News in Detail
New projects by Prof. Dr Margot Mieskes
„BigScience - Research Workshop on Large Multilingual Models and Datasets“ aims to demonstrate a different way to build, explore and share large language models and large
research artefacts in general within the AI/NLP research communities. It is led by HuggingFace, the AI community that builds, trains and implements state-of-the-art models based on the open source machine learning reference.
Furthermore, Prof. Dr. Mieskes is involved in the project "ReproHum: Investigating the Reproducibility of Human Assessments in NLP" (led by the University of Aberdeen; funding: EPSRC UK). The main aims of the ReproHum project are to develop a methodological
framework for testing the reproducibility of human assessments in NLP and a multi-laboratory paradigm for conducting such tests in practice, with the first study of its kind being conducted in NLP.
Another current project of Prof. Dr. Mieskes is "Multi-CrowsPairs - a multilingual database to
detect biases and stereotypes in language models" (in cooperation with the Université de Sorbonne, France). It is well known that language models, such as those used in machine
language processing to generate texts, are full of biases and prejudices. In the project "Multi-CrowsPairs", in cooperation with various partners, she is creating a data set in different
languages (e.g. German, Spanish, Chinese, Arabic, Portuguese, etc.) in which various biases are listed. These are used to check language models of the respective languages for their prejudices and biases.