Why We Need Culturally Aware AI
Biases in data can be both explicit and implicit. Explicitly, "The Dutch Seventeenth Century" and "The Dutch Golden Age" are pseudo-synonymous and refer to a particular era of Dutch history. Implicitly, the "Golden Age" moniker is contested due to the fact that the geopolitical and economic expansion came with great costs, such as the slave trade. A simple two-word phrase can carry strong contestations, and entire research fields, such as post-colonial studies, are devoted to them. However, these sometimes subtle (and sometimes not so subtle) differences in voice are as yet not often found in the results of automatic analyses or datasets created using automated methods.
The reason for this is that current AI technologies and data representations often reflect the popular or majority vote. This is an inherent artefact of the frequentist bias of many statistical analysis methods resulting in simplified representations of the world in which diverse perspectives are underrepresented. With the use of AI-infused tools ever increasing, as well as the diverse audiences using these tools, this bias needs to be addressed in both the algorithms analyzing data, as well as in the resulting representations. In this lecture, I will discuss how the Cultural AI Lab is working towards mitigating such biases.
Marieke van Erp is a Language Technology and Semantic Web expert with a penchant for interdisciplinary research. She leads the Digital Humanities Lab at KNAW Humanities Cluster (Amsterdam) and is a co-founder and scientific director of the Cultural AI Lab, a collaboration between various heritage and research institutes in the Netherlands. The Lab is aimed at developing socio-technological AI systems that are implicitly or explicitly aware of the subtle and subjective complexity of human culture. Van Erp’s work spans many different domains, from olfactory, socio-economic, and maritime history to data journalism and cultural heritage but is always driven by language understanding and knowledge representation.
Wegen der weiterhin bestehenden Einschränkungen wird das Kolloquium im Online-Format stattfinden. Für Details und aktuelle Informationen siehe https://isis.tu-berlin.de/course/view.php?id=20896.