Contentious Descriptions and Bias in Published Photographic Heritage Collections: A Digital Humanities Approach

In this paper, we discuss the approach taken in the Digital Europe project “De-Bias - Detecting and cur(at)ing harmful language in cultural heritage collections”1. This project tries to address a growing problem for heritage organizations. With the massive digitization efforts and the push for Open Access, many institutions publish their digital catalogs online. In doing so, these heritage organizations become publishers, which entails specific responsibilities. However, often the historical descriptions of these objects stem from distant times, and use a language that is no longer valid today. In this way, un unwanted performativity arises, where possibly harmful content is published. This issue is exacerbated by the fact that GenAI algorithms tend to be trained by these large open collections, such as Europeana. They risk learning foul language in doing so. The De-bias consortium, composed of aggregators (such as Europeana, European Fashion Heritage Association and Michael Culture, Archives Portal Europe, French Ministry of Culture), archives and collection holders (such as Deutsche Film Institute - the project lead and NISV) and academic experts both from Cultural Studies, History and Digital Humanities (DigitGLAM KU Leuven) as well as engineering (ThinkCode, Datoptron) and community experts (ECCOM) is trying to address this issue by developing a tool to automatically detect problematic terms in cultural heritage metadata and provide information about their problematic background. It will use vocabularies that combine offensive language with contextual information and suggestions for appropriate terms. The chosen approach is a typical Digital Humanities setting, in which different disciplines are Integrated in one process flow, in 6 phases that showcase an interesting example of the pivotal role of Digital Humanities in interdisciplinary research. First of all, the vocabulary is based on both a literature scoping as well as new research by colleagues from history and cultural studies at Leuven University, who researched and compiled a list of 600 terms in 6 languages of contentious terms. Second, this vocabulary has been validated with communities, using both anthropological as well as oral history methods, in collaboration with heritage institutions and community organizations. For this, a specific workshop methodology involving a “Wheel of Bias” has been developed. In a third step, the vocabulary is transformed into a knowledge graph and subsequently integrated in the developed AI tool by the engineering teams, which is based on Named Entity Recognition algorithms. A second round of validation is on the interface and implementation and interface, which is done according to UX design principles. The last step is the actual integration in the Europeana platform. The tool will be both available integrated in the Europeana workflow as well as being offered as a standalone tool. Separately, the vocabulary will be integrated into the EU VocBench vocabulary system as Linked Open Data, as well as published as an independent resource accessible through an API.