Automated Classification of DFG Subject Areas with Deep Learning

The TIB Knowledge Loom and the Open Research Knowledge Graph (ORKG) aim to organize and interlink scientific knowledge. A practical bottleneck in such knowledge graph construction is the assignment of research content to the correct subject areas, which is traditionally done through manual expert review. As the number of projects grows, this workflow becomes costly and difficult to scale. This thesis investigates whether transformer-based language models can automate DFG subject area classification from project descriptions. We compare traditional TF–IDF approaches with fine-tuned transformer encoders on a dataset of project titles and abstracts from the GEPRIS database. The results indicate that modern transformer models can support scalable and consistent subject area assignment, reducing the manual effort required for research management and knowledge graph construction.