Art-historical archives often contain large amounts of documents, such as auction
catalogues, letters, or newspapers. These documents enhance our understanding of both
the artists themselves and their work. Providing them in a digital way can make them more
accessible to art enthusiasts and researchers worldwide. In addition, the digital format
allows for new ways of searching and filtering, provided the necessary metadata is available.

Machine learning methods can help acquire this metadata automatically, without human
work. One such method is language identification. It is useful both on its own, e.g., by
enabling the filtering of documents by language, and as a preprocessing step for other tasks,
such as optical character recognition (OCR).

Many state-of-the-art machine learning methods rely on large amounts of labeled training
data. However, while we do have a large dataset of historical documents given to us by our
project partner, the Wildenstein Plattner Institute (WPI), we do not have labels available.
Thus, this project aims to solve the language identification task without labels.

The goal of this project is to develop a method for language identification that does not
require a large, labeled dataset. For this we are going to:

• Familiarize ourselves with the state of the art in the field
• Compare different approaches with varying degrees of supervision and decide on the
  most feasible
• Implement the selected approach and evaluate its performance on the data given to
  us by the WPI
• Develop a prototype to showcase the results

• IT-Systems Engineering
• Data Engineering
• Digital Health
• Cybersecurity

• Jona Otholt
• Hendrik Rätz
• PD Dr. Haojin Yang
• Prof. Dr. Christoph Meinel