automatically detect problematic terms in cultural heritage metadata and provide information about their problematic background. The DE-BIAS project aims to cultivate a more inclusive and respectful approach to the description of digital heritage collections and tell the stories and histories of minority communities.

Kerstin Herlt is the EU project coordinator at the Deutsches Filminstitut & Filmmuseum (DFF) in Frankfurt and the DE-BIAS project coordinator. The European Fashion Heritage Association (EFHA) in Florence is another partner organisation. Marta Franceschini, who is Head of Communications and Editorial has been actively involved in the development of the tool. For this article, we spoke to both of them about how bias finds its way into collection descriptions, the limitations of AI in detecting harmful language and the importance of community involvement.

By Lucy Rowan

Could you explain the societal significance of developing the DE-BIAS tool? 

Kerstin: Not only did we want to raise awareness about offensive language within the collection descriptions of cultural heritage institutions, but we also wanted to promote a more inclusive and up-to-date approach to describe these collections. Therefore, we have been developing a tool that automatically detects and flags offensive terms. The project creates a list of vocabulary, which we call "bias vocabulary" that combines offensive language with contextual information and, if appropriate, also suggests alternative terms. The most important thing with this project is that the vocabulary will be co-created with the communities affected and impacted by discriminatory language regarding their religion, race, gender, sexual orientation, and/or ethnicity. 

The main case study for the project is Europeana's digital collections. But we also will make this tool available for cultural heritage institutions so they can check their collection descriptions. We are not removing or changing any terminology, it is up to them to make those decisions about how to deal with that. 

Why is this important for society? I think acknowledging the identity and cultural memory of communities might not repair historical or even actual injustices such as colonialism or gender discrimination. Providing a more accurate and respectful representation of cultural heritage can at least help to create a better cultural understanding. 

Marta: Yeah, I think the benefit is not just for the people involved in the project. It's also sustainable for the future. Who uses these data sets? Who sees and uses them for research or even personal research? It might simply be someone who's looking for something interesting online. The idea is to create a more inclusive and open environment so that everybody can have all the information, and also track the changes in the ways these objects have been catalogued for museum collections. 

So it's definitely about acknowledging the past and how the tradition of collecting and cataloguing has been carried out in different areas in Europe. Then it's about trying to rectify and show how we can move forward. How can looking at the past inform collecting and cataloguing in the future?

What role is the DFF playing in the project and what work are you doing with communities?

Kerstin: All in all, the consortium consists of eleven partners from all across Europe. Most of them are connected to Europeana in one way or another - being an aggregator or a partner in Europeana. Those partners are working with communities and are focusing on main themes within vocabulary: migration, colonial past, gender, sexual identity, ethnicity and ethno-religious identity. 

As for the DFF, we will look into the anti-semitism in language, first and foremost within German. The co-creation work with the community and the experts will not be so much focussed on identifying biased terminology, but more on how we can close the gap in our local databases and the collections to give more visibility, for example, to Jewish filmmaking and Jewish film production. This will probably also impact the way we describe these collections in the future, and what terms we will use to describe them. 
 

What is the role of the EFHA in the process and how are communities involved in your work? 

Marta: We are coordinating the communication and dissemination of the project. So we are working on the ways the project reaches out and how its output is communicated to an unspecified and general audience. Let's say this is the main section. 

We are also in charge of detecting and working on the gender and sexuality bias in collections. Gender and sexuality, especially sexual identity is kind of concealed in our records, it's a real visibility issue. Gender is something that has been dictated through garments, clothing and apparel. We're reaching out to different communities in the UK and in Italy, to try and understand where the bias is or where the problems are exactly. 

For us, the problem is more misinterpretation and misuse of terms that come in and out of fashion. Something that was considered appropriate ten years ago, is no longer appropriate today. The main issue is that certain data has been concealed or not mentioned properly because it was problematic for the cataloguer, for instance. 

How does ethno-religious-bias such as anti-Semitism make its way into collection language? 

Kerstin: Cultural heritage institutions are mainly concerned with historical objects. So anti-Semitism as an ideology finds its way into the collections and their descriptions through the objects. In Germany, this is of course a special case because of the Nazi Regime and the Second World War. But as far as we are concerned at DFF, there are many propaganda films produced during the Nazi regime, but they are now banned. If you want to show this kind of film, you need to provide an educational or research purpose with context and an introduction to why the content is racist and harmful. Whether the same is true for other countries, I am not so sure. I would expect, some sort of anti-Semitic propaganda can also be seen in collections, including non-German ones. 

It's important to note that Germany is a very specific case language-wise. You also cannot really translate these terms or their bias into English. Everything is very context-dependent.
 
In language more generally, you can easily stumble across anti-Semitism on social media. There is a lot of research that has been done, which demonstrates that it is more covert, and more of a combination of words instead of specific terms. So it's rather the context that you associate "Jewishness" with, for example, with money or the science movement. That's typically the way anti-semitism makes its way into language in a more insidious way online. 

How does sexuality/gender-bias find itself in collection language? 

Marta: In several ways, but the most worrying is the absence of it. The more we look for information on sexuality and gender, the more we realise that it's difficult to find anything appropriate because the words used typically reflect the historical moment and the geography of where the entry was created. Who wrote the description also plays an important role. 

We're very aware that we are dealing with historical documents and dealing with one language alone already comes with its challenges. So when Europeana translates the records into different languages, things become much more complicated. We've seen this when comparing Italian and English. A word that has offensive connotations in English, might not in Italian. The words are cultural artefacts within themselves. 

Let's take the example of the word "queer" for instance. In English, it historically was a derogatory word that then got reappropriated by the community. In English, it could be considered an insult, but in Italy, the word arrived through LGBTQ+ theory so its connotations are only positive. Language and its nuances develop very quickly - this new notion of "Queerness” only came about in the last ten years. 

I have to say that invisibility is a bias in itself because people who identify as "queer", might type in that word in a database and may be faced with pages and pages of unrelated content. It's actually very common that the words in our theme (gender/sexuality) sometimes mean something else as well. So it is also about detecting where they're used in a derogatory way and maybe bringing them forward because visibility for the community is a big issue. Not only do they want to be seen, but they also want their mistreatments to be acknowledged and seen. 

Marta mentioned the role that the geopolitical location of a collection plays in collection descriptions and the bias found? Kerstin do you have any examples from your research? 

Kerstin: I did some research on ethnographic films and I discovered the interview with a researcher from Angola. He was a filmmaker, who was looking for films from the Civil War in Angola. Whilst he was searching collections in Germany, he couldn't find anything because he had put in the “wrong” keyword - he was looking for "liberation" and "revolution" films. Instead, the films were catalogued as "terrorism" and "rebellion". So he didn't manage to find the films in the collection. 

Perspective and someone’s socio-political stance/upbringing are factors in the way you label things, whether you consider a movement as liberation or as an act of terrorism. And this is something you can find throughout history and across different geographical locations. 
 
Why is it important to have communities at the heart of these projects? 

Marta: What we want to have at the end of the project is a tool, which can be used by someone effectively. It's not something that will run itself. That's why having people in our communities involved in the project as intermediaries, gives us direct exposure and puts us in contact with community members themselves. It's a group effort. It's so important not just to develop the tool, but to be informed by the needs of the community and the people who will actively use it - communities, professionals, curators and conservators. 

I think that we're still in the process of understanding how, after detecting the words, what we're going to do next. We have discussed the prospect of the tool having a short description that explains why the word was highlighted or detected by the DE-BIAS tool. Those definitions have to come out of conversations with people who are part of these communities and who are connected with these objects. In this sense, I think about it as a way of writing a history of the future. It's not just for past objects because museums acquire new things daily and are constantly adding to their databases. 

As we said, language is situated geographically and chronologically, so if these objects are still described today, we also want to furnish and assist curators and conservators on how to catalogue the future, and expand collections in a more inclusive way. This cannot be done by gatekeeping information. It's something that has to be co-created. So it's not just the vocabulary that has to be co-created, but ultimately all of the knowledge.

What are the limitations of using AI to detect harmful language? At what point do humans need to be involved in the process and why?

Kerstin: Marta has already mentioned this, but one word has the potential to be offensive in one language or not in another. Meanings also change with time. The tool we are building with the technological partners is mainly based on natural language processing. So it's an AI power tool, but we always say it's more "soft AI" - there's no algorithm that we are going to train with data, and if so, I think we would end up reproducing and perpetuating bias in collection descriptions instead. If you assume that language is biased by nature, in a way that also implies that language produces different meanings and terminology always has multiple connotations. So one term can be considered harmful for one person, but not for another one. So context is important. 

That's why you need people, and in our case, communities and experts - some sort of "humanity" in the loop. I don't think this work can be fully automated. What can be, however, the running of the tool through millions of records. It's so great that you can manage such a large amount of records. In the end, it helps to reflect on specific terms to contextualise them. It ultimately comes down to the curator or the archivist - the person who has to decide how to deal with that, and that's the process that cannot be automated. 

Beyond the DE-BIAS project, what more needs to be done to improve the bias we find in language regarding minority groups? What changes would you like to see? 

Kerstin: I would like to see more awareness, which I think is still lacking. Of course, some progress has been made in the last few years. Sometimes there is a lack of time and resources, but I would still like to see more action - archivists and curators should spend more time evaluating their catalogues and labels for exhibits. 

On a larger scale, it's really about reconsidering collection strategies and exhibition strategies as a whole. It's not only about reflecting on the language but also about what objects we collect and from whom. What's our collection strategy? How do we reflect the cultural production of communities in our world? In Germany, we have a large migrant community - a very big Turkish community, which is not aptly reflected in our collections. 

I think something that can also help improve this is to have more diversity managers in museums. We need people who make the connections between the staff and the communities to create joint projects. They need to train our staff on diversity and we also need more diversity among the staff members. I think that will have an impact on how we name and describe objects. Some progress has been made already, but there's still a long way to go. In Germany alone, we already see controversial debates and backlash over the use of gender-neutral language. So I think in the end, it's more than just changing mindsets. 

Marta: Yeah, just echoing that. I think the change that needs to happen is institutional, if not political. The problem is within the system. If we don't change the system we are just fixing the facade without really looking at the building underneath - papering over cracks. I think that the strategies to implement systemic change can be grassroots - they can grow from the bottom to the top. That's what we're trying to foster in a way with this project. 

I think that visibility is a very important issue. We need to be listening to people who are affected by these issues because the more "professional take" on these topics tends to attach labels and create more biases. This is one of the reasons why we think that our methodology has to be dictated by the community - how do they want to be involved in the project? Not us telling them. Otherwise, what's the point? 
Ultimately, national institutions need the budget to run these projects, but also enough to pay their curators, conservators and archivists for their daily tasks. So if there's no support from political powers, it's very hard for them to make a difference. We see it all the time - some great curators and conservators work a lot outside of their regular hours to look back at collections and see what they can do to improve them. 

When it comes to the curating of collections, there should also be a long-term, sustainable effort to gain the trust of communities. They don't want to be used for one specialist exhibition together and then left behind. Building this relationship shouldn't be special. It should be an inherent daily task. It should be internalised in the role and the job description. 

Kerstin Herlt Kerstin Herlt works as EU project coordinator at the DFF - Deutsches Filminstitut & Filmmuseum in Frankfurt am Main. Current projects and initiatives include EFG - The European Film Gateway which aggregates film heritage data to Europeana and DE-BIAS, which promotes an AI-supported and inclusive approach to cultural heritage collections descriptions. Kerstin Herlt holds a master's degree in Roman Philology, Sociology and European Media Studies.  










Marta Franceschini Marta Franceschini is a design historian and researcher. She serves as Head of Communications and Editorial for the European Fashion Heritage Association. Marta holds a PhD from Università Iuav di Venezia and regularly lectures material culture, research methodologies and curation. She collaborates with brand archives and museums and has worked on various exhibitions, including Fashioning Masculinities: The Art of Menswear at the V&A.